[Perf] (CI): rewrite one rolling ccache entry per job on the default branch instead of saving one per run

.github/scripts/fetch-trace-fixture-lfs.sh

@@ -1,6 +1,10 @@
 #!/usr/bin/env bash
 set -euo pipefail
 
+script_dir="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+# shellcheck source=trace-fixture-lib.sh
+. "${script_dir}/trace-fixture-lib.sh"
+
 if [ "$#" -lt 1 ] || [ "$#" -gt 2 ]; then
   echo "usage: $0 <trace-case> [fixture-dir]" >&2
   exit 2
@@ -62,57 +66,6 @@ if [ "${case_name}" = "OpenRA" ]; then
   exit 0
 fi
 
-get_lfs_metadata() {
-  local file="$1"
-  local pointer
-  local expected_oid
-  local expected_size
-
-  if ! pointer="$(git show "HEAD:${file}" 2>/dev/null)"; then
-    echo "failed to read tracked fixture metadata: ${file}" >&2
-    return 1
-  fi
-  if ! grep -q '^version https://git-lfs.github.com/spec/v1$' <<< "${pointer}"; then
-    echo "tracked fixture is not a Git LFS pointer: ${file}" >&2
-    return 1
-  fi
-
-  expected_oid="$(awk '$1 == "oid" && $2 ~ /^sha256:/ { sub(/^sha256:/, "", $2); print $2 }' <<< "${pointer}")"
-  expected_size="$(awk '$1 == "size" { print $2 }' <<< "${pointer}")"
-  if ! [[ "${expected_oid}" =~ ^[0-9a-f]{64}$ ]] || ! [[ "${expected_size}" =~ ^[0-9]+$ ]]; then
-    echo "invalid Git LFS pointer metadata: ${file}" >&2
-    return 1
-  fi
-
-  printf '%s %s\n' "${expected_oid}" "${expected_size}"
-}
-
-verify_fixture_file() {
-  local downloaded_file="$1"
-  local display_name="$2"
-  local expected_oid="$3"
-  local expected_size="$4"
-  local actual_oid
-  local actual_size
-
-  if [ ! -f "${downloaded_file}" ]; then
-    echo "fixture file is missing: ${display_name}" >&2
-    return 1
-  fi
-
-  actual_size="$(wc -c < "${downloaded_file}" | tr -d '[:space:]')"
-  if [ "${actual_size}" != "${expected_size}" ]; then
-    echo "fixture size mismatch for ${display_name}: expected ${expected_size}, got ${actual_size}" >&2
-    return 1
-  fi
-
-  actual_oid="$(sha256sum "${downloaded_file}" | awk '{ print $1 }')"
-  if [ "${actual_oid}" != "${expected_oid}" ]; then
-    echo "fixture SHA-256 mismatch for ${display_name}: expected ${expected_oid}, got ${actual_oid}" >&2
-    return 1
-  fi
-}
-
 fetch_file_from_mirror() {
   local file="$1"
   local url="$2"

.github/scripts/trace-fixture-cache.sh

@@ -0,0 +1,117 @@
+#!/usr/bin/env bash
+# Cache-side helper for trace fixtures.
+#
+#   key <case> [fixture-dir]     derive the actions/cache key and path list
+#   verify <case> [fixture-dir]  check restored fixtures against their pointers
+#   reset <case> [fixture-dir]   drop restored fixtures, leaving the pointers
+#
+# The cache key is content-addressed on the Git LFS pointer oids tracked at
+# HEAD, which are readable from a plain checkout without smudging. Fixture
+# content therefore maps 1:1 onto a key: unchanged content hits, changed
+# content is a new key and thus a miss, and the download path handles it. The
+# key deliberately carries no restore-keys prefix in the workflow - a fixture
+# that does not match the pointer exactly must never be restored.
+set -euo pipefail
+
+script_dir="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+# shellcheck source=trace-fixture-lib.sh
+. "${script_dir}/trace-fixture-lib.sh"
+
+# Bump when the key derivation changes in a way that must invalidate old
+# entries; the content digest alone would not notice a format change.
+key_schema="v1"
+
+if [ "$#" -lt 2 ] || [ "$#" -gt 3 ]; then
+  echo "usage: $0 <key|verify|reset> <trace-case> [fixture-dir]" >&2
+  exit 2
+fi
+
+command_name="$1"
+case_name="$2"
+fixture_dir="${3:-tools/trace_replay/fixtures}"
+python_bin="${PYTHON:-python3}"
+
+if ! command -v "${python_bin}" >/dev/null 2>&1 && command -v python >/dev/null 2>&1; then
+  python_bin=python
+fi
+
+mapfile -t files < <(trace_fixture_files "${case_name}" "${fixture_dir}" "${python_bin}")
+if [ "${#files[@]}" -eq 0 ]; then
+  echo "no fixture files declared for trace case: ${case_name}" >&2
+  exit 1
+fi
+
+# Writes "name=value" to $GITHUB_OUTPUT when running under Actions, and to
+# stdout otherwise so the script stays runnable (and testable) off-CI.
+emit_output() {
+  local name="$1"
+  local value="$2"
+  if [ -n "${GITHUB_OUTPUT:-}" ]; then
+    if [[ "${value}" == *$'\n'* ]]; then
+      local delimiter="ghadelim_$(date +%s%N)_$$"
+      {
+        printf '%s<<%s\n' "${name}" "${delimiter}"
+        printf '%s\n' "${value}"
+        printf '%s\n' "${delimiter}"
+      } >> "${GITHUB_OUTPUT}"
+    else
+      printf '%s=%s\n' "${name}" "${value}" >> "${GITHUB_OUTPUT}"
+    fi
+  fi
+  printf '%s=%s\n' "${name}" "${value}"
+}
+
+sanitize_case() {
+  printf '%s' "$1" | sed 's/[^A-Za-z0-9._-]/_/g'
+}
+
+case "${command_name}" in
+  key)
+    manifest=""
+    for file in "${files[@]}"; do
+      # A case whose fixtures are committed directly rather than through Git LFS
+      # (OpenRA) has no pointer oid to key on, and nothing to download either.
+      # Report it as uncacheable so the workflow skips the cache entirely.
+      if ! metadata="$(get_lfs_metadata "${file}" 2>/dev/null)"; then
+        echo "trace case ${case_name} is not stored in Git LFS; skipping fixture cache" >&2
+        emit_output "cacheable" "false"
+        emit_output "key" ""
+        exit 0
+      fi
+      read -r expected_oid expected_size <<< "${metadata}"
+      manifest+="$(basename "${file}") ${expected_oid} ${expected_size}"$'\n'
+    done
+
+    digest="$(printf '%s' "${manifest}" | sha256sum | awk '{ print substr($1, 1, 16) }')"
+    safe_case="$(sanitize_case "${case_name}")"
+
+    emit_output "cacheable" "true"
+    emit_output "key" "trace-fixture-${key_schema}-${safe_case}-${digest}"
+    emit_output "paths" "$(printf '%s\n' "${files[@]}")"
+    ;;
+
+  verify)
+    for file in "${files[@]}"; do
+      metadata="$(get_lfs_metadata "${file}")"
+      read -r expected_oid expected_size <<< "${metadata}"
+      verify_fixture_file "${file}" "${file}" "${expected_oid}" "${expected_size}"
+    done
+    echo "Verified ${#files[@]} fixture file(s) for ${case_name} against the tracked Git LFS pointers."
+    ;;
+
+  reset)
+    # Put the working tree back to the pointer files a fresh checkout would
+    # have, so that a rejected cache entry falls through to exactly the same
+    # download path a cache miss takes.
+    for file in "${files[@]}"; do
+      rm -f "${file}" "${file}.tmp"
+    done
+    git checkout -- "${files[@]}"
+    echo "Reset ${#files[@]} fixture file(s) for ${case_name} to their tracked Git LFS pointers."
+    ;;
+
+  *)
+    echo "unknown command: ${command_name}" >&2
+    exit 2
+    ;;
+esac

.github/scripts/trace-fixture-lib.sh

@@ -0,0 +1,73 @@
+#!/usr/bin/env bash
+# Shared helpers for trace-fixture handling: reading the in-tree Git LFS pointer
+# metadata and verifying a fixture file against it. Sourced by
+# fetch-trace-fixture-lfs.sh (verify after download) and by
+# trace-fixture-cache.sh (cache key derivation and verify after cache restore),
+# so both paths agree on what a valid fixture is.
+
+# Reads the Git LFS pointer tracked at HEAD for a fixture path and prints
+# "<oid> <size>". Fails if the tracked blob is not a well-formed LFS pointer.
+get_lfs_metadata() {
+  local file="$1"
+  local pointer
+  local expected_oid
+  local expected_size
+
+  if ! pointer="$(git show "HEAD:${file}" 2>/dev/null)"; then
+    echo "failed to read tracked fixture metadata: ${file}" >&2
+    return 1
+  fi
+  if ! grep -q '^version https://git-lfs.github.com/spec/v1$' <<< "${pointer}"; then
+    echo "tracked fixture is not a Git LFS pointer: ${file}" >&2
+    return 1
+  fi
+
+  expected_oid="$(awk '$1 == "oid" && $2 ~ /^sha256:/ { sub(/^sha256:/, "", $2); print $2 }' <<< "${pointer}")"
+  expected_size="$(awk '$1 == "size" { print $2 }' <<< "${pointer}")"
+  if ! [[ "${expected_oid}" =~ ^[0-9a-f]{64}$ ]] || ! [[ "${expected_size}" =~ ^[0-9]+$ ]]; then
+    echo "invalid Git LFS pointer metadata: ${file}" >&2
+    return 1
+  fi
+
+  printf '%s %s\n' "${expected_oid}" "${expected_size}"
+}
+
+# Checks an on-disk fixture against the size and SHA-256 from its LFS pointer.
+verify_fixture_file() {
+  local downloaded_file="$1"
+  local display_name="$2"
+  local expected_oid="$3"
+  local expected_size="$4"
+  local actual_oid
+  local actual_size
+
+  if [ ! -f "${downloaded_file}" ]; then
+    echo "fixture file is missing: ${display_name}" >&2
+    return 1
+  fi
+
+  actual_size="$(wc -c < "${downloaded_file}" | tr -d '[:space:]')"
+  if [ "${actual_size}" != "${expected_size}" ]; then
+    echo "fixture size mismatch for ${display_name}: expected ${expected_size}, got ${actual_size}" >&2
+    return 1
+  fi
+
+  actual_oid="$(sha256sum "${downloaded_file}" | awk '{ print $1 }')"
+  if [ "${actual_oid}" != "${expected_oid}" ]; then
+    echo "fixture SHA-256 mismatch for ${display_name}: expected ${expected_oid}, got ${actual_oid}" >&2
+    return 1
+  fi
+}
+
+# Prints the fixture file paths of a trace case, one per line. Strips CR so the
+# result is usable when python emits CRLF (Git Bash on Windows).
+trace_fixture_files() {
+  local case_name="$1"
+  local fixture_dir="$2"
+  local python_bin="${3:-python3}"
+
+  "${python_bin}" tools/trace_replay/trace_cases.py \
+    --format fixture-files \
+    --case "${case_name}" \
+    --fixture-root "${fixture_dir}" | tr -d '\r'
+}

.github/workflows/apk.yml

@@ -11,6 +11,9 @@ on:
 jobs:
   build:
     runs-on: ubuntu-latest
+    permissions:
+      actions: write
+      contents: read
     env:
       CCACHE_BASEDIR: ${{ github.workspace }}
       CCACHE_COMPRESS: "true"
@@ -41,12 +44,11 @@ jobs:
           gradle-version: 8.10.2
 
       - name: Restore ccache
-        uses: actions/cache@v5
+        uses: actions/cache/restore@v5
         with:
           path: .ccache
-          key: ${{ runner.os }}-apk-${{ github.job }}-ccache-${{ github.ref_name }}-${{ github.run_id }}
+          key: ${{ runner.os }}-apk-${{ github.job }}-ccache-v1
           restore-keys: |
-            ${{ runner.os }}-apk-${{ github.job }}-ccache-${{ github.ref_name }}-
             ${{ runner.os }}-apk-${{ github.job }}-ccache-
 
       - name: Install ccache
@@ -125,6 +127,28 @@ jobs:
         if: always()
         run: ccache --show-stats
 
+      # Rewrite one rolling entry per job on the default branch. The upload stays
+      # cumulative - it carries every object restored at the top of this run plus
+      # the few TUs that actually changed - but Actions cache keys are immutable,
+      # so the superseded blob has to be released before the same key can be
+      # re-uploaded. Running after the build means a failed build leaves the
+      # existing entry untouched. The other trigger branches restore this entry
+      # rather than each writing a ~4 GB one of their own.
+      - name: Release superseded ccache entry
+        if: github.ref_name == github.event.repository.default_branch
+        env:
+          GH_TOKEN: ${{ github.token }}
+          CACHE_KEY: ${{ runner.os }}-apk-${{ github.job }}-ccache-v1
+        run: gh cache delete "${CACHE_KEY}" || true
+
+      - name: Save ccache
+        if: github.ref_name == github.event.repository.default_branch
+        continue-on-error: true
+        uses: actions/cache/save@v5
+        with:
+          path: .ccache
+          key: ${{ runner.os }}-apk-${{ github.job }}-ccache-v1
+
       - name: Verify APK metadata and packaging
         run: |
           AAPT2="$(find "$ANDROID_HOME/build-tools" -name aapt2 -type f | sort -V | tail -n 1)"
@@ -201,9 +225,41 @@ jobs:
       - name: Checkout repo
         uses: actions/checkout@v6
 
+      - name: Derive trace fixture cache key
+        id: fixture-key
+        run: bash .github/scripts/trace-fixture-cache.sh key '${{ matrix.case }}'
+
+      - name: Restore trace fixture cache
+        id: fixture-cache
+        if: steps.fixture-key.outputs.cacheable == 'true'
+        uses: actions/cache/restore@v5
+        with:
+          path: ${{ steps.fixture-key.outputs.paths }}
+          key: ${{ steps.fixture-key.outputs.key }}
+
+      - name: Verify restored trace fixture
+        id: fixture-verify
+        if: steps.fixture-cache.outputs.cache-hit == 'true'
+        run: |
+          if bash .github/scripts/trace-fixture-cache.sh verify '${{ matrix.case }}'; then
+            echo "ok=true" >> "$GITHUB_OUTPUT"
+          else
+            echo "ok=false" >> "$GITHUB_OUTPUT"
+            echo "::warning::Cached fixture for ${{ matrix.case }} failed verification; falling back to the download path"
+            bash .github/scripts/trace-fixture-cache.sh reset '${{ matrix.case }}'
+          fi
+
       - name: Fetch trace fixture
+        if: steps.fixture-verify.outputs.ok != 'true'
         run: bash .github/scripts/fetch-trace-fixture-lfs.sh '${{ matrix.case }}'
 
+      - name: Save trace fixture cache
+        if: steps.fixture-key.outputs.cacheable == 'true' && steps.fixture-cache.outputs.cache-hit != 'true'
+        uses: actions/cache/save@v5
+        with:
+          path: ${{ steps.fixture-key.outputs.paths }}
+          key: ${{ steps.fixture-key.outputs.key }}
+
       - name: Stage trace fixture
         run: |
           safe_case="$(printf '%s' '${{ matrix.case }}' | sed 's/[^A-Za-z0-9._-]/_/g')"
@@ -379,6 +435,9 @@ jobs:
           if [ "${{ matrix.backend.name }}" = "DirectGLES" ] && [ "${{ matrix.case.name }}" = "minecraft-1.21.4-fabric-iris-bliss-in-world" ]; then
             extra_retrace_args+=(--avoid-angle-llvmpipe-sampler-mipmap-min-filter)
           fi
+          if [ "${{ matrix.backend.name }}" = "DirectGLES" ] && [ "${{ matrix.case.avoid_angle_llvmpipe_explicit_lod_bias || false }}" = "true" ]; then
+            extra_retrace_args+=(--avoid-angle-llvmpipe-explicit-lod-bias)
+          fi
           if [ "${{ matrix.case.coherent_as_flush || false }}" = "true" ]; then
             extra_retrace_args+=(--coherent-as-flush)
           fi
@@ -411,6 +470,24 @@ jobs:
           run_retrace || retrace_status=$?
           if [ "${retrace_status}" -eq 75 ]; then
             echo "::warning::Android emulator infrastructure failed; restarting it and retrying this retrace once."
+            # Surface-lost is retried rather than failed, so it would otherwise
+            # be invisible. Report it per job - a healthy run prints nothing and
+            # a rate spike shows up as a row per affected case.
+            reason_file="android-retrace-result/infrastructure-failure-reason.txt"
+            surface_lost_retries=0
+            if [ -f "${reason_file}" ]; then
+              surface_lost_retries="$(grep -c 'angle-surface-lost' "${reason_file}" || true)"
+            fi
+            if [ "${surface_lost_retries}" -gt 0 ]; then
+              echo "surface-lost retries: ${surface_lost_retries} (${{ matrix.backend.name }}, ${{ matrix.case.name }})" \
+                >> "${GITHUB_STEP_SUMMARY}"
+            fi
+            # The restart truncates EMULATOR_LOG, and the attempt that lost the
+            # emulator is the one worth reading - the retry usually only shows
+            # the wreckage. Keep the first attempt's log before it is clobbered.
+            if [ -f "${EMULATOR_LOG}" ]; then
+              cp "${EMULATOR_LOG}" "${EMULATOR_LOG}.first-attempt" || true
+            fi
             sh android-plugin/run-avd-ci.sh stop \
               --avd-name "${AVD_NAME}" \
               --emulator-log "${EMULATOR_LOG}" \
@@ -450,6 +527,13 @@ jobs:
           if [ -f "${EMULATOR_LOG}" ]; then
             cp "${EMULATOR_LOG}" android-retrace-result/diagnostics/emulator.log
           fi
+          if [ -f "${EMULATOR_LOG}.first-attempt" ]; then
+            cp "${EMULATOR_LOG}.first-attempt" android-retrace-result/diagnostics/emulator-first-attempt.log
+          fi
+          # A vanished emulator looks identical whether the host OOM killer took
+          # qemu or the renderer faulted. These two say which.
+          free -h > android-retrace-result/diagnostics/host-memory.txt 2>&1 || true
+          sudo dmesg -T 2>/dev/null | tail -300 > android-retrace-result/diagnostics/host-dmesg.txt || true
 
       - name: Stop Emulator
         if: always()
@@ -531,22 +615,41 @@ jobs:
           )
 
           if ((${#failed_cases[@]})); then
-            echo "Retaining fixtures for failed retrace case(s):"
+            echo "Retaining fixtures and results for failed retrace case(s):"
             printf '  %s\n' "${!failed_cases[@]}"
           else
-            echo "All retrace jobs succeeded; no fixtures need to be retained."
+            echo "All retrace jobs succeeded; nothing needs to be retained."
           fi
 
           deleted=0
           retained=0
           while IFS=$'\t' read -r artifact_id artifact_name; do
+            keep=0
             if [[ "${artifact_name}" == MobileGL-trace-fixture-* ]]; then
               case_name="${artifact_name#MobileGL-trace-fixture-}"
               if [[ -v "failed_cases[${case_name}]" ]]; then
-                echo "Retaining ${artifact_name} (${artifact_id}) for failed retrace."
-                ((retained += 1))
-                continue
+                keep=1
               fi
+            elif [[ "${artifact_name}" == MobileGL-android-retrace-result-* ]]; then
+              # The result artifact carries mobilegl.log, retrace.log, logcat,
+              # the emulator log and the actual/diff images - the only record of
+              # why a retrace failed. Its name ends in -<backend>-<case>, so a
+              # suffix match on the case name keeps both backends' results for a
+              # case that failed on either of them, which is what a comparison
+              # needs. The match is anchored at the end, so a case name that is a
+              # prefix of a longer one does not retain the longer one's results.
+              for case_name in "${!failed_cases[@]}"; do
+                if [[ "${artifact_name}" == *-"${case_name}" ]]; then
+                  keep=1
+                  break
+                fi
+              done
+            fi
+
+            if ((keep)); then
+              echo "Retaining ${artifact_name} (${artifact_id}) for failed retrace."
+              ((retained += 1))
+              continue
             fi
 
             echo "Deleting ${artifact_name} (${artifact_id})"

.github/workflows/test.yml

@@ -1,4 +1,4 @@
-name: Test
+name: Test
 
 on:
   push:
@@ -11,6 +11,9 @@ on:
 jobs:
   build-linux:
     runs-on: ubuntu-latest
+    permissions:
+      actions: write
+      contents: read
     env:
       BUILD_DIR: build-linux
       CCACHE_BASEDIR: ${{ github.workspace }}
@@ -34,12 +37,11 @@ jobs:
         uses: lukka/get-cmake@v4.3.3
 
       - name: Restore ccache
-        uses: actions/cache@v5
+        uses: actions/cache/restore@v5
         with:
           path: .ccache
-          key: ${{ runner.os }}-test-${{ github.job }}-ccache-${{ github.ref_name }}-${{ github.run_id }}
+          key: ${{ runner.os }}-test-${{ github.job }}-ccache-v1
           restore-keys: |
-            ${{ runner.os }}-test-${{ github.job }}-ccache-${{ github.ref_name }}-
             ${{ runner.os }}-test-${{ github.job }}-ccache-
 
       - name: Prepare Vulkan SDK
@@ -83,6 +85,8 @@ jobs:
             -DMOBILEGL_LOG_ACTIVE_LEVEL=MOBILEGL_LOG_LEVEL_INFO \
             -DMOBILEGL_BUILD_TEST=ON \
             -DMOBILEGL_BUILD_BENCHMARK=ON \
+            -DMOBILEGL_BUILD_INTEGRATION_TEST=ON \
+            -DMOBILEGL_ITEST_VK_ICD=/usr/share/vulkan/icd.d/lvp_icd.json \
             -DMOBILEGL_BUILD_TRACE_REPLAY=OFF \
             -DBENCHMARK_DOWNLOAD_DEPENDENCIES=ON \
             -DBENCHMARK_ENABLE_TESTING=OFF \
@@ -95,6 +99,28 @@ jobs:
         if: always()
         run: ccache --show-stats
 
+      # Rewrite one rolling entry per job on the default branch. The upload stays
+      # cumulative - it carries every object restored at the top of this run plus
+      # the few TUs that actually changed - but Actions cache keys are immutable,
+      # so the superseded blob has to be released before the same key can be
+      # re-uploaded. Running after the build means a failed build leaves the
+      # existing entry untouched. The other trigger branches restore this entry
+      # rather than each writing one of their own.
+      - name: Release superseded ccache entry
+        if: github.ref_name == github.event.repository.default_branch
+        env:
+          GH_TOKEN: ${{ github.token }}
+          CACHE_KEY: ${{ runner.os }}-test-${{ github.job }}-ccache-v1
+        run: gh cache delete "${CACHE_KEY}" || true
+
+      - name: Save ccache
+        if: github.ref_name == github.event.repository.default_branch
+        continue-on-error: true
+        uses: actions/cache/save@v5
+        with:
+          path: .ccache
+          key: ${{ runner.os }}-test-${{ github.job }}-ccache-v1
+
       - name: Package Linux runtime
         run: |
           mkdir -p ci-artifacts
@@ -110,6 +136,7 @@ jobs:
             "${BUILD_DIR}/CTestTestfile.cmake" \
             "${BUILD_DIR}/MobileGL/MG_Test" \
             "${BUILD_DIR}/MobileGL/MG_Benchmark" \
+            "${BUILD_DIR}/MobileGL/MG_IntegrationTest" \
             "${SHARED_LIBS[@]}"
 
       - name: Upload Linux runtime
@@ -159,12 +186,102 @@ jobs:
       - name: Test
         working-directory: build-linux
         run: |
+          ulimit -c unlimited
+          sudo sysctl -w kernel.core_pattern='/tmp/core.%e.%p'
           if [ "${{ secrets.ACTIONS_STEP_DEBUG }}" = "true" ]; then
             ctest -V -L unit --no-tests=error
           else
             ctest --output-on-failure -L unit --no-tests=error
           fi
 
+      - name: Upload core dumps
+        if: failure()
+        uses: actions/upload-artifact@v7
+        with:
+          name: unit-core-dumps
+          path: /tmp/core.*
+          if-no-files-found: ignore
+
+  integration:
+    runs-on: ubuntu-latest
+    needs: build-linux
+
+    steps:
+      - name: Checkout repo
+        uses: actions/checkout@v6
+
+      - name: Get CMake
+        uses: lukka/get-cmake@v4.3.3
+
+      - name: Install runtime dependencies
+        # Same set as the benchmark job, for the same reason: the scenarios bring
+        # up real headless EGL (llvmpipe) and Vulkan (lavapipe) contexts, and
+        # libegl-mesa0 - the EGL vendor library behind glvnd's libegl1 dispatch -
+        # only arrives as a Recommends.
+        run: |
+          sudo apt-get update
+          sudo apt-get install -y libvulkan1 libegl1 libegl-mesa0 libgles2 libgl1-mesa-dri mesa-vulkan-drivers
+
+      - name: Download Linux runtime
+        uses: actions/download-artifact@v8
+        with:
+          name: mobilegl-linux-runtime
+          path: .
+
+      - name: Unpack Linux runtime
+        run: tar -xzf mobilegl-linux-runtime.tgz
+
+      - name: Normalize CTest command paths
+        run: |
+          python - <<'PY'
+          from pathlib import Path
+          import re
+
+          for path in Path('build-linux').rglob('CTestTestfile.cmake'):
+              text = path.read_text()
+              text = re.sub(r'"[^"]*/cmake-[^"]*/bin/cmake"', '"cmake"', text)
+              path.write_text(text)
+          PY
+
+      - name: Integration scenarios
+        working-directory: build-linux
+        # REQUIRE_GPU makes a driverless runner FAIL instead of skipping every
+        # scenario - an all-skip run is otherwise indistinguishable from a pass,
+        # which is how a five-month-old draw-dropping bug survived unseen until
+        # this lane existed.
+        #
+        # The lavapipe ICD pin lives in the build-linux configure
+        # (-DMOBILEGL_ITEST_VK_ICD), NOT here: the configure bakes it into each
+        # test's ctest ENVIRONMENT property, and a property entry OVERRIDES the
+        # job environment - a VK_ICD_FILENAMES exported here would be silently
+        # ignored while looking like it works. This lane runs on lavapipe
+        # deterministically, not on whichever of the eight Mesa ICDs a GPU-less
+        # runner enumerates first.
+        #
+        # Cores are armed so that any crash - the harness pre-flight child's
+        # included - leaves /tmp/core.*, which the failure-only step below ships
+        # as an artifact. Analyzing a downloaded core against the runtime
+        # artifact's binary in an ubuntu-24.04 userspace reproduces the exact
+        # crash stack without burning a CI round on an in-workflow debugger.
+        env:
+          MOBILEGL_ITEST_REQUIRE_GPU: "1"
+        run: |
+          ulimit -c unlimited
+          sudo sysctl -w kernel.core_pattern='/tmp/core.%e.%p'
+          if [ "${{ secrets.ACTIONS_STEP_DEBUG }}" = "true" ]; then
+            ctest -V -L integration-gpu --no-tests=error
+          else
+            ctest --output-on-failure -L integration-gpu --no-tests=error
+          fi
+
+      - name: Upload core dumps
+        if: failure()
+        uses: actions/upload-artifact@v7
+        with:
+          name: integration-core-dumps
+          path: /tmp/core.*
+          if-no-files-found: ignore
+
   benchmark:
     runs-on: ubuntu-latest
     needs: build-linux
@@ -208,7 +325,18 @@ jobs:
 
       - name: Benchmark
         working-directory: build-linux
-        run: ctest -V -C Release -L benchmark --no-tests=error
+        run: |
+          ulimit -c unlimited
+          sudo sysctl -w kernel.core_pattern='/tmp/core.%e.%p'
+          ctest -V -C Release -L benchmark --no-tests=error
+
+      - name: Upload core dumps
+        if: failure()
+        uses: actions/upload-artifact@v7
+        with:
+          name: benchmark-core-dumps
+          path: /tmp/core.*
+          if-no-files-found: ignore
 
   build-retrace:
     runs-on: ubuntu-latest
@@ -216,6 +344,9 @@ jobs:
       - build-linux
       - test
       - benchmark
+    permissions:
+      actions: write
+      contents: read
     env:
       BUILD_DIR: build-retrace
       CCACHE_BASEDIR: ${{ github.workspace }}
@@ -240,12 +371,11 @@ jobs:
         uses: lukka/get-cmake@v4.3.3
 
       - name: Restore ccache
-        uses: actions/cache@v5
+        uses: actions/cache/restore@v5
         with:
           path: .ccache
-          key: ${{ runner.os }}-test-${{ github.job }}-ccache-${{ github.ref_name }}-${{ github.run_id }}
+          key: ${{ runner.os }}-test-${{ github.job }}-ccache-v1
           restore-keys: |
-            ${{ runner.os }}-test-${{ github.job }}-ccache-${{ github.ref_name }}-
             ${{ runner.os }}-test-${{ github.job }}-ccache-
 
       - name: Prepare Vulkan SDK
@@ -311,6 +441,21 @@ jobs:
         if: always()
         run: ccache --show-stats
 
+      - name: Release superseded ccache entry
+        if: github.ref_name == github.event.repository.default_branch
+        env:
+          GH_TOKEN: ${{ github.token }}
+          CACHE_KEY: ${{ runner.os }}-test-${{ github.job }}-ccache-v1
+        run: gh cache delete "${CACHE_KEY}" || true
+
+      - name: Save ccache
+        if: github.ref_name == github.event.repository.default_branch
+        continue-on-error: true
+        uses: actions/cache/save@v5
+        with:
+          path: .ccache
+          key: ${{ runner.os }}-test-${{ github.job }}-ccache-v1
+
       - name: Normalize CTest command paths
         run: |
           python - <<'PY'
@@ -366,9 +511,41 @@ jobs:
       - name: Checkout repo
         uses: actions/checkout@v6
 
+      - name: Derive trace fixture cache key
+        id: fixture-key
+        run: bash .github/scripts/trace-fixture-cache.sh key '${{ matrix.case }}'
+
+      - name: Restore trace fixture cache
+        id: fixture-cache
+        if: steps.fixture-key.outputs.cacheable == 'true'
+        uses: actions/cache/restore@v5
+        with:
+          path: ${{ steps.fixture-key.outputs.paths }}
+          key: ${{ steps.fixture-key.outputs.key }}
+
+      - name: Verify restored trace fixture
+        id: fixture-verify
+        if: steps.fixture-cache.outputs.cache-hit == 'true'
+        run: |
+          if bash .github/scripts/trace-fixture-cache.sh verify '${{ matrix.case }}'; then
+            echo "ok=true" >> "$GITHUB_OUTPUT"
+          else
+            echo "ok=false" >> "$GITHUB_OUTPUT"
+            echo "::warning::Cached fixture for ${{ matrix.case }} failed verification; falling back to the download path"
+            bash .github/scripts/trace-fixture-cache.sh reset '${{ matrix.case }}'
+          fi
+
       - name: Fetch trace fixture
+        if: steps.fixture-verify.outputs.ok != 'true'
         run: bash .github/scripts/fetch-trace-fixture-lfs.sh '${{ matrix.case }}'
 
+      - name: Save trace fixture cache
+        if: steps.fixture-key.outputs.cacheable == 'true' && steps.fixture-cache.outputs.cache-hit != 'true'
+        uses: actions/cache/save@v5
+        with:
+          path: ${{ steps.fixture-key.outputs.paths }}
+          key: ${{ steps.fixture-key.outputs.key }}
+
       - name: Stage trace fixture
         run: |
           safe_case="$(printf '%s' '${{ matrix.case }}' | sed 's/[^A-Za-z0-9._-]/_/g')"
@@ -456,6 +633,8 @@ jobs:
       - name: Retrace and validate
         working-directory: build-retrace/tools/trace_replay
         run: |
+          ulimit -c unlimited
+          sudo sysctl -w kernel.core_pattern='/tmp/core.%e.%p'
           if [ '${{ matrix.backend }}' = 'DirectVulkan' ]; then
             export MOBILEGL_MAGMA_R11G11B10F_FALLBACK=1
           fi
@@ -470,6 +649,14 @@ jobs:
           fi
           ctest -V --no-tests=error -R '^MobileGLTraceReplay\.${{ matrix.case }}\.${{ matrix.backend }}$'
 
+      - name: Upload core dumps
+        if: failure()
+        uses: actions/upload-artifact@v7
+        with:
+          name: retrace-core-dumps-${{ matrix.backend }}-${{ matrix.case }}
+          path: /tmp/core.*
+          if-no-files-found: ignore
+
       - name: Upload actual image
         if: always()
         uses: actions/upload-artifact@v7

.gitmodules

@@ -7,9 +7,6 @@
 [submodule "3rdparty/SPIRV-Cross"]
 	path = 3rdparty/SPIRV-Cross
 	url = https://github.com/KhronosGroup/SPIRV-Cross.git
-[submodule "include/FastSTL"]
-	path = include/FastSTL
-	url = https://github.com/MobileGL-Dev/FastSTL.git
 [submodule "3rdparty/tracy"]
 	path = 3rdparty/tracy
 	url = https://github.com/wolfpld/tracy.git
@@ -34,6 +31,6 @@
 [submodule "3rdparty/asio"]
 	path = 3rdparty/asio
 	url = https://github.com/chriskohlhoff/asio.git
-[submodule "3rdparty/libfork"]
-	path = 3rdparty/libfork
-	url = https://github.com/ConorWilliams/libfork.git
+[submodule "include/ska"]
+	path = include/ska
+	url = https://github.com/MobileGL-Dev/flat_hash_map.git

CMakeLists.txt

@@ -20,6 +20,81 @@ set(MOBILEGL_VULKAN_LIBRARY "" CACHE FILEPATH "Vulkan loader/MoltenVK library to
 if (ANDROID)
     set(MOBILEGL_BUILD_TEST       OFF CACHE BOOL "Build MobileGL tests"      FORCE)
     set(MOBILEGL_BUILD_BENCHMARK  OFF CACHE BOOL "Build MobileGL benchmarks" FORCE)
+
+    # ------- Android API level policy: minimum 26, decided here and only here -------
+    # MobileGL ships against API 26: the codebase must not use any API introduced
+    # after 26. That usage constraint is enforced where it is real - the shipping
+    # gradle build compiles at minSdk 26, where a newer API is simply undeclared
+    # and fails to compile. Configuring at a HIGHER level is therefore allowed
+    # (nothing in the tree may rely on it), but a LOWER level would change the
+    # libc contract underneath the shipped library and is refused.
+    #
+    # This has to live at configure time because the level cannot be corrected
+    # from a source header. A `#define __ANDROID_API__ 26` in a common header
+    # only rewrites the macro for the bionic headers that happen to be included
+    # after it; any libc++ header pulled in earlier has already latched its
+    # feature macros at the real configure-time level. libc++ and bionic then
+    # disagree about which symbols exist - libc++ calls e.g.
+    # pthread_cond_clockwait while bionic, re-read at the lowered level, has
+    # hidden its declaration. MobileGL/Defines.h carried exactly that pin from
+    # the first commit until it was removed; this guard is what replaces it.
+    #
+    # Read the level back from the compiler target triple first. Its trailing
+    # number (aarch64-none-linux-android26) is precisely what clang turns into
+    # __ANDROID_API__, so it cannot disagree with the compile itself, and it is
+    # already past every NDK normalisation step - codename aliases, "latest",
+    # and per-ABI minimum pull-ups. ANDROID_PLATFORM_LEVEL is the fallback for
+    # generators/languages where the triple variable is not populated.
+    #
+    # Note CMAKE_SYSTEM_VERSION is deliberately NOT consulted: it holds the API
+    # level only under the NDK's newer toolchain path, and is a meaningless 1
+    # when ANDROID_USE_LEGACY_TOOLCHAIN_FILE is on (which is what AGP has been
+    # defaulting to). Reading it would fail every legacy-mode build.
+    set(MOBILEGL_ANDROID_API_LEVEL 26)
+
+    set(_mobilegl_android_api "")
+    foreach (_mobilegl_api_triple "${CMAKE_CXX_COMPILER_TARGET}"
+                                  "${CMAKE_C_COMPILER_TARGET}")
+        if (NOT _mobilegl_android_api AND
+            _mobilegl_api_triple MATCHES "-android([0-9]+)$")
+            set(_mobilegl_android_api "${CMAKE_MATCH_1}")
+        endif()
+    endforeach()
+
+    foreach (_mobilegl_api_var ANDROID_PLATFORM_LEVEL ANDROID_NATIVE_API_LEVEL
+                               ANDROID_PLATFORM)
+        if (NOT _mobilegl_android_api AND ${_mobilegl_api_var})
+            string(REGEX REPLACE "^android-" ""
+                   _mobilegl_android_api "${${_mobilegl_api_var}}")
+        endif()
+    endforeach()
+
+    if (NOT _mobilegl_android_api MATCHES "^[0-9]+$")
+        message(FATAL_ERROR
+            "MobileGL: could not determine the Android API level (got "
+            "\"${_mobilegl_android_api}\"). Configure with the NDK toolchain "
+            "file and -DANDROID_PLATFORM=android-${MOBILEGL_ANDROID_API_LEVEL}.")
+    elseif (_mobilegl_android_api LESS MOBILEGL_ANDROID_API_LEVEL)
+        message(FATAL_ERROR
+            "MobileGL requires at least Android API ${MOBILEGL_ANDROID_API_LEVEL}, "
+            "but this build resolved to API ${_mobilegl_android_api}.\n"
+            "Configure with -DANDROID_PLATFORM=android-${MOBILEGL_ANDROID_API_LEVEL} "
+            "(gradle builds get this from minSdk ${MOBILEGL_ANDROID_API_LEVEL}, so "
+            "check that minSdk instead of adding an override).")
+    elseif (_mobilegl_android_api GREATER MOBILEGL_ANDROID_API_LEVEL)
+        message(STATUS
+            "MobileGL: configuring at Android API ${_mobilegl_android_api} "
+            "(> shipping minimum ${MOBILEGL_ANDROID_API_LEVEL}). Allowed, but the "
+            "tree must not use post-${MOBILEGL_ANDROID_API_LEVEL} APIs - the "
+            "minSdk-${MOBILEGL_ANDROID_API_LEVEL} gradle build is the enforcing "
+            "compile.")
+    endif()
+
+    message(STATUS "MobileGL: Android API level ${_mobilegl_android_api}")
+
+    unset(_mobilegl_android_api)
+    unset(_mobilegl_api_var)
+    unset(_mobilegl_api_triple)
 endif()
 
 option(MOBILEGL_ENABLE_LTO "Build with ThinLTO/IPO" OFF)
@@ -205,9 +280,12 @@ set(SOURCE_FILES
     MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/PackDoubleVertexInputsPass.cpp
     MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/RebaseInstanceIndexPass.cpp
     MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/NormalizeRectCoordinatesPass.cpp
+    MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/PrivateToEntryLocalPass.cpp
+    MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/StripUniformLocationsPass.cpp
     MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/StripUboMemberRelaxedPrecisionPass.cpp
     MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/StripNoPerspectivePass.cpp
     MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/EmulateNoPerspectivePass.cpp
+    MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/LegalizeFragmentOutputIndexPass.cpp
 
     MobileGL/MG_Util/BackendLoaders/OpenGL/Loader.cpp
     MobileGL/MG_Util/BackendLoaders/Vulkan/Loader.cpp
@@ -296,6 +374,7 @@ set(SOURCE_FILES
     MobileGL/MG_State/GLState/TextureState/TextureState.cpp
     MobileGL/MG_State/GLState/ProgramState/ProgramObject.cpp
     MobileGL/MG_State/GLState/ProgramState/ProgramLinkTask.cpp
+    MobileGL/MG_State/GLState/ProgramState/ProgramSpirvTask.cpp
     MobileGL/MG_State/GLState/ProgramState/ShaderCompileTask.cpp
     MobileGL/MG_State/GLState/ProgramState/ShaderObject.cpp
     MobileGL/MG_State/GLState/ProgramState/ShaderPreprocessCache.cpp
@@ -373,13 +452,6 @@ set(MOBILEGL_INCLUDE_DIR
         # MG_Util/Async/ShaderCompilePool.cpp includes it, and it stays behind that file's
         # pimpl so no consumer target needs this path.
         ${CMAKE_SOURCE_DIR}/3rdparty/asio/asio/include
-        # The second shader-compile execution engine (MOBILEGL_ASYNC_POOL=libfork), on the
-        # same terms as Asio above: header-only, no add_subdirectory (its CMakeLists only
-        # declares an INTERFACE target plus install/test scaffolding we do not want), no link
-        # target, and reachable from exactly one translation unit. libfork's own
-        # target_compile_features asks for cxx_std_23, which this project already sets
-        # globally, so its C++20 coroutines need no per-source standard override.
-        ${CMAKE_SOURCE_DIR}/3rdparty/libfork/include
 )
 
 add_library(${CMAKE_PROJECT_NAME} SHARED

MobileGL/Config.h

@@ -66,12 +66,11 @@ namespace MobileGL::MG_Config {
     //   - DISPLAY: X11 session variable, not MobileGL configuration.
     //   - MOBILEGL_LOG_FILE_PATH: log-file init runs before MG_ConfigLoader::Init
     //     (see MG_Util/Debug/Log.cpp).
-    //   - MOBILEGL_ASYNC_POOL: a ShaderCompilePool is constructed by binaries that never call
-    //     MobileGL::Initialize() and so never run MG_ConfigLoader::Init - MG_Test's
-    //     JobNodeTest builds pools directly, and it is the suite that runs the whole async
-    //     matrix against both execution engines. Mirroring it here would resolve to the
-    //     default in exactly the tests that exist to tell the engines apart (see
-    //     MG_Util/Async/ShaderCompilePool.cpp, DetectAsyncPoolEngine).
+    //   - MOBILEGL_VALIDATE_SPIRV: test suites like SpirvPassTest exercise
+    //     ShaderCompiler without ever running MobileGL::Initialize(), and every
+    //     Initialize() re-runs MG_ConfigLoader::Init, which would clobber a
+    //     programmatic override stored here (see ShaderCompiler.cpp,
+    //     SpirvValidationEnabled).
     struct FeaturesTable {
         // MOBILEGL_DISABLE_TIMERQUERY: do not advertise or use GPU timer queries.
         Bool DisableTimerQuery = false;
@@ -90,6 +89,13 @@ namespace MobileGL::MG_Config {
         // MOBILEGL_AVOID_SAMPLER_MIPMAP_MIN_FILTER: avoid mipmap min filters in samplers,
         // resolves certain rendering bugs on ANGLE + llvmpipe.
         Bool AvoidSamplerMipmapMinFilter = false;
+        // MOBILEGL_AVOID_EXPLICIT_LOD_BIAS: leave an already-explicit LOD argument alone when
+        // emulating GL_TEXTURE_LOD_BIAS, instead of adding the bias uniform to it. Injecting
+        // the uniform turns a compile-time-constant LOD into a runtime expression, which
+        // sends ANGLE + llvmpipe down a mip-selection path that dereferences a NULL
+        // descriptor and kills the process. Deviates from spec (Vulkan adds the bias to
+        // OpImageSampleExplicitLod), so it is an avoidance for that stack only.
+        Bool AvoidExplicitLodBias = false;
         // MOBILEGL_COHERENT_AS_FLUSH: app-compat for engines (e.g. Flywheel) that write
         // GPU-read data through persistent GL_MAP_FLUSH_EXPLICIT_BIT maps they never
         // flush. Persistent FLUSH_EXPLICIT map requests are rewritten to coherent
@@ -142,6 +148,19 @@ namespace MobileGL::MG_Config {
         // MOBILEGL_ASYNC_SHADER_COMPILE_THREADS: shader-compile worker count. 0 (unset) means
         // auto, which is min(4, big cores); an explicit value is honoured as given.
         Uint32 AsyncShaderCompileThreads = 0;
+        // MOBILEGL_ASYNC_OPTIMISTIC_SHADER_STATUS: while a compile job is still in flight,
+        // glGetShaderiv(GL_COMPILE_STATUS) answers GL_TRUE and the shader info log reads
+        // empty, WITHOUT joining the job (latched per compile - see
+        // ShaderObject::TakeOptimisticCompileAnswer). A deliberate, bounded spec violation:
+        // a real failure still fails the program link with the compile log quoted. It
+        // exists for applications that compile hundreds of shaders serially and read the
+        // status right after each glCompileShader - Iris's shader-pack load - where those
+        // per-shader joins are what serializes the batch on its main path (Iris's gbuffer
+        // phase issues no program-level query between programs; program-level LINK_STATUS
+        // and the program info log still join truthfully, so paths that check each link
+        // immediately stay serial by their own construction). Off by default; never
+        // advertise it.
+        QuirkOverride AsyncOptimisticShaderStatus = QuirkOverride::Auto;
     };
     extern FeaturesTable Features;
 } // namespace MobileGL::MG_Config

MobileGL/ConfigLoader.cpp

@@ -171,6 +171,7 @@ namespace MobileGL::MG_ConfigLoader {
         features.MagmaFramesInFlight = QueryEnvUint32("MOBILEGL_MAGMA_FRAMESINFLIGHT", 3, 1, 64);
         features.AvoidSamplerMipmapMinFilter =
             QueryEnvFlag("MOBILEGL_AVOID_SAMPLER_MIPMAP_MIN_FILTER");
+        features.AvoidExplicitLodBias = QueryEnvFlag("MOBILEGL_AVOID_EXPLICIT_LOD_BIAS");
         features.CoherentAsFlush = QueryEnvFlag("MOBILEGL_COHERENT_AS_FLUSH");
         features.TraceSkipAutodestroy = QueryEnvFlag("MOBILEGL_TRACE_SKIP_AUTODESTROY");
         features.DisableUboRing = QueryEnvFlag("MOBILEGL_DISABLE_UBO_RING");
@@ -183,6 +184,8 @@ namespace MobileGL::MG_ConfigLoader {
         features.EsprytMultiDrawMode = QueryEnvGLESMultiDrawMode("MOBILEGL_ESPRYT_MULTIDRAW_MODE");
         features.AsyncShaderCompile = QueryEnvQuirkOverride("MOBILEGL_ASYNC_SHADER_COMPILE");
         features.AsyncShaderCompileThreads = QueryEnvUint32("MOBILEGL_ASYNC_SHADER_COMPILE_THREADS", 0, 0, 64);
+        features.AsyncOptimisticShaderStatus =
+            QueryEnvQuirkOverride("MOBILEGL_ASYNC_OPTIMISTIC_SHADER_STATUS");
     }
 
     inline void InitBackendType() {

MobileGL/Defines.h

@@ -9,10 +9,20 @@
 #pragma once
 
 // ============== Platform-specific definitions and macros ============== //
-#ifdef __ANDROID__
-#undef __ANDROID_API__
-#define __ANDROID_API__ 26 // force Android API level to 26 for compatibility
-#endif
+// No __ANDROID_API__ pin here on purpose. The effective API level is owned by
+// the build system (gradle minSdk 26 -> -DANDROID_PLATFORM=android-26, enforced
+// by the configure-time guard in CMakeLists.txt), not by a macro.
+//
+// History: this used to `#define __ANDROID_API__ 26` to *raise* the level back
+// when the build configured something lower, so that pthread_getname_np (which
+// bionic guards with __INTRODUCED_IN(26)) would be declared. Once a later
+// change added an `#undef` in front of it, the same line started *lowering* the
+// level whenever the build configured higher than 26 - and that is an
+// include-order split-brain, not a compatibility knob: a TU that includes any
+// libc++ header before Includes.h latches libc++'s feature macros at the
+// configure-time level, and only the bionic headers pulled in afterwards see
+// the lowered value. The two halves then disagree (e.g. libc++ believes
+// pthread_cond_clockwait exists while bionic has since hidden its declaration).
 
 #ifdef _WIN32
 #ifndef NOMINMAX
@@ -37,6 +47,19 @@
 #define MOBILEGL_WGL_API MOBILEGL_API
 
 // ====================== MobileGL configurations ======================= //
+// The numeric log levels live here, not only in Log.h: MOBILEGL_ASSERT below compares
+// MOBILEGL_LOG_ACTIVE_LEVEL against MOBILEGL_LOG_LEVEL_DEBUG, and in a translation unit
+// that includes Defines.h without Log.h both tokens would silently evaluate to 0 in the
+// preprocessor conditional - enabling the assert in exactly the INFO-level builds it is
+// documented to be compiled out of. Log.h redefines them identically, which is legal.
+#ifndef MOBILEGL_LOG_LEVEL_DEBUG
+#define MOBILEGL_LOG_LEVEL_DEBUG 0
+#define MOBILEGL_LOG_LEVEL_WARN 1
+#define MOBILEGL_LOG_LEVEL_ERROR 2
+#define MOBILEGL_LOG_LEVEL_INFO 3
+#define MOBILEGL_LOG_LEVEL_FATAL 4
+#endif
+
 #ifndef MOBILEGL_LOG_ACTIVE_LEVEL
 #define MOBILEGL_LOG_ACTIVE_LEVEL MOBILEGL_LOG_LEVEL_INFO
 #endif

MobileGL/Includes.h

@@ -49,8 +49,8 @@
 #include <stacktrace>
 #endif
 
-// Include FastSTL
-#include <FastSTL/UnorderedMap.h>
+// Include ska::flat_hash_map
+#include <ska/flat_hash_map.hpp>
 
 // Include xxHash
 #include <xxhash.h>

MobileGL/MG_Backend/DirectGLES/BackendObject_DirectGLES.cpp

@@ -1151,6 +1151,16 @@ namespace MobileGL::MG_Backend::DirectGLES {
         m_dynamicParameters.MaxComputeUniformBlocks = m_GLESCapabilities.MaxComputeUniformBlocks;
         m_dynamicParameters.MaxComputeWorkGroupInvocations = m_GLESCapabilities.MaxComputeWorkGroupInvocations;
         m_dynamicParameters.MaxShaderStorageBufferBindings = m_GLESCapabilities.MaxShaderStorageBufferBindings;
+        // This is the number glGetIntegerv(GL_MAX_TEXTURE_BUFFER_SIZE) hands the application, and
+        // on a host without buffer textures it is knowingly a floor MobileGL cannot honour rather
+        // than a driver answer (m_GLESCapabilities.MaxTextureBufferSizeIsDriverReported says
+        // which). Reporting 0 instead was considered and rejected: MobileGL advertises an OpenGL
+        // 4.x context, where buffer textures are core and the limit has a spec minimum of 65536,
+        // so 0 is not a legal answer and applications are not written to survive it. GL offers no
+        // way to say "this core feature is missing", so the honesty is carried outside the limit:
+        // FillInGLESCapabilities logs the tier, glTexBuffer and the program build each name the
+        // missing capability at MGLOG_I, and the driver POST carries a "Buffer textures" row that
+        // FAILs on this tier.
         m_dynamicParameters.MaxTextureBufferSize = m_GLESCapabilities.MaxTextureBufferSize;
         m_dynamicParameters.TextureBufferOffsetAlignment = m_GLESCapabilities.TextureBufferOffsetAlignment;
         m_dynamicParameters.MaxUniformBufferBindings = m_GLESCapabilities.MaxUniformBufferBindings;

MobileGL/MG_Backend/DirectGLES/DirectGLES.cpp

@@ -605,10 +605,11 @@ namespace MobileGL::MG_Backend::DirectGLES {
             // Cached address of g_xfbObjects[g_currentXfbName]: PrepareForDraw consults
             // CurrentXfb on EVERY draw (StartPendingTransformFeedback) and the map
             // lookup was pure per-draw overhead for the overwhelmingly common no-capture
-            // case. FastSTL's open addressing keeps values in the bucket array, so ANY
-            // insert can rehash and move them (and erase/clear can too): every site that
-            // mutates the map or rebinds the current name resets this to null instead of
-            // reasoning about stability, and CurrentXfb re-resolves lazily.
+            // case. Open addressing keeps values in the bucket array, so ANY insert can
+            // rehash and move them - and erase moves them too, by shifting the rest of the
+            // probe cluster into the hole, which reaches entries other than the erased one.
+            // Every site that mutates the map or rebinds the current name resets this to
+            // null instead of reasoning about stability, and CurrentXfb re-resolves lazily.
             XfbObjectState* g_currentXfbState = nullptr;
 
             XfbObjectState& CurrentXfb() {
@@ -883,7 +884,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
             if (it->second.esId != 0 && g_GLESFuncs.glDeleteTransformFeedbacks != nullptr) {
                 g_GLESFuncs.glDeleteTransformFeedbacks(1, &it->second.esId);
             }
-            g_currentXfbState = nullptr; // erase can move values (open addressing)
+            g_currentXfbState = nullptr; // erase shifts the probe cluster, moving other entries
             g_xfbObjects.erase(it);
             // The frontend reverts to the default object when the bound one is deleted.
             if (g_currentXfbName == name) {
@@ -1214,7 +1215,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
             if (g_unitTextureSyncListValid &&
                 g_unitTextureSyncListContextId == keys.contextId &&
                 g_unitTextureSyncListMaxUnit == maxTouchedUnit &&
-                g_unitTextureSyncListContextGeneration == g_textureContextGeneration &&
+                g_unitTextureSyncListContextGeneration == g_backendContextGeneration &&
                 g_unitTextureSyncListEpoch == unitBindingsEpoch &&
                 g_unitTextureSyncListSamplingGeneration == samplingGeneration &&
                 PairingsIntact(g_unitTextureSyncList)) {
@@ -1246,7 +1247,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
                 }
                 g_unitTextureSyncListContextId = keys.contextId;
                 g_unitTextureSyncListMaxUnit = maxTouchedUnit;
-                g_unitTextureSyncListContextGeneration = g_textureContextGeneration;
+                g_unitTextureSyncListContextGeneration = g_backendContextGeneration;
                 g_unitTextureSyncListEpoch = unitBindingsEpoch;
                 g_unitTextureSyncListSamplingGeneration = samplingGeneration;
                 g_unitTextureSyncListValid = true;
@@ -1274,7 +1275,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
                     g_fboTextureSyncListSlotVersion == fboSlotVersion &&
                     g_fboTextureSyncListObjectVersion == fboObjectVersion &&
                     g_fboTextureSyncListContextId == keys.contextId &&
-                    g_fboTextureSyncListContextGeneration == g_textureContextGeneration &&
+                    g_fboTextureSyncListContextGeneration == g_backendContextGeneration &&
                     PairingsIntact(g_fboTextureSyncList);
                 if (fboListValid) {
                     for (const auto& entry : g_fboTextureSyncList) {
@@ -1302,7 +1303,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
                     g_fboTextureSyncListSlotVersion = fboSlotVersion;
                     g_fboTextureSyncListObjectVersion = fboObjectVersion;
                     g_fboTextureSyncListContextId = keys.contextId;
-                    g_fboTextureSyncListContextGeneration = g_textureContextGeneration;
+                    g_fboTextureSyncListContextGeneration = g_backendContextGeneration;
                 }
             } else {
                 g_fboTextureSyncListFbo = nullptr;
@@ -2028,7 +2029,9 @@ namespace MobileGL::MG_Backend::DirectGLES {
             g_currentDrawFrontendProgram = nullptr;
             g_currentDrawBackendProgram = nullptr;
 
-            if (!currentProgram || !currentProgram->GetLinkStatus()) {
+            // ... || !GetSpirvStatus(): see BackendProgramObjectImpl::SyncToBackend - a
+            // program whose SPIR-V never arrived is linked but not drawable.
+            if (!currentProgram || !currentProgram->GetLinkStatus() || !currentProgram->GetSpirvStatus()) {
                 g_GLESFuncs.glUseProgram(0);
                 g_lastUsedBackendProgramId = 0;
                 return;
@@ -2075,11 +2078,30 @@ namespace MobileGL::MG_Backend::DirectGLES {
             // A link-version mismatch means the program was relinked: the backend
             // shaders and every cache built by CacheResourceLocations (block
             // indices, sampler locations, UBO upload gate) are stale.
+            //
+            // The storage-block signature is the same shape of condition: ES cannot move a
+            // storage block's binding after link, so glShaderStorageBlockBinding is honoured by
+            // baking the effective binding into the generated ESSL - which makes a program built
+            // against a different override set stale. It is compared HERE rather than acted on in
+            // the entry point because that one must never trigger a build (see
+            // ShaderStorageBlockBinding below). The signature is over the values, so an
+            // application that re-sets the same bindings every frame rebuilds nothing.
+            //
+            // The image-unit generation is a third of the same shape, and it used to be
+            // carried by accident: glUniform1i on an image uniform bumped the program's backend
+            // state version, which was in the program-pipeline composite's cache key, so a
+            // pipeline draw got a whole NEW composite object and therefore a fresh twin. Keying
+            // that cache on the link version instead (ProgramPipelineObject) removed the
+            // accident - and it never covered the monolithic glUseProgram path at all - so the
+            // dependency is stated here instead.
             if (!twin->GetBackendProgramId() ||
                 twin->GetSyncedLinkVersion() != currentProgram->GetLinkVersion() ||
+                twin->GetSyncedImageUnitVersion() != currentProgram->GetImageUnitVersion() ||
                 twin->GetSnormFallbackClampOutputMask() != g_snormFallbackClampOutputMask ||
                 twin->GetUnormFallbackClampOutputMask() != g_unormFallbackClampOutputMask ||
-                twin->GetFragColorBroadcastCount() != g_fragColorBroadcastCount) {
+                twin->GetFragColorBroadcastCount() != g_fragColorBroadcastCount ||
+                twin->GetShaderStorageBlockBindingSignature() !=
+                    ComputeShaderStorageBlockBindingSignature(*currentProgram)) {
                 twin->SyncToBackend(currentProgram);
             }
             g_currentDrawFrontendProgram = currentProgram.get();
@@ -2421,7 +2443,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
             static_cast<SizeT>(maxTouchedUnit + 1) * sizeof(SamplerImpl::g_boundSamplersCache[0]);
         if (g_unitSamplerWalkValid && g_unitSamplerWalkContextId == keys.contextId &&
             g_unitSamplerWalkEpoch == keys.unitBindingsEpoch && g_unitSamplerWalkMaxUnit == maxTouchedUnit &&
-            g_unitSamplerWalkContextGeneration == TextureImpl::g_textureContextGeneration &&
+            g_unitSamplerWalkContextGeneration == g_backendContextGeneration &&
             std::memcmp(g_unitSamplerWalkRows.data(), SamplerImpl::g_boundSamplersCache.data(), rowBytes) == 0) {
             return;
         }
@@ -2442,7 +2464,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
         g_unitSamplerWalkContextId = keys.contextId;
         g_unitSamplerWalkEpoch = keys.unitBindingsEpoch;
         g_unitSamplerWalkMaxUnit = maxTouchedUnit;
-        g_unitSamplerWalkContextGeneration = TextureImpl::g_textureContextGeneration;
+        g_unitSamplerWalkContextGeneration = g_backendContextGeneration;
         std::memcpy(g_unitSamplerWalkRows.data(), SamplerImpl::g_boundSamplersCache.data(), rowBytes);
         g_unitSamplerWalkValid = true;
     }
@@ -2552,7 +2574,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
                                memo.programBackendStateVersion ==
                                    (currentProgram ? currentProgram->GetBackendStateVersion() : 0) &&
                                memo.programLinked == (currentProgram && currentProgram->GetLinkStatus()) &&
-                               memo.contextGeneration == TextureImpl::g_textureContextGeneration;
+                               memo.contextGeneration == g_backendContextGeneration;
         // Short-circuited: the shadow compare is only meaningful once the key (and with it the
         // snapshotted row count) matches.
         if (!keysMatch || std::memcmp(memo.boundTextures.data(), TextureImpl::g_boundTexturesCache.data(),
@@ -2567,7 +2589,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
                 memo.programLifetimeId = currentProgram ? currentProgram->GetLifetimeId() : 0;
                 memo.programBackendStateVersion = currentProgram ? currentProgram->GetBackendStateVersion() : 0;
                 memo.programLinked = currentProgram && currentProgram->GetLinkStatus();
-                memo.contextGeneration = TextureImpl::g_textureContextGeneration;
+                memo.contextGeneration = g_backendContextGeneration;
                 std::memcpy(memo.boundTextures.data(), TextureImpl::g_boundTexturesCache.data(), shadowBytes);
                 memo.valid = true;
             }
@@ -2589,7 +2611,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
     static void BindCurrentProgramWithResources(
         const SharedPtr<MG_State::GLState::ProgramObject>& currentProgram,
         const TextureImpl::DrawTextureSyncKeys& keys) {
-        if (currentProgram && currentProgram->GetLinkStatus()) {
+        if (currentProgram && currentProgram->GetLinkStatus() && currentProgram->GetSpirvStatus()) {
 #ifdef TRACY_ENABLE
             ZoneScopedNC("BindCurrentProgram", TRACY_ZONECOLOR_BACKEND);
 #endif
@@ -2742,7 +2764,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
                         samplerPassMemo.unitBindingsEpoch == keys.unitBindingsEpoch &&
                         samplerPassMemo.samplingGeneration == keys.samplingGeneration &&
                         samplerPassMemo.backendStateVersion == programBackendStateVersion &&
-                        samplerPassMemo.textureContextGeneration == TextureImpl::g_textureContextGeneration;
+                        samplerPassMemo.textureContextGeneration == g_backendContextGeneration;
                     if (samplerPassClean) {
                         for (Uint i = 0; i < samplerPassMemo.count; ++i) {
                             if (SamplerImpl::g_boundSamplersCache[samplerPassMemo.units[i]] !=
@@ -2841,7 +2863,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
                             samplerPassMemo.unitBindingsEpoch = keys.unitBindingsEpoch;
                             samplerPassMemo.samplingGeneration = keys.samplingGeneration;
                             samplerPassMemo.backendStateVersion = programBackendStateVersion;
-                            samplerPassMemo.textureContextGeneration = TextureImpl::g_textureContextGeneration;
+                            samplerPassMemo.textureContextGeneration = g_backendContextGeneration;
                             samplerPassMemo.valid = true;
                         }
                     }
@@ -2859,7 +2881,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
     // is pinned for the duration. Prefers the per-draw stash those preparations wrote.
     static PrgramImpl::BackendProgramObjectImpl* GetCurrentBackendProgram() {
         const auto& currentProgram = MG_State::pGLContext->GetProgramForDraw();
-        if (!currentProgram || !currentProgram->GetLinkStatus()) {
+        if (!currentProgram || !currentProgram->GetLinkStatus() || !currentProgram->GetSpirvStatus()) {
             return nullptr;
         }
         if (PrgramImpl::g_currentDrawFrontendProgram == currentProgram.get()) {
@@ -3008,8 +3030,10 @@ namespace MobileGL::MG_Backend::DirectGLES {
         ZoneScopedC(TRACY_ZONECOLOR_BACKEND);
 #endif
         // Single per-dispatch program resolve and texture-key capture, as in
-        // PrepareForDraw (nothing below can move either).
-        const auto& currentProgram = MG_State::pGLContext->GetProgramForDraw();
+        // PrepareForDraw (nothing below can move either). The DISPATCH accessor: with a
+        // pipeline bound this is its compute stage program, which is a whole program on its
+        // own - the graphics composite a draw builds carries no compute stage.
+        const auto& currentProgram = MG_State::pGLContext->GetProgramForDispatch();
         const TextureImpl::DrawTextureSyncKeys textureKeys = TextureImpl::CaptureDrawTextureSyncKeys();
 
         BufferImpl::SyncComputeBuffers(includeDispatchIndirectBuffer);
@@ -3017,7 +3041,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
         TextureImpl::SyncImageTextureBindings();
         PrgramImpl::SyncCurrentProgram(currentProgram);
 
-        if (!currentProgram || !currentProgram->GetLinkStatus()) {
+        if (!currentProgram || !currentProgram->GetLinkStatus() || !currentProgram->GetSpirvStatus()) {
             g_GLESFuncs.glUseProgram(0);
             PrgramImpl::g_lastUsedBackendProgramId = 0;
             return;
@@ -3603,12 +3627,12 @@ namespace MobileGL::MG_Backend::DirectGLES {
             return false;
         }
 
-        if (s_resolveContextGeneration != TextureImpl::g_textureContextGeneration) {
+        if (s_resolveContextGeneration != g_backendContextGeneration) {
             // The ids belonged to a dead context; the context reclaimed them with it.
             s_resolveFramebuffer = 0;
             s_resolveRenderbuffer = 0;
             s_resolveFormat = 0;
-            s_resolveContextGeneration = TextureImpl::g_textureContextGeneration;
+            s_resolveContextGeneration = g_backendContextGeneration;
         }
         if (s_resolveFramebuffer == 0) {
             g_GLESFuncs.glGenFramebuffers(1, &s_resolveFramebuffer);
@@ -3756,7 +3780,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
         }
 
         static Bool EnsureResources() {
-            if (s_contextGeneration != TextureImpl::g_textureContextGeneration) {
+            if (s_contextGeneration != g_backendContextGeneration) {
                 // The ids belonged to a dead context; the context reclaimed them with it.
                 s_framebuffer = 0;
                 s_texture = 0;
@@ -3767,7 +3791,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
                 s_depthProgram = 0;
                 s_stencilProgram = 0;
                 s_programsFailed = false;
-                s_contextGeneration = TextureImpl::g_textureContextGeneration;
+                s_contextGeneration = g_backendContextGeneration;
             }
             if (s_programsFailed) {
                 return false;
@@ -5149,8 +5173,19 @@ namespace MobileGL::MG_Backend::DirectGLES {
                           const SharedPtr<MG_State::GLState::ITextureObject>& dstTexture,
                           GLenum dstTarget, GLint dstLevel, GLint dstX, GLint dstY, GLint dstZ,
                           GLsizei srcWidth, GLsizei srcHeight, GLsizei srcDepth) {
-        auto& srcBackendTexture = TextureImpl::SyncTextureObjectToBackend(srcTexture);
-        auto& dstBackendTexture = TextureImpl::SyncTextureObjectToBackend(dstTexture);
+        // BY VALUE, not by reference. SyncTextureObjectToBackend hands back a reference to a
+        // slot inside the backend texture registry, and the second call mutates that very map:
+        // GetOrCreate indexes it (an insert relocates entries - by rehashing, and also by
+        // robin-hood displacement well under the load factor), and Find drops any
+        // entry whose state object has expired - which, with the map open-addressed and erasing
+        // by shifting the probe cluster backwards, relocates entries other than the erased one.
+        // Either way a reference taken by the first call is stale by the time the second returns,
+        // and it is read four more times below. Copying the SharedPtr costs two refcount bumps on
+        // a path that is already doing a texture copy.
+        const SharedPtr<TextureImpl::BackendTextureObject> srcBackendTexture =
+            TextureImpl::SyncTextureObjectToBackend(srcTexture);
+        const SharedPtr<TextureImpl::BackendTextureObject> dstBackendTexture =
+            TextureImpl::SyncTextureObjectToBackend(dstTexture);
 
         const Bool srcIsDepth = MG_Util::IsDepthFormatInternalFormat(srcTexture->GetFormat());
         const Bool dstIsDepth = MG_Util::IsDepthFormatInternalFormat(dstTexture->GetFormat());
@@ -7198,6 +7233,83 @@ namespace MobileGL::MG_Backend::DirectGLES {
                g_GLESFuncs.glGetQueryObjectui64vEXT;
     }
 
+    namespace {
+        // The entry point the resolved tier's support ships, or null when there is none.
+        MG_External::GLES::glTexBuffer_PTR ResolveTexBufferEntryPoint() {
+            using Tier = MG_External::GLESCapabilities::TextureBufferTier;
+            switch (g_GLESCapabilities.TextureBufferSupport) {
+            case Tier::ExtensionEXT:
+                return g_GLESFuncs.glTexBufferEXT ? g_GLESFuncs.glTexBufferEXT : g_GLESFuncs.glTexBuffer;
+            case Tier::ExtensionOES:
+                return g_GLESFuncs.glTexBufferOES ? g_GLESFuncs.glTexBufferOES : g_GLESFuncs.glTexBuffer;
+            case Tier::CoreEs32:
+                return g_GLESFuncs.glTexBuffer;
+            case Tier::None:
+            default:
+                return nullptr;
+            }
+        }
+
+        MG_External::GLES::glTexBufferRange_PTR ResolveTexBufferRangeEntryPoint() {
+            using Tier = MG_External::GLESCapabilities::TextureBufferTier;
+            switch (g_GLESCapabilities.TextureBufferSupport) {
+            case Tier::ExtensionEXT:
+                return g_GLESFuncs.glTexBufferRangeEXT ? g_GLESFuncs.glTexBufferRangeEXT
+                                                       : g_GLESFuncs.glTexBufferRange;
+            case Tier::ExtensionOES:
+                return g_GLESFuncs.glTexBufferRangeOES ? g_GLESFuncs.glTexBufferRangeOES
+                                                       : g_GLESFuncs.glTexBufferRange;
+            case Tier::CoreEs32:
+                return g_GLESFuncs.glTexBufferRange;
+            case Tier::None:
+            default:
+                return nullptr;
+            }
+        }
+    } // namespace
+
+    Bool AreBufferTexturesSupported() {
+        // Both halves matter. The tier is what the driver ADVERTISES, and it is only meaningful
+        // once the capabilities have been filled in; the resolved pointer is what MobileGL can
+        // actually call, through the spelling that tier's support ships. Gating on the
+        // unsuffixed name alone would call an entry point an EXT/OES driver never exported.
+        return g_GLESCapabilities.TextureBufferSupport !=
+                   MG_External::GLESCapabilities::TextureBufferTier::None &&
+               ResolveTexBufferEntryPoint() != nullptr;
+    }
+
+    void CallTexBuffer(GLenum target, GLenum internalFormat, GLuint buffer) {
+        MG_External::GLES::glTexBuffer_PTR entryPoint = ResolveTexBufferEntryPoint();
+        if (entryPoint == nullptr) {
+            return;
+        }
+        entryPoint(target, internalFormat, buffer);
+    }
+
+    Bool CallTexBufferRange(GLenum target, GLenum internalFormat, GLuint buffer, GLintptr offset, GLsizeiptr size) {
+        MG_External::GLES::glTexBufferRange_PTR entryPoint = ResolveTexBufferRangeEntryPoint();
+        if (entryPoint == nullptr) {
+            return false;
+        }
+        entryPoint(target, internalFormat, buffer, offset, size);
+        return true;
+    }
+
+    const char* GetBufferTextureTierName() {
+        using Tier = MG_External::GLESCapabilities::TextureBufferTier;
+        switch (g_GLESCapabilities.TextureBufferSupport) {
+        case Tier::CoreEs32:
+            return "core (ES 3.2)";
+        case Tier::ExtensionEXT:
+            return "GL_EXT_texture_buffer";
+        case Tier::ExtensionOES:
+            return "GL_OES_texture_buffer";
+        case Tier::None:
+        default:
+            return "unsupported";
+        }
+    }
+
     BackendQueryHandle BeginTimeElapsedQuery() {
         // Query objects can only be created on the thread that owns the ES
         // context (MC's F3 profiler queries on the render thread, which
@@ -7480,7 +7592,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
         PixelStoreImpl::InvalidatePackStateCache();
         // Texture ids belong to the dying context; wrappers destroyed later must
         // not glDeleteTextures a recycled name in a successor context.
-        ++TextureImpl::g_textureContextGeneration;
+        ++g_backendContextGeneration;
         g_backendContextOwnerThread.store(std::thread::id{}, std::memory_order_release);
         // Outstanding fence handles now refer to a dead context; treat them as
         // signaled from here on.

MobileGL/MG_Backend/DirectGLES/DirectGLES.h

@@ -117,6 +117,24 @@ namespace MobileGL::MG_Backend::DirectGLES {
     // capability read needs no current ES context, and it stays false until
     // the ES capabilities have been filled in.
     Bool AreTimerQueriesSupported();
+    // True when the host ES driver can back a GL_TEXTURE_BUFFER at all - ES 3.2 core, or
+    // EXT/OES_texture_buffer, with glTexBuffer resolved. Desktop GL has had buffer textures as
+    // core since 3.1, so the frontend advertises them unconditionally and an app may call
+    // glTexBuffer whenever it likes; this is the only thing standing between that call and a
+    // null entry point. False also means every shader declaring a samplerBuffer is
+    // uncompilable on this driver, which the program build reports by name.
+    Bool AreBufferTexturesSupported();
+    // Human-readable name of the buffer-texture tier for diagnostics and the driver POST:
+    // "core (ES 3.2)", "GL_EXT_texture_buffer", "GL_OES_texture_buffer" or "unsupported".
+    const char* GetBufferTextureTierName();
+    // glTexBuffer / glTexBufferRange through whichever spelling this driver's buffer-texture
+    // support actually ships: the unsuffixed names are ES 3.2 core, while an EXT/OES driver
+    // exports glTexBuffer{,Range}EXT / OES. Callers must have checked
+    // AreBufferTexturesSupported() first. CallTexBufferRange reports whether it could honour
+    // the range - no tier is required to expose the range form, and the whole-buffer form is
+    // the documented fallback.
+    void CallTexBuffer(GLenum target, GLenum internalFormat, GLuint buffer);
+    Bool CallTexBufferRange(GLenum target, GLenum internalFormat, GLuint buffer, GLintptr offset, GLsizeiptr size);
     // GL timer-query objects, backed by GL_EXT_disjoint_timer_query. The
     // creators return null (the frontend then falls back to an immediately
     // available zero result) when the calling thread does not own the ES

MobileGL/MG_Backend/DirectGLES/Managers.cpp

@@ -33,6 +33,8 @@
 #include <regex>
 
 namespace MobileGL::MG_Backend::DirectGLES {
+    Uint g_backendContextGeneration = 1;
+
     constexpr Bool PREFER_MAP_BUFFER_RANGE_FOR_BUFFER_SYNC = false;
     constexpr const char* BASE_INSTANCE_UNIFORM_NAME = "mg_BaseInstance";
     constexpr const char* DRAW_ID_UNIFORM_NAME = "mg_DrawID";
@@ -53,6 +55,12 @@ namespace MobileGL::MG_Backend::DirectGLES {
         return g_GLESCapabilities.AvoidSamplerMipmapMinFilter;
     }
 
+    static Bool ShouldAvoidExplicitLodBiasOnAngleLlvmpipe() {
+        // IsAngleLlvmpipeRenderer combined with the MOBILEGL_AVOID_EXPLICIT_LOD_BIAS
+        // feature toggle, both resolved in FillInGLESCapabilities.
+        return g_GLESCapabilities.AvoidExplicitLodBias;
+    }
+
     static GLenum ResolveBackendMinFilter(const SamplerParameters& samplerParams,
                                           Bool avoidMipmapMinFilter) {
         GLenum filter = MG_Util::ConvertSamplerFilterModeToGLEnum(samplerParams.minFilter,
@@ -1646,7 +1654,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
             ZoneScopedC(TRACY_ZONECOLOR_BACKEND);
 #endif
             g_GLESFuncs.glGenTextures(1, &m_backendTextureId);
-            m_contextGeneration = g_textureContextGeneration;
+            m_contextGeneration = g_backendContextGeneration;
             if (m_backendTextureId == 0) {
                 MGLOG_E("Failed to generate texture object.");
                 MGLOG_E("ES glGetError(): %s", MG_Util::ConvertGLEnumToString(g_GLESFuncs.glGetError()).c_str());
@@ -1673,7 +1681,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
                     }
                 }
             }
-            if (m_contextGeneration == g_textureContextGeneration && g_GLESFuncs.glDeleteTextures) {
+            if (m_contextGeneration == g_backendContextGeneration && g_GLESFuncs.glDeleteTextures) {
                 g_GLESFuncs.glDeleteTextures(1, &m_backendTextureId);
             }
             m_backendTextureId = 0;
@@ -1712,7 +1720,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
         void BackendTextureObject::RecreateBackendTexture() {
             if (m_backendTextureId != 0) {
                 ScratchFBOImpl::NoteTextureIdDeleted(m_backendTextureId);
-                if (m_contextGeneration == g_textureContextGeneration) {
+                if (m_contextGeneration == g_backendContextGeneration) {
                     g_GLESFuncs.glDeleteTextures(1, &m_backendTextureId);
                 }
                 for (auto& unitCache : g_boundTexturesCache) {
@@ -1725,7 +1733,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
             }
 
             g_GLESFuncs.glGenTextures(1, &m_backendTextureId);
-            m_contextGeneration = g_textureContextGeneration;
+            m_contextGeneration = g_backendContextGeneration;
             if (m_backendTextureId == 0) {
                 MGLOG_E("Failed to regenerate texture object.");
                 MGLOG_E("ES glGetError(): %s", MG_Util::ConvertGLEnumToString(g_GLESFuncs.glGetError()).c_str());
@@ -2777,6 +2785,29 @@ namespace MobileGL::MG_Backend::DirectGLES {
                                                        &glType, TextureTarget::TextureBuffer);
 
                 if (needsRegeneration) {
+                    // Desktop GL has had buffer textures core since 3.1 and MobileGL advertises a
+                    // 4.x context, so glTexBuffer is a legal call the app may make on any driver -
+                    // but ES only gained them in 3.2, and g_GLESFuncs.glTexBuffer is simply null
+                    // below that without EXT/OES_texture_buffer. Calling it was an unconditional
+                    // null dereference. There is no conformant way to refuse the call (it is valid
+                    // in the context MobileGL claims), so the texture is left unbacked and the
+                    // reason is stated once per respecify at a level that survives the shipped
+                    // INFO build - MGLOG_E is compiled out there, which is exactly how this class
+                    // of defect stays invisible.
+                    if (!AreBufferTexturesSupported()) {
+                        if (m_bufferTextureUnsupportedReported) {
+                            break;
+                        }
+                        m_bufferTextureUnsupportedReported = true;
+                        MGLOG_I("Texture buffer %u cannot be backed: this ES driver has no buffer "
+                                "textures (%s). Every draw sampling it will read zero and every "
+                                "shader declaring a samplerBuffer will fail to compile. MobileGL "
+                                "still advertises GL_MAX_TEXTURE_BUFFER_SIZE = %d because an "
+                                "OpenGL 4.x context may not report 0.",
+                                stateTextureObject->GetExternalIndex(), GetBufferTextureTierName(),
+                                g_GLESCapabilities.MaxTextureBufferSize);
+                        break;
+                    }
                     MGLOG_D("Texture state changed significantly or not initialized, regenerating texture buffer with "
                             "ID: %u, buffer ID: %u, buffer size: %zu, format: %s",
                             m_backendTextureId, backendId, buffer->GetSize(),
@@ -2787,17 +2818,19 @@ namespace MobileGL::MG_Backend::DirectGLES {
                     // is absent).
                     const SizeT rangeOffset = textureBufferObject->GetBufferRangeOffset();
                     const SizeT rangeSize = textureBufferObject->GetBufferRangeSizeInBytes();
+                    // Through CallTexBuffer/CallTexBufferRange rather than g_GLESFuncs directly:
+                    // the unsuffixed entry points are the ES 3.2 core spelling, and a driver
+                    // whose buffer textures come from EXT/OES_texture_buffer exports the
+                    // suffixed ones instead. The dispatchers pick whichever this tier ships.
                     if (rangeOffset == 0 && rangeSize == buffer->GetSize()) {
-                        g_GLESFuncs.glTexBuffer(GL_TEXTURE_BUFFER, glInternalFormat, backendId);
-                    } else if (g_GLESFuncs.glTexBufferRange != nullptr) {
-                        g_GLESFuncs.glTexBufferRange(GL_TEXTURE_BUFFER, glInternalFormat, backendId,
-                                                     static_cast<GLintptr>(rangeOffset),
-                                                     static_cast<GLsizeiptr>(rangeSize));
-                    } else {
-                        MGLOG_E("Texture buffer %u names a sub-range but the driver has no "
+                        CallTexBuffer(GL_TEXTURE_BUFFER, glInternalFormat, backendId);
+                    } else if (!CallTexBufferRange(GL_TEXTURE_BUFFER, glInternalFormat, backendId,
+                                                   static_cast<GLintptr>(rangeOffset),
+                                                   static_cast<GLsizeiptr>(rangeSize))) {
+                        MGLOG_I("Texture buffer %u names a sub-range but the driver has no "
                                 "glTexBufferRange; binding the whole buffer instead",
                                 stateTextureObject->GetExternalIndex());
-                        g_GLESFuncs.glTexBuffer(GL_TEXTURE_BUFFER, glInternalFormat, backendId);
+                        CallTexBuffer(GL_TEXTURE_BUFFER, glInternalFormat, backendId);
                     }
                     DebugImpl::ErrorLopper::Loop(
                         [file = __FILE__, line = __LINE__, func = __func__, glInternalFormat, backendId](GLenum err) {
@@ -2809,7 +2842,14 @@ namespace MobileGL::MG_Backend::DirectGLES {
                 break;
             }
             default:
-                THROW_UNIMPL_EXCEPTION;
+                // TextureStorageType is {Mipmap, Buffer}, both handled above, so this is a
+                // backstop for a state object that grew a new storage kind. Skipping the upload
+                // renders wrong; throwing unwinds through the C GL ABI and kills the process.
+                MGLOG_I("DirectGLES texture sync: no upload path for storage type %d on texture %u; "
+                        "skipping this sync",
+                        static_cast<int>(stateTextureObject->GetStorageType()),
+                        stateTextureObject->GetExternalIndex());
+                break;
             }
 
             DebugImpl::ErrorLopper::Loop([file = __FILE__, line = __LINE__, func = __func__](GLenum err) {
@@ -3074,7 +3114,6 @@ namespace MobileGL::MG_Backend::DirectGLES {
         }
 
         Uint g_activeTextureUnit = 0;
-        Uint g_textureContextGeneration = 1;
         Array<Array<BackendTextureObject*, (SizeT)TextureTarget::TextureTargetCount>,
               MG_State::GLState::TextureState::MAX_TEXTURE_IMAGE_UNITS>
             g_boundTexturesCache;
@@ -3093,6 +3132,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
                 m_backendColorSlots[i] = GL_COLOR_ATTACHMENT0 + i;
             }
             g_GLESFuncs.glGenFramebuffers(1, &m_backendFBOId);
+            m_contextGeneration = g_backendContextGeneration;
             if (m_backendFBOId == 0) {
                 MGLOG_E("Failed to generate framebuffer object.");
                 MGLOG_E("ES glGetError(): %s", MG_Util::ConvertGLEnumToString(g_GLESFuncs.glGetError()).c_str());
@@ -3101,6 +3141,22 @@ namespace MobileGL::MG_Backend::DirectGLES {
             }
         }
 
+        BackendFramebufferObject::~BackendFramebufferObject() {
+            if (InProcessTeardown()) {
+                return; // see InProcessTeardown(): the driver may be unloaded already
+            }
+            if (m_backendFBOId == 0) {
+                return;
+            }
+            // Scrub the binding shadow whether or not the id can still be deleted: a
+            // recycled name must never satisfy the shadow's dedup.
+            NoteFramebufferIdDeleted(m_backendFBOId);
+            if (m_contextGeneration == g_backendContextGeneration && g_GLESFuncs.glDeleteFramebuffers) {
+                g_GLESFuncs.glDeleteFramebuffers(1, &m_backendFBOId);
+            }
+            m_backendFBOId = 0;
+        }
+
         void BackendFramebufferObject::Bind(FramebufferTarget target) const {
 #ifdef TRACY_ENABLE
             ZoneScopedC(TRACY_ZONECOLOR_BACKEND);
@@ -3156,6 +3212,17 @@ namespace MobileGL::MG_Backend::DirectGLES {
             return g_driverFBOBindings[idx];
         }
 
+        void NoteFramebufferIdDeleted(Uint id) {
+            if (id == 0) {
+                return;
+            }
+            for (SizeT idx = 0; idx < g_driverFBOBindings.size(); ++idx) {
+                if (g_driverFBOBindingKnown[idx] && g_driverFBOBindings[idx] == id) {
+                    g_driverFBOBindings[idx] = 0; // glDeleteFramebuffers reverts a bound FBO to 0
+                }
+            }
+        }
+
         void InvalidateFramebufferBindingCache() {
             g_driverFBOBindings = {0, 0};
             g_driverFBOBindingKnown = {false, false};
@@ -4146,6 +4213,31 @@ namespace MobileGL::MG_Backend::DirectGLES {
             }
         }
 
+        Uint64 ComputeShaderStorageBlockBindingSignature(
+            const MG_State::GLState::ProgramObject& stateProgramObject) {
+            const auto& overrides = stateProgramObject.GetShaderStorageBlockBindingOverrides();
+            if (overrides.empty()) return 0; // the overwhelming majority of programs
+            // Order-independent on purpose: the source is an UnorderedMap, so any signature that
+            // depended on iteration order would differ between two identical override sets and
+            // rebuild the program for nothing.
+            //
+            // Built from the VALUES, not from a change counter, so re-setting a block to the
+            // binding it already carries produces the same signature and forces no rebuild - an
+            // application that calls glShaderStorageBlockBinding every frame with unchanged
+            // arguments must not retranspile every frame.
+            Uint64 signature = 0;
+            for (const auto& [blockName, binding] : overrides) {
+                if (binding < 0) continue; // never rebound; the declared qualifier still stands
+                Uint64 entry = std::hash<String>{}(blockName);
+                // Mixed rather than merely summed with the name hash: name and binding must not
+                // be able to trade places between two entries and cancel out.
+                entry ^= (static_cast<Uint64>(static_cast<Uint32>(binding)) + 0x9e3779b97f4a7c15ull +
+                          (entry << 6) + (entry >> 2));
+                signature += entry; // commutative combine
+            }
+            return signature;
+        }
+
         void BackendProgramObjectImpl::SyncToBackend(
             const SharedPtr<MG_State::GLState::ProgramObject>& stateProgramObject) {
 #ifdef TRACY_ENABLE
@@ -4155,9 +4247,18 @@ namespace MobileGL::MG_Backend::DirectGLES {
                 MGLOG_E("State program object is null, skipping backend sync.");
                 return;
             }
+            // Recorded before either early return below, so Use() can always name the GL
+            // program a no-op draw belongs to - including the "linked but not drawable" exit.
+            m_frontendProgramId = stateProgramObject->GetExternalIndex();
 
-            if (!stateProgramObject->GetLinkStatus()) {
-                MGLOG_E("Program object is not linked, skipping backend sync. State program ID: %u",
+            // GetSpirvStatus() as well as GetLinkStatus(): a program whose phase-B job was
+            // cancelled (teardown) or whose optimizer run failed is fully linked and fully
+            // queryable, but has no SPIR-V to build a driver program out of. GL cannot retract
+            // a LINK_STATUS it already reported true, so "linked but not drawable" is the
+            // answer, and this is where the ES backend expresses it.
+            if (!stateProgramObject->GetLinkStatus() || !stateProgramObject->GetSpirvStatus()) {
+                MGLOG_E("Program object is not linked or has no generated SPIR-V, skipping backend sync. State "
+                        "program ID: %u",
                         stateProgramObject->GetExternalIndex());
                 return;
             }
@@ -4172,6 +4273,11 @@ namespace MobileGL::MG_Backend::DirectGLES {
             m_snormFallbackClampOutputMask = g_snormFallbackClampOutputMask;
             m_unormFallbackClampOutputMask = g_unormFallbackClampOutputMask;
             m_fragColorBroadcastCount = g_fragColorBroadcastCount;
+            // The generated ESSL bakes these in (see the SetShaderStorageBlockBinding call in the
+            // transpile loop below), so the set they were generated against is part of what makes
+            // this build current - the draw path compares the signature and rebuilds on a change.
+            const auto& storageBlockBindingOverrides = stateProgramObject->GetShaderStorageBlockBindingOverrides();
+            m_shaderStorageBlockBindingSignature = ComputeShaderStorageBlockBindingSignature(*stateProgramObject);
 
             // Detach all existing shaders
             GLint attachedCount = 0;
@@ -4222,6 +4328,26 @@ namespace MobileGL::MG_Backend::DirectGLES {
                 String source;
                 auto& spirvCode = shaderSpirvs[index];
 
+                // A samplerBuffer is core in the OpenGL 3.1+ context MobileGL advertises but needs
+                // ES 3.2 or EXT/OES_texture_buffer on the host. Without it SPIRV-Cross emits
+                // `#extension GL_EXT_texture_buffer : require` and the driver rejects both that
+                // and the isamplerBuffer keyword - the program never links and every draw using it
+                // becomes a silent no-op. Say so here, naming the stage, instead of leaving a
+                // driver info log the shipped INFO build compiles out (MGLOG_E is inactive there).
+                // Gated on the capability so the module walk never runs on a healthy driver.
+                if (!AreBufferTexturesSupported() &&
+                    MG_Util::ShaderTranspiler::ShaderCompiler::ModuleDeclaresBufferTextureSampler(spirvCode)) {
+                    MGLOG_I("Program %u stage %s samples a buffer texture, which this ES driver "
+                            "cannot provide (%s). The shader will not compile and the program will "
+                            "not link; every draw using it is a no-op.",
+                            m_backendProgramId,
+                            MG_Util::ConvertGLEnumToString(glShaderType).c_str(),
+                            GetBufferTextureTierName());
+                    m_backendProgramUsable = false;
+                    g_GLESFuncs.glDeleteShader(backendShaderId);
+                    continue;
+                }
+
                 // ESSL cannot express gl_DrawID/gl_BaseInstance/gl_BaseVertex; demote them to
                 // plain globals (mg_*) before handing the module to SPIRV-Cross.
                 Vector<unsigned int> loweredSpirv;
@@ -4271,6 +4397,22 @@ namespace MobileGL::MG_Backend::DirectGLES {
                     effectiveSpirv = &rectLoweredSpirv;
                 }
 
+                // GLSL ES demands a constant integral expression to index a fragment output
+                // array; SPIR-V does not, so a shader that writes coeff[i] from a loop
+                // reaches SPIRV-Cross intact and comes out as ESSL a strict driver rejects
+                // outright ("array indexes for fragment outputs must be constant integral
+                // expressions"), linking no program and silently no-oping every draw that
+                // uses it. Mesa accepts it, ANGLE does not - which is the whole of the
+                // improved-transparency-minecraft-26.3 failure. Fold or lower the index here,
+                // on the ESSL path only: the same module is legal for DirectVulkan.
+                Vector<unsigned int> outputIndexSpirv;
+                if (glShaderType == GL_FRAGMENT_SHADER &&
+                    MG_Util::ShaderTranspiler::ShaderCompiler::LegalizeFragmentOutputIndexingForEssl(
+                        *effectiveSpirv, outputIndexSpirv) &&
+                    !outputIndexSpirv.empty()) {
+                    effectiveSpirv = &outputIndexSpirv;
+                }
+
                 MG_Util::ShaderTranspiler::SpvcSession spvcSession(*effectiveSpirv,
                     MG_Util::ShaderTranspiler::SessionUsageBit::Transpile);
 
@@ -4284,6 +4426,15 @@ namespace MobileGL::MG_Backend::DirectGLES {
 
                 spvcSession.SetOptions(options);
 
+                // ES fixes a storage block's binding at link from its layout(binding=) qualifier
+                // and has no glShaderStorageBlockBinding to move it afterwards, so a rebinding
+                // can only be honoured by printing it INTO the qualifier. Rewriting the Binding
+                // decoration before SPIRV-Cross emits is what does that; RemoveLayoutBinding
+                // then deliberately preserves the qualifier for `buffer` declarations.
+                if (!storageBlockBindingOverrides.empty()) { // empty for almost every program
+                    spvcSession.SetShaderStorageBlockBinding(storageBlockBindingOverrides);
+                }
+
                 const char* result = nullptr;
                 spvcSession.Compile(&result);
 
@@ -4299,12 +4450,29 @@ namespace MobileGL::MG_Backend::DirectGLES {
 
                 source = result;
 
+                // Position in the chain is arbitrary: this is the only header-level rewrite, it
+                // edits #extension directives and never the body, and the replacement is the
+                // same length and stays an #extension line - so it commutes with every pass
+                // below, including ForceSupporterOutput's scan for the last directive. First,
+                // because a header concern reads better before the body ones.
+                source = RetargetTextureBufferExtension(std::move(source),
+                                                        g_GLESCapabilities.TextureBufferSupport);
+
                 source = RebindImageUniformsToFrontendUnits(std::move(source), stateProgramObject);
+                // Wedged between those two on purpose:
+                //  * AFTER RebindImageUniformsToFrontendUnits, so the binding it copies onto
+                //    both halves of a split image is already the frontend texture unit (and so
+                //    that pass never has to reason about the alias it introduces);
+                //  * BEFORE RemoveLayoutBinding, whose keepBindingRegex recognises an image
+                //    declaration and preserves its binding - an image unit cannot be set from
+                //    the API in ES, so the qualifier is the only binding mechanism there is,
+                //    and both halves of the pair have to still be carrying theirs when it runs.
+                source = SplitReadWriteImageUniforms(source);
                 source = RemoveLayoutBinding(source);
                 source = ProcessOutColorLocations(source);
                 source = ForceFlatIntegerVaryings(source, glShaderType);
                 source = BroadcastLegacyFragColor(std::move(source), glShaderType, m_fragColorBroadcastCount);
-                source = EmulateTextureLodBias(source);
+                source = EmulateTextureLodBias(source, ShouldAvoidExplicitLodBiasOnAngleLlvmpipe());
                 source = EmulateBaseInstanceInVertexShader(std::move(source), glShaderType);
                 source = PromoteDrawParameterGlobalsToUniforms(std::move(source), glShaderType);
                 source = ForceSupporterOutput(source);
@@ -4342,13 +4510,41 @@ namespace MobileGL::MG_Backend::DirectGLES {
                     Vector<GLchar> log(static_cast<SizeT>(logLength) + 1, '\0');
                     g_GLESFuncs.glGetShaderInfoLog(backendShaderId, logLength, nullptr, log.data());
                     log.back() = '\0';
-                    MGLOG_E("Shader compilation failed for backend ID %u: %s", backendShaderId, log.data());
+                    // MGLOG_I, deliberately. Every CI, retrace and release build compiles at
+                    // MOBILEGL_LOG_LEVEL_INFO, where MGLOG_E and MGLOG_W expand to nothing
+                    // (Log.h orders DEBUG < WARN < ERROR < INFO), so this diagnostic used to
+                    // exist only in debug builds: the Android retrace artifact carried 294
+                    // INFO lines and zero ERROR lines while two generated shaders were being
+                    // rejected outright, and the lane could not say why it was rendering an
+                    // empty translucent layer. A shader the driver refuses is never noise.
+                    MGLOG_I("Shader compilation failed. State program ID: %u, stage: %s, backend shader ID: "
+                            "%u, driver log: %s",
+                            stateProgramObject->GetExternalIndex(),
+                            MG_Util::ConvertGLEnumToString(glShaderType).c_str(), backendShaderId,
+                            log.data());
                     m_backendProgramUsable = false;
+                    // Nothing will ever attach this one, so nothing else can free it.
+                    g_GLESFuncs.glDeleteShader(backendShaderId);
                     continue;
                 }
 
                 MGLOG_D("Attaching shader ID: %u to program %u", backendShaderId, m_backendProgramId);
                 g_GLESFuncs.glAttachShader(m_backendProgramId, backendShaderId);
+                // Hand the shader's lifetime to the program, immediately and unconditionally.
+                //
+                // glDeleteShader only FLAGS a shader; the driver frees it when it is attached to
+                // nothing. Flagging it here is what makes the program own it, so deleting the
+                // program (or the detach loop above, on a relink) is what actually frees it.
+                // Without this call every program build leaked its shader objects for the process
+                // lifetime, and a relink leaked them twice - the detach loop above dropped the
+                // program's reference to shaders nothing had flagged, so they became unreachable
+                // AND undeletable. The GL swizzle conformance test builds 1,296 programs per case,
+                // so a handful of cases left tens of thousands of live driver shaders behind and
+                // the driver started mis-serving them (KHR-GL33/GL40.texture_swizzle.smoke_*).
+                // Same class of defect as the missing framebuffer/renderbuffer/sampler destructors
+                // fixed in Wave 1, and the last of that family: this is the one backend GL object
+                // MobileGL creates without an owning wrapper to destroy it.
+                g_GLESFuncs.glDeleteShader(backendShaderId);
 
                 MGLOG_D("Processed shader source length: %zu", source.length());
             }
@@ -4387,8 +4583,11 @@ namespace MobileGL::MG_Backend::DirectGLES {
                 Vector<GLchar> log(static_cast<SizeT>(logLength) + 1, '\0');
                 g_GLESFuncs.glGetProgramInfoLog(m_backendProgramId, logLength, nullptr, log.data());
                 log.back() = '\0';
-                MGLOG_E("Program %u linking failed for %u: %s", stateProgramObject->GetExternalIndex(),
-                        m_backendProgramId, log.data());
+                // MGLOG_I for the same reason as the compile failure above: a program that
+                // links nothing no-ops every draw that uses it, and that has to be readable
+                // in an INFO-level artifact.
+                MGLOG_I("Program linking failed. State program ID: %u, backend program ID: %u, driver log: %s",
+                        stateProgramObject->GetExternalIndex(), m_backendProgramId, log.data());
             } else {
                 MGLOG_D("Program linked successfully. ID: %u", m_backendProgramId);
             }
@@ -4419,18 +4618,38 @@ namespace MobileGL::MG_Backend::DirectGLES {
             }
 
             CacheResourceLocations(stateProgramObject);
-            // AFTER the link, because glShaderStorageBlockBinding needs the driver's linked
-            // interface. This is the only place Espryt applies a rebinding: the frontend
-            // record is authoritative and the glShaderStorageBlockBinding entry point itself
-            // deliberately never forces a program build (see DirectGLES.cpp), so a rebinding
-            // requested while no backend program existed yet arrives here instead.
+            // NOT the mechanism that makes a rebinding work - the transpiled qualifier above is.
+            // glShaderStorageBlockBinding is a GL 4.3 entry point that no real ES driver exposes,
+            // so this replay is a no-op almost everywhere; it stays because it is still correct
+            // (and cheaper than a rebuild) on a driver that does expose it, e.g. a desktop GL
+            // driver used as the ES backend. AFTER the link either way, because it needs the
+            // driver's linked interface.
             ReseedShaderStorageBlockBindings(m_backendProgramId, *stateProgramObject);
             m_syncedLinkVersion = stateProgramObject->GetLinkVersion();
+            m_syncedImageUnitVersion = stateProgramObject->GetImageUnitVersion();
 
             m_isInitialized = true;
             MGLOG_D("Program sync completed. backend ID %u", m_backendProgramId);
         }
 
+        namespace {
+            // The GL name of the array element that lives at `location`, given the reflection
+            // name reported for it. Reflection reports one name per UNIFORM ("goku[0]") but
+            // one location per ELEMENT, so a caller walking locations sees the same name
+            // repeatedly; this turns it back into "goku[k]". Anything that is not an array
+            // (or whose base location cannot be resolved) comes back unchanged, so the only
+            // behaviour that moves is the array case.
+            String SubscriptUniformNameForElement(const MG_State::GLState::ProgramObject& program, const String& name,
+                                                  Uint location) {
+                if (name.size() < 3 || name.compare(name.size() - 3, 3, "[0]") != 0) return name;
+                const Int base = program.GetUniformLocation(name);
+                if (base < 0 || static_cast<Uint>(base) > location) return name;
+                const Uint element = location - static_cast<Uint>(base);
+                if (element == 0) return name;
+                return name.substr(0, name.size() - 3) + "[" + std::to_string(element) + "]";
+            }
+        } // namespace
+
         // Resolves every name-based resource lookup once per link so the per-draw path
         // (BindCurrentProgramWithResources) never issues glGetUniformBlockIndex /
         // glGetUniformLocation string queries; block-to-binding-point assignments are
@@ -4493,7 +4712,17 @@ namespace MobileGL::MG_Backend::DirectGLES {
                     // is an INVALID_OPERATION.
                     continue;
                 }
-                const Int backendLoc = g_GLESFuncs.glGetUniformLocation(m_backendProgramId, name.c_str());
+                // Reflection names an array uniform after its FIRST element ("goku[0]") at
+                // every location the array spans, so asking the driver for that one name
+                // once per location hands back the same backend location N times. The
+                // per-draw pass then issues N glUniform1i calls against it and only the
+                // last element's unit survives - "layout(binding = 1) uniform sampler2D
+                // goku[7]" ended up with goku[0] on unit 7 and goku[1..6] still on 0.
+                // Address each element by its own name instead; the frontend already
+                // reserves one location per element, so the element index is the distance
+                // from the array's base location.
+                const String elementName = SubscriptUniformNameForElement(*stateProgramObject, name, loc);
+                const Int backendLoc = g_GLESFuncs.glGetUniformLocation(m_backendProgramId, elementName.c_str());
                 if (backendLoc < 0) continue;
                 SamplerUniformBinding binding;
                 binding.frontendLocation = loc;
@@ -4502,8 +4731,8 @@ namespace MobileGL::MG_Backend::DirectGLES {
                 binding.lastAssignedUnit = -1;
                 // Present only for the samplers EmulateTextureLodBias actually rewrote; the
                 // pass names it after the sampler, which SPIRV-Cross preserves verbatim.
-                binding.lodBiasLocation =
-                    g_GLESFuncs.glGetUniformLocation(m_backendProgramId, (String(LOD_BIAS_UNIFORM_PREFIX) + name).c_str());
+                binding.lodBiasLocation = g_GLESFuncs.glGetUniformLocation(
+                    m_backendProgramId, (String(LOD_BIAS_UNIFORM_PREFIX) + elementName).c_str());
                 binding.lastAssignedLodBias = 0.0f;
                 m_samplerUniformBindings.push_back(binding);
             }
@@ -4522,6 +4751,17 @@ namespace MobileGL::MG_Backend::DirectGLES {
             if (g_lastUsedBackendProgramId == programToBind) {
                 return;
             }
+            if (!m_backendProgramUsable) {
+                // MGLOG_I, not MGLOG_W: at MOBILEGL_LOG_LEVEL_INFO - the level the shipped
+                // fordebug builds compile at - only I and F survive, and this is precisely the
+                // line those builds need. Every draw made with this program renders nothing and
+                // raises no GL error, so without it the only symptom is a framebuffer that kept
+                // its clear colour. The early return above keeps it to at most one line per
+                // program state change, not one per draw.
+                MGLOG_I("Backend program for GL program %u is unusable (a shader failed to transpile, "
+                        "compile or link); binding program 0 - draws with it will render nothing",
+                        m_frontendProgramId);
+            }
             MGLOG_D("Using program %u", programToBind);
             g_GLESFuncs.glUseProgram(programToBind);
             g_lastUsedBackendProgramId = programToBind;
@@ -4557,6 +4797,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
             ZoneScopedC(TRACY_ZONECOLOR_BACKEND);
 #endif
             g_GLESFuncs.glGenSamplers(1, &m_backendSamplerId);
+            m_contextGeneration = g_backendContextGeneration;
             if (m_backendSamplerId == 0) {
                 MGLOG_E("Failed to generate sampler object.");
                 MGLOG_E("ES glGetError(): %s", MG_Util::ConvertGLEnumToString(g_GLESFuncs.glGetError()).c_str());
@@ -4565,6 +4806,26 @@ namespace MobileGL::MG_Backend::DirectGLES {
             }
         }
 
+        BackendSamplerObject::~BackendSamplerObject() {
+            if (InProcessTeardown()) {
+                return; // see InProcessTeardown(): the driver may be unloaded already
+            }
+            if (m_backendSamplerId == 0) {
+                return;
+            }
+            // Scrub the unit shadow whether or not the id can still be deleted - the next
+            // twin can land on this heap address and would otherwise false-skip its Bind.
+            for (auto& boundSampler : g_boundSamplersCache) {
+                if (boundSampler == this) {
+                    boundSampler = nullptr; // glDeleteSamplers unbinds from every unit
+                }
+            }
+            if (m_contextGeneration == g_backendContextGeneration && g_GLESFuncs.glDeleteSamplers) {
+                g_GLESFuncs.glDeleteSamplers(1, &m_backendSamplerId);
+            }
+            m_backendSamplerId = 0;
+        }
+
         void BackendSamplerObject::SyncToBackend(
             const SharedPtr<MG_State::GLState::SamplerObject>& stateSamplerObject) {
 #ifdef TRACY_ENABLE
@@ -4680,12 +4941,28 @@ namespace MobileGL::MG_Backend::DirectGLES {
             ZoneScopedC(TRACY_ZONECOLOR_BACKEND);
 #endif
             g_GLESFuncs.glGenRenderbuffers(1, &m_backendRBOId);
+            m_contextGeneration = g_backendContextGeneration;
             if (m_backendRBOId == 0) {
                 MGLOG_E("Failed to generate renderbuffer object.");
                 MGLOG_E("ES glGetError(): %s", MG_Util::ConvertGLEnumToString(g_GLESFuncs.glGetError()).c_str());
             }
         }
 
+        BackendRenderbufferObject::~BackendRenderbufferObject() {
+            if (InProcessTeardown()) {
+                return; // see InProcessTeardown(): the driver may be unloaded already
+            }
+            if (m_backendRBOId == 0) {
+                return;
+            }
+            // No driver-level renderbuffer-binding shadow exists (Bind() always issues the
+            // call), so there is nothing to scrub here - only the id to release.
+            if (m_contextGeneration == g_backendContextGeneration && g_GLESFuncs.glDeleteRenderbuffers) {
+                g_GLESFuncs.glDeleteRenderbuffers(1, &m_backendRBOId);
+            }
+            m_backendRBOId = 0;
+        }
+
         void BackendRenderbufferObject::Bind() const {
 #ifdef TRACY_ENABLE
             ZoneScopedC(TRACY_ZONECOLOR_BACKEND);

MobileGL/MG_Backend/DirectGLES/Managers.h

@@ -36,6 +36,14 @@ namespace MobileGL::MG_Backend::DirectGLES {
     Bool InProcessTeardown();
     void EnsureProcessTeardownSentinel();
 
+    // Generation of the backend ES context that owns the driver ids currently handed
+    // out. Bumped exactly once per DestroyEGLContext. Every backend twin that owns a
+    // driver name (texture, framebuffer, renderbuffer, sampler) stamps this at
+    // construction and compares it in its destructor: a twin outliving its context
+    // must NOT glDelete* its id, because a successor context may already have recycled
+    // that name and the delete would take out a live object of the new context.
+    extern Uint g_backendContextGeneration;
+
     // Which optional pieces of state a draw needs synchronized before it is issued.
     // Index/indirect buffer syncs and the instancing-related work are skipped for
     // draws that provably cannot read them.
@@ -121,6 +129,11 @@ namespace MobileGL::MG_Backend::DirectGLES {
 
         // Null when no live state object owns this key. The result points into the map, so
         // it stays valid only until the next GetOrCreate/Find/CollectGarbage on this registry.
+        // Take that literally, including for Find: the map is open-addressed and erases by
+        // shifting the rest of the probe cluster into the hole, so an erase relocates entries
+        // OTHER than the erased one - and Find erases, whenever it lands on a key whose state
+        // object has expired. Callers that need the twin across another registry call must copy
+        // the BackendPtr out (or keep only the pointee, which is heap-allocated and never moves).
         BackendPtr* Find(StateObject* stateObj) {
             const auto entryIt = m_entries.find(stateObj);
             if (entryIt == m_entries.end()) {
@@ -613,6 +626,11 @@ namespace MobileGL::MG_Backend::DirectGLES {
             Bool m_isInitialized = false;
             Bool m_imageBindableStorageRequired = false;
             Bool m_backendStorageImmutable = false;
+            // Latches the "this driver has no buffer textures" report to once per texture. The
+            // report is emitted from the respecify path, which bails before recording the state
+            // it was asked to apply - so without the latch the texture stays permanently dirty
+            // and every draw of every frame logs the same line.
+            Bool m_bufferTextureUnsupportedReported = false;
             StateTextureBasicInfo m_prevTextureInfo;
             // Frontend content version at the last completed mipmap sync. The per-draw
             // clean probe compares this before rebuilding shape info and scanning
@@ -657,15 +675,20 @@ namespace MobileGL::MG_Backend::DirectGLES {
                      MG_State::GLState::TextureState::MAX_TEXTURE_IMAGE_UNITS>
             g_boundTexturesCache;
         extern Uint g_activeTextureUnit;
-        // Bumped when the backend ES context is destroyed; texture ids stamped with
-        // an older generation belong to a dead context and must not be deleted.
-        extern Uint g_textureContextGeneration;
     } // namespace TextureImpl
 
     namespace FramebufferImpl {
         class BackendFramebufferObject {
         public:
             BackendFramebufferObject();
+            // Deletes the driver framebuffer and scrubs the binding shadow. Without it every
+            // frontend glDeleteFramebuffers leaked one ES framebuffer for the process lifetime;
+            // an app that creates a framebuffer per readback (GL CTS packed_pixels does ~3300
+            // per case) walked the driver into hundreds of megabytes of dead framebuffers and
+            // out of the resources a later attachment needs.
+            ~BackendFramebufferObject();
+            BackendFramebufferObject(const BackendFramebufferObject&) = delete;
+            BackendFramebufferObject& operator=(const BackendFramebufferObject&) = delete;
             void SyncToBackend(const SharedPtr<MG_State::GLState::FramebufferObject>& stateFBOObject,
                                FramebufferTarget asTarget);
             // Apply only this FBO's read buffer (glReadBuffer) to the backend. Split out so it can
@@ -680,6 +703,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
 
         private:
             Uint m_backendFBOId = 0;
+            Uint m_contextGeneration = 0;
 
             /* this will save buffers in its original form,
                reversion, absence or not consecutive are all allowed, as long as GL spec allows it
@@ -821,6 +845,10 @@ namespace MobileGL::MG_Backend::DirectGLES {
         void BindFramebufferId(GLenum fbTarget, Uint id);
         Uint CurrentFramebufferBinding(FramebufferTarget target);
         void InvalidateFramebufferBindingCache();
+        // A driver framebuffer id is about to be deleted: ES reverts every target that
+        // currently binds it to 0, so the binding shadow has to follow or the next
+        // BindFramebufferId(0) would be deduped away and leave the deleted name bound.
+        void NoteFramebufferIdDeleted(Uint id);
     } // namespace FramebufferImpl
 
     // Shared scratch framebuffers for the readback/copy/blit emulation paths, with a
@@ -1010,6 +1038,11 @@ namespace MobileGL::MG_Backend::DirectGLES {
             Uint32 GetSnormFallbackClampOutputMask() const { return m_snormFallbackClampOutputMask; }
             Uint32 GetUnormFallbackClampOutputMask() const { return m_unormFallbackClampOutputMask; }
             Uint GetFragColorBroadcastCount() const { return m_fragColorBroadcastCount; }
+            // Signature of the glShaderStorageBlockBinding override set the generated ESSL was
+            // transpiled against (ES can only express a storage-block binding as the declared
+            // qualifier, so the overrides are baked into the source). A mismatch means the
+            // program is stale exactly like the clamp masks above.
+            Uint64 GetShaderStorageBlockBindingSignature() const { return m_shaderStorageBlockBindingSignature; }
 
             Bool HasGlobalUboBlock() const { return m_globalUboBackendBlockIndex >= 0; }
             const Vector<Int>& GetUniformBlockBackendIndices() const { return m_uniformBlockBackendIndices; }
@@ -1025,11 +1058,22 @@ namespace MobileGL::MG_Backend::DirectGLES {
             // Frontend link version this backend program (and its resource caches) was
             // built from; a mismatch means every link-derived cache here is stale.
             Uint32 GetSyncedLinkVersion() const { return m_syncedLinkVersion; }
+            // Image-uniform unit generation this backend program was GENERATED against.
+            // Separate from the link version because it is not link state: ES forbids
+            // glUniform1i on an image uniform, so RebindImageUniformsToFrontendUnits bakes the
+            // unit into the ESSL, and a program built before glUniform1i moved that unit is as
+            // stale as one built before a relink - while the sampler half, which really is
+            // re-issued per draw, needs nothing of the sort.
+            Uint32 GetSyncedImageUnitVersion() const { return m_syncedImageUnitVersion; }
 
         private:
             void CacheResourceLocations(const SharedPtr<MG_State::GLState::ProgramObject>& stateProgramObject);
 
             Uint m_backendProgramId = 0;
+            // GL name of the frontend program this was last synced from; diagnostics only, so
+            // an unusable backend program can be traced back to the glCreateProgram id the app
+            // knows it by.
+            Uint m_frontendProgramId = 0;
             Uint m_backendGlobalUBOId = 0;
             Int m_baseInstanceUniformLocation = -1;
             Int m_drawIdUniformLocation = -1;
@@ -1040,6 +1084,8 @@ namespace MobileGL::MG_Backend::DirectGLES {
             // Draw buffers a legacy gl_FragColor write has to reach (see
             // PrgramImpl::BroadcastLegacyFragColor); 1 keeps the plain single-output shader.
             Uint m_fragColorBroadcastCount = 1;
+            // 0 is the signature of an empty override set, i.e. what almost every program has.
+            Uint64 m_shaderStorageBlockBindingSignature = 0;
             Bool m_isInitialized = false;
             Bool m_backendProgramUsable = false;
 
@@ -1050,6 +1096,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
             Uint32 m_lastUploadedGlobalUboVersion = ~0u;
             BufferImpl::UboRingAllocation m_globalUboRingAllocation;
             Uint32 m_syncedLinkVersion = ~0u;
+            Uint32 m_syncedImageUnitVersion = ~0u;
             SamplerPassMemo m_samplerPassMemo;
         };
 
@@ -1073,26 +1120,45 @@ namespace MobileGL::MG_Backend::DirectGLES {
         // on the backend program (eliminated as unused, or the driver lacks the entry
         // points), which is not an error - GL_BUFFER_BINDING is served from the frontend
         // record either way.
+        //
+        // NOT how a rebinding reaches the shader. glShaderStorageBlockBinding has no ES
+        // equivalent and is absent from every real ES driver, so this is a no-op there;
+        // SyncToBackend bakes the effective binding into the ESSL it generates instead
+        // (SpvcSession::SetShaderStorageBlockBinding). This is kept as the cheaper path on
+        // a driver that does happen to expose the entry point.
         Bool ApplyShaderStorageBlockBinding(Uint backendProgramId, const String& blockName, Uint binding);
         // Replays every glShaderStorageBlockBinding recorded on the program onto a backend
-        // program that was just built. The frontend record is authoritative (only the
-        // shader's DECLARED binding survives in the SPIR-V), so without this replay any
-        // rebuild would silently revert rebound blocks. Mirrors DirectVulkan's
-        // reseed-on-rebuild in BuildProgramResourceCache.
+        // program that was just built - best effort, on the same "only where the driver has
+        // the entry point" terms as ApplyShaderStorageBlockBinding above. Mirrors
+        // DirectVulkan's reseed-on-rebuild in BuildProgramResourceCache.
         void ReseedShaderStorageBlockBindings(Uint backendProgramId,
                                               const MG_State::GLState::ProgramObject& stateProgramObject);
+        // Order-independent digest of the program's glShaderStorageBlockBinding overrides.
+        // The generated ESSL carries them (ES has no way to move a storage block's binding
+        // after link), so a program built against a different set is stale and the draw path
+        // has to rebuild it. Computed from the values, so re-setting a block to the binding it
+        // already has costs nothing. 0 when nothing was ever rebound.
+        Uint64 ComputeShaderStorageBlockBindingSignature(
+            const MG_State::GLState::ProgramObject& stateProgramObject);
     } // namespace PrgramImpl
 
     namespace SamplerImpl {
         class BackendSamplerObject {
         public:
             BackendSamplerObject();
+            // Deletes the driver sampler and clears the units whose binding shadow still names
+            // this twin (a recycled heap address would otherwise false-skip a later Bind).
+            // Frontend glDeleteSamplers used to leak the backend id for the process lifetime.
+            ~BackendSamplerObject();
+            BackendSamplerObject(const BackendSamplerObject&) = delete;
+            BackendSamplerObject& operator=(const BackendSamplerObject&) = delete;
             void SyncToBackend(const SharedPtr<MG_State::GLState::SamplerObject>& stateSamplerObject);
             void Bind(Uint unit);
             Uint GetBackendSamplerId() const;
 
         private:
             Uint m_backendSamplerId = 0;
+            Uint m_contextGeneration = 0;
             Bool m_isInitialized = false;
             SamplerParameters m_cacheSamplerParameters;
             Uint16 m_syncedSamplerVersion = 0;
@@ -1110,12 +1176,18 @@ namespace MobileGL::MG_Backend::DirectGLES {
         class BackendRenderbufferObject {
         public:
             BackendRenderbufferObject();
+            // Deletes the driver renderbuffer; frontend glDeleteRenderbuffers used to leak it
+            // (with its whole image allocation) for the process lifetime.
+            ~BackendRenderbufferObject();
+            BackendRenderbufferObject(const BackendRenderbufferObject&) = delete;
+            BackendRenderbufferObject& operator=(const BackendRenderbufferObject&) = delete;
             void SyncToBackend(const SharedPtr<MG_State::GLState::RenderbufferObject>& stateRBOObject);
             Uint GetBackendRenderbufferId() const { return m_backendRBOId; }
             void Bind() const;
 
         private:
             Uint m_backendRBOId = 0;
+            Uint m_contextGeneration = 0;
             Bool m_isInitialized = false;
             TextureInternalFormat m_cacheInternalFormat = TextureInternalFormat::Unknown;
             Int m_cacheWidth = 0;

MobileGL/MG_Backend/DirectGLES/Utils.cpp

@@ -21,6 +21,7 @@
 #include <MG_Util/Math/HalfFloat.h>
 #include <MG_Util/Math/SmallFloat.h>
 
+#include <algorithm>
 #include <cmath>
 #include <cctype>
 #include <cstring>
@@ -434,6 +435,89 @@ namespace MobileGL::MG_Backend::DirectGLES {
             return result;
         }
 
+        String RetargetTextureBufferExtension(String glslCode,
+                                              MG_External::GLESCapabilities::TextureBufferTier tier) {
+#ifdef TRACY_ENABLE
+            ZoneScopedC(TRACY_ZONECOLOR_BACKEND);
+#endif
+            // SPIRV-Cross hardcodes the EXT spelling: CompilerGLSL::type_to_glsl emits
+            // require_extension_internal("GL_EXT_texture_buffer") for any Dim=Buffer image
+            // whenever it targets ESSL below 320, with no OES alternative and no way to
+            // configure it. GL_OES_texture_buffer is functionally identical but is a separate
+            // directive, and `#extension <name> : require` on a name the driver does not
+            // advertise is a hard compile error - so on an OES-only driver the emitted shader
+            // fails to compile for the sake of one token.
+            //
+            // Line comments are excluded by the directive check below; a `#extension` line inside
+            // a /* */ block is not, and would be rewritten. That is harmless (it stays a comment)
+            // and is not worth a preprocessor-aware scan here.
+            //
+            // Deliberately a directive rewrite and nothing more. The alternative - teaching the
+            // SPIR-V to stop asking for the extension - is not available: the requirement is
+            // synthesized by SPIRV-Cross from the image type itself, not carried in the module,
+            // so there is nothing upstream to strip. Everything about the shader body that
+            // actually uses the buffer texture is identical between the two extensions.
+            using Tier = MG_External::GLESCapabilities::TextureBufferTier;
+            if (tier != Tier::ExtensionOES) {
+                return glslCode;
+            }
+            static constexpr const char* kExtName = "GL_EXT_texture_buffer";
+            static constexpr const char* kOesName = "GL_OES_texture_buffer";
+            constexpr SizeT kExtNameLength = 21; // strlen("GL_EXT_texture_buffer")
+            static_assert(sizeof("GL_EXT_texture_buffer") - 1 == kExtNameLength, "name length drifted");
+            static_assert(sizeof("GL_OES_texture_buffer") - 1 == kExtNameLength,
+                          "the two spellings must be the same length for the in-place replace");
+
+            // Only rewrite the name where it is the whole subject of an #extension directive.
+            // Two separate guards, both load-bearing:
+            //   * the directive check, so a line-comment mentioning the name is left alone;
+            //   * the identifier-boundary check, because GL_EXT_texture_buffer is a PREFIX of
+            //     GL_EXT_texture_buffer_object - a different, real extension that SPIRV-Cross
+            //     emits from the same `case DimBuffer:` on its legacy-desktop branch. Without
+            //     the boundary this pass would silently rewrite a request for that extension
+            //     into a request for a GL_OES_texture_buffer_object that does not exist.
+            const auto isIdentifierChar = [](char c) {
+                return std::isalnum(static_cast<unsigned char>(c)) != 0 || c == '_';
+            };
+            SizeT searchFrom = 0;
+            while (true) {
+                const SizeT hit = glslCode.find(kExtName, searchFrom);
+                if (hit == String::npos) {
+                    break;
+                }
+                searchFrom = hit + kExtNameLength;
+
+                // Identifier boundary on both sides, so the name is not a fragment of a longer one.
+                if (hit > 0 && isIdentifierChar(glslCode[hit - 1])) {
+                    continue;
+                }
+                if (hit + kExtNameLength < glslCode.size() && isIdentifierChar(glslCode[hit + kExtNameLength])) {
+                    continue;
+                }
+
+                // Walk back to the start of the line and require that it is an #extension
+                // directive, allowing whitespace between '#' and the keyword.
+                SizeT lineStart = glslCode.rfind('\n', hit);
+                lineStart = (lineStart == String::npos) ? 0 : lineStart + 1;
+                SizeT cursor = lineStart;
+                while (cursor < hit && std::isspace(static_cast<unsigned char>(glslCode[cursor]))) {
+                    ++cursor;
+                }
+                if (cursor >= hit || glslCode[cursor] != '#') {
+                    continue;
+                }
+                ++cursor;
+                while (cursor < hit && std::isspace(static_cast<unsigned char>(glslCode[cursor]))) {
+                    ++cursor;
+                }
+                if (glslCode.compare(cursor, 9, "extension") != 0) {
+                    continue;
+                }
+                glslCode.replace(hit, kExtNameLength, kOesName);
+            }
+            return glslCode;
+        }
+
         String RemoveLayoutBinding(const String& glslCode) {
 #ifdef TRACY_ENABLE
             ZoneScopedC(TRACY_ZONECOLOR_BACKEND);
@@ -470,6 +554,352 @@ namespace MobileGL::MG_Backend::DirectGLES {
             return result;
         }
 
+        namespace {
+            Bool IsImagePassIdentifierChar(char c) {
+                return std::isalnum(static_cast<unsigned char>(c)) || c == '_';
+            }
+
+            // Occurrences of `identifier` in `code` that are whole identifiers, i.e. not the
+            // tail or head of a longer one. "goku" must not find "goku_hd" or "my_goku".
+            SizeT CountIdentifierOccurrences(const String& code, const String& identifier) {
+                if (identifier.empty()) return 0;
+                SizeT count = 0;
+                for (SizeT pos = code.find(identifier); pos != String::npos;
+                     pos = code.find(identifier, pos + 1)) {
+                    if (pos > 0 && IsImagePassIdentifierChar(code[pos - 1])) continue;
+                    const SizeT after = pos + identifier.size();
+                    if (after < code.size() && IsImagePassIdentifierChar(code[after])) continue;
+                    ++count;
+                }
+                return count;
+            }
+
+            Bool ContainsIdentifier(const String& code, const String& identifier) {
+                return CountIdentifierOccurrences(code, identifier) > 0;
+            }
+
+            // The image format layout qualifiers ESSL accepts (GLSL ES 3.20 4.4.7 table 4.6 -
+            // the ES-legal subset of what SPIRV-Cross's format_to_glsl can print). The
+            // readonly/writeonly rule only applies to a declaration that carries one of them.
+            Bool IsImageFormatQualifier(const String& token) {
+                static constexpr StringView FORMATS[] = {
+                    "rgba32f",      "rgba16f",      "rg32f",       "rg16f",        "r11f_g11f_b10f",
+                    "r32f",         "r16f",         "rgba16",      "rgb10_a2",     "rgba8",
+                    "rg16",         "rg8",          "r16",         "r8",           "rgba16_snorm",
+                    "rgba8_snorm",  "rg16_snorm",   "rg8_snorm",   "r16_snorm",    "r8_snorm",
+                    "rgba32i",      "rgba16i",      "rgba8i",      "rg32i",        "rg16i",
+                    "rg8i",         "r32i",         "r16i",        "r8i",          "rgba32ui",
+                    "rgba16ui",     "rgb10_a2ui",   "rgba8ui",     "rg32ui",       "rg16ui",
+                    "rg8ui",        "r32ui",        "r16ui",       "r8ui",
+                };
+                for (const StringView format : FORMATS) {
+                    if (token == format) return true;
+                }
+                return false;
+            }
+
+            // "Except for image variables qualified with the format qualifiers r32f, r32i, and
+            // r32ui, image variables must specify either memory qualifier readonly or the
+            // memory qualifier writeonly." (GLSL ES 3.20 4.10)
+            Bool IsMemoryQualifierExemptImageFormat(const String& token) {
+                return token == "r32f" || token == "r32i" || token == "r32ui";
+            }
+
+            // Comma-separated contents of a layout(...) list, each entry trimmed.
+            Vector<String> SplitLayoutQualifierList(const String& layout) {
+                Vector<String> tokens;
+                SizeT start = 0;
+                while (start <= layout.size()) {
+                    SizeT comma = layout.find(',', start);
+                    const Bool last = comma == String::npos;
+                    String token = layout.substr(start, last ? String::npos : comma - start);
+                    const SizeT first = token.find_first_not_of(" \t\r\n");
+                    if (first == String::npos) {
+                        token.clear();
+                    } else {
+                        token = token.substr(first, token.find_last_not_of(" \t\r\n") - first + 1);
+                    }
+                    if (!token.empty()) tokens.push_back(Move(token));
+                    if (last) break;
+                    start = comma + 1;
+                }
+                return tokens;
+            }
+
+            // Trims both ends and collapses every internal whitespace run to one space, so a
+            // qualifier list or array suffix can be spliced back into a rebuilt declaration
+            // whatever the original spacing was.
+            String NormalizeDeclarationSpacing(const String& text) {
+                String out;
+                out.reserve(text.size());
+                Bool pendingSpace = false;
+                for (const char c : text) {
+                    if (std::isspace(static_cast<unsigned char>(c))) {
+                        pendingSpace = !out.empty();
+                        continue;
+                    }
+                    if (pendingSpace) out += ' ';
+                    pendingSpace = false;
+                    out += c;
+                }
+                return out;
+            }
+
+            // How an image builtin touches the image it is handed.
+            enum class ImageBuiltinAccess { None, Load, Store, Unknown };
+
+            ImageBuiltinAccess ClassifyImageBuiltin(const String& name) {
+                if (name == "imageStore") return ImageBuiltinAccess::Store;
+                if (name == "imageLoad") return ImageBuiltinAccess::Load;
+                // imageAtomic* both reads and writes, but ES only defines the atomics on
+                // r32i/r32ui/r32f images - exactly the formats the rule above exempts - so this
+                // pass has already skipped any declaration they can legally appear on. Load is
+                // enough to keep the classification total without ever being acted upon.
+                if (name.compare(0, 11, "imageAtomic") == 0) return ImageBuiltinAccess::Load;
+                if (name == "imageSize" || name == "imageSamples") return ImageBuiltinAccess::None;
+                // Some other identifier that starts with "image" and is being called: not a
+                // shape this pass can reason about, so it poisons the declaration instead of
+                // being guessed at.
+                return ImageBuiltinAccess::Unknown;
+            }
+
+            struct ImageUniformDecl {
+                String name;
+                String writeName;   // the writeonly half's name, when split
+                String layout;      // raw contents of layout(...)
+                String qualifiers;  // memory/precision qualifiers, normalized, no trailing space
+                String type;        // image2D, uimage2DArray, ...
+                String arraySuffix; // "" or "[7]"
+                SizeT declStart = 0;
+                SizeT declLength = 0;
+                SizeT referenceCount = 0; // uses this pass recognized and accounted for
+                Bool loaded = false;
+                Bool stored = false;
+                Bool unknownUse = false;
+                Bool split = false;
+            };
+
+            // A rebuilt declaration. Keeps SPIRV-Cross's own word order (`uniform readonly
+            // highp image2D`) so the image-rebinding regex in Managers.cpp still matches what
+            // comes out of here, whichever order the two passes end up running in.
+            String BuildImageDeclaration(const ImageUniformDecl& decl, const char* memoryQualifier,
+                                         const String& variableName) {
+                String out = "layout(" + decl.layout + ") uniform ";
+                out += memoryQualifier;
+                out += ' ';
+                if (!decl.qualifiers.empty()) {
+                    out += decl.qualifiers;
+                    out += ' ';
+                }
+                out += decl.type;
+                out += ' ';
+                out += variableName;
+                out += decl.arraySuffix;
+                out += ';';
+                return out;
+            }
+
+            // A name for the writeonly half that no identifier in the shader (and no other
+            // half already minted) can collide with.
+            String MakeImageWriteAliasName(const String& name, const String& source,
+                                           const Vector<String>& taken) {
+                String candidate = String(IMAGE_WRITE_ALIAS_PREFIX) + name;
+                // "__" anywhere in an identifier is reserved (GLSL ES 3.20 3.7), which a name
+                // that already starts with '_' would otherwise produce.
+                for (SizeT doubled = candidate.find("__"); doubled != String::npos;
+                     doubled = candidate.find("__", doubled)) {
+                    candidate.erase(doubled, 1);
+                }
+                auto isTaken = [&](const String& identifier) {
+                    if (ContainsIdentifier(source, identifier)) return true;
+                    for (const auto& other : taken) {
+                        if (other == identifier) return true;
+                    }
+                    return false;
+                };
+                while (isTaken(candidate)) candidate += 'X';
+                return candidate;
+            }
+
+            struct ImageSourceEdit {
+                SizeT start;
+                SizeT length;
+                String text;
+            };
+        } // namespace
+
+        String SplitReadWriteImageUniforms(const String& glslCode) {
+#ifdef TRACY_ENABLE
+            ZoneScopedC(TRACY_ZONECOLOR_BACKEND);
+#endif
+            if (glslCode.find("image") == String::npos) {
+                return glslCode;
+            }
+
+            // layout(...) uniform <memory/precision qualifiers> <image type> <name>[array];
+            // The qualifier alternation is order-free even though SPIRV-Cross emits a fixed
+            // order (to_qualifiers_glsl: storage, then coherent/restrict/readonly/writeonly,
+            // then precision), and the array group is repeated so a hypothetical multi-
+            // dimensional image array survives the round trip intact.
+            static const std::regex imageDeclRegex(
+                R"(layout\s*\(([^)]*)\)\s*uniform\s+)"
+                R"(((?:(?:readonly|writeonly|coherent|volatile|restrict|highp|mediump|lowp)\s+)*))"
+                R"(([iu]?image[A-Za-z0-9_]*)\s+([A-Za-z_][A-Za-z0-9_]*)\s*((?:\[[^\]]*\]\s*)*);)");
+
+            Vector<ImageUniformDecl> decls;
+            for (std::sregex_iterator it(glslCode.begin(), glslCode.end(), imageDeclRegex), last; it != last; ++it) {
+                const std::smatch& match = *it;
+                const String qualifiers = match[2].str();
+                // Already legal: SPIRV-Cross decided one way, leave it alone.
+                if (ContainsIdentifier(qualifiers, "readonly") || ContainsIdentifier(qualifiers, "writeonly")) {
+                    continue;
+                }
+
+                Bool hasFormat = false;
+                Bool exemptFormat = false;
+                for (const String& token : SplitLayoutQualifierList(match[1].str())) {
+                    if (!IsImageFormatQualifier(token)) continue;
+                    hasFormat = true;
+                    exemptFormat = IsMemoryQualifierExemptImageFormat(token);
+                }
+                // No format qualifier at all is a different (and, in ES, unconditionally
+                // illegal) shape that GL_EXT_shader_image_load_formatted would be needed for;
+                // SPIRV-Cross refuses to emit it for an ES target, so nothing to do here.
+                if (!hasFormat || exemptFormat) continue;
+
+                ImageUniformDecl decl;
+                decl.layout = match[1].str();
+                decl.qualifiers = NormalizeDeclarationSpacing(qualifiers);
+                decl.type = match[3].str();
+                decl.name = match[4].str();
+                decl.arraySuffix = NormalizeDeclarationSpacing(match[5].str());
+                decl.declStart = static_cast<SizeT>(match.position(0));
+                decl.declLength = match[0].str().size();
+                decls.push_back(Move(decl));
+            }
+            if (decls.empty()) {
+                return glslCode;
+            }
+
+            auto findDecl = [&decls](const String& name) -> SizeT {
+                for (SizeT i = 0; i < decls.size(); ++i) {
+                    if (decls[i].name == name) return i;
+                }
+                return decls.size();
+            };
+
+            // Walk every `image*(` call and attribute its first argument to a declaration.
+            struct StoreSite {
+                SizeT declIndex;
+                SizeT start;
+                SizeT length;
+            };
+            Vector<StoreSite> storeSites;
+            for (SizeT pos = glslCode.find("image"); pos != String::npos; pos = glslCode.find("image", pos + 1)) {
+                if (pos > 0 && IsImagePassIdentifierChar(glslCode[pos - 1])) continue; // uimage2D, myimageFoo
+                SizeT tokenEnd = pos;
+                while (tokenEnd < glslCode.size() && IsImagePassIdentifierChar(glslCode[tokenEnd])) ++tokenEnd;
+                const String builtin = glslCode.substr(pos, tokenEnd - pos);
+
+                const SizeT openParen = glslCode.find_first_not_of(" \t\r\n", tokenEnd);
+                if (openParen == String::npos || glslCode[openParen] != '(') continue; // a type, not a call
+
+                const SizeT argStart = glslCode.find_first_not_of(" \t\r\n", openParen + 1);
+                if (argStart == String::npos) continue;
+                if (!std::isalpha(static_cast<unsigned char>(glslCode[argStart])) && glslCode[argStart] != '_') {
+                    continue; // an expression, not a bare variable - it names no image of ours
+                }
+                SizeT argEnd = argStart;
+                while (argEnd < glslCode.size() && IsImagePassIdentifierChar(glslCode[argEnd])) ++argEnd;
+
+                const SizeT declIndex = findDecl(glslCode.substr(argStart, argEnd - argStart));
+                if (declIndex == decls.size()) continue;
+                ImageUniformDecl& decl = decls[declIndex];
+                ++decl.referenceCount;
+
+                // The operand has to be the bare variable, optionally subscripted. Anything
+                // else (a member access, a call result) is a shape this pass cannot rewrite.
+                SizeT after = glslCode.find_first_not_of(" \t\r\n", argEnd);
+                if (after != String::npos && glslCode[after] == '[') {
+                    Int depth = 0;
+                    SizeT scan = after;
+                    for (; scan < glslCode.size(); ++scan) {
+                        if (glslCode[scan] == '[') ++depth;
+                        else if (glslCode[scan] == ']' && --depth == 0) break;
+                    }
+                    after = scan >= glslCode.size() ? String::npos
+                                                    : glslCode.find_first_not_of(" \t\r\n", scan + 1);
+                }
+                const char nextChar = after == String::npos ? '\0' : glslCode[after];
+                if (nextChar != ',' && nextChar != ')') {
+                    decl.unknownUse = true;
+                    continue;
+                }
+
+                switch (ClassifyImageBuiltin(builtin)) {
+                case ImageBuiltinAccess::Load:
+                    decl.loaded = true;
+                    break;
+                case ImageBuiltinAccess::Store:
+                    decl.stored = true;
+                    storeSites.push_back({declIndex, argStart, argEnd - argStart});
+                    break;
+                case ImageBuiltinAccess::None:
+                    break;
+                default:
+                    decl.unknownUse = true;
+                    break;
+                }
+            }
+
+            // Every mention of the name has to be one this pass saw, or the split would leave
+            // a store pointing at the readonly half. One occurrence is the declaration itself.
+            for (auto& decl : decls) {
+                if (CountIdentifierOccurrences(glslCode, decl.name) != decl.referenceCount + 1) {
+                    decl.unknownUse = true;
+                }
+            }
+
+            Vector<ImageSourceEdit> edits;
+            Vector<String> takenAliases;
+            for (auto& decl : decls) {
+                if (decl.unknownUse) continue; // leave it exactly as it was; no guessing
+                if (decl.loaded && decl.stored) {
+                    decl.writeName = MakeImageWriteAliasName(decl.name, glslCode, takenAliases);
+                    takenAliases.push_back(decl.writeName);
+                    decl.split = true;
+                    edits.push_back({decl.declStart, decl.declLength,
+                                     BuildImageDeclaration(decl, "readonly", decl.name) + "\n" +
+                                         BuildImageDeclaration(decl, "writeonly", decl.writeName)});
+                } else if (decl.stored) {
+                    edits.push_back({decl.declStart, decl.declLength,
+                                     BuildImageDeclaration(decl, "writeonly", decl.name)});
+                } else {
+                    // Loaded only, or only ever handed to imageSize (or unused): readonly is
+                    // the qualifier that keeps every one of those legal.
+                    edits.push_back({decl.declStart, decl.declLength,
+                                     BuildImageDeclaration(decl, "readonly", decl.name)});
+                }
+            }
+            for (const StoreSite& site : storeSites) {
+                const ImageUniformDecl& decl = decls[site.declIndex];
+                if (!decl.split) continue;
+                edits.push_back({site.start, site.length, decl.writeName});
+            }
+            if (edits.empty()) {
+                return glslCode;
+            }
+
+            // Back to front, so an earlier edit's offsets stay valid.
+            std::sort(edits.begin(), edits.end(),
+                      [](const ImageSourceEdit& a, const ImageSourceEdit& b) { return a.start > b.start; });
+            String result = glslCode;
+            for (const ImageSourceEdit& edit : edits) {
+                result.replace(edit.start, edit.length, edit.text);
+            }
+            return result;
+        }
+
         namespace {
             // How a lookup carries its level of detail, and how many arguments it takes
             // before the optional bias.
@@ -527,7 +957,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
             }
         } // namespace
 
-        String EmulateTextureLodBias(const String& glslCode) {
+        String EmulateTextureLodBias(const String& glslCode, Bool avoidExplicitLodBias) {
 #ifdef TRACY_ENABLE
             ZoneScopedC(TRACY_ZONECOLOR_BACKEND);
 #endif
@@ -588,6 +1018,11 @@ namespace MobileGL::MG_Backend::DirectGLES {
                 if (samplerIt == samplerNames.end()) continue;
 
                 const String& biasName = samplerIt->second;
+                if (form->explicitLodArg >= 0 && avoidExplicitLodBias) {
+                    // The lookup already names its level; leaving it alone keeps a constant
+                    // LOD constant. Costs the bias on explicit-LOD lookups only.
+                    continue;
+                }
                 if (form->explicitLodArg >= 0) {
                     // Explicit LOD: the bias adds to it, as Vulkan does for
                     // OpImageSampleExplicitLod and as the CTS reference expects.

MobileGL/MG_Backend/DirectGLES/Utils.h

@@ -130,7 +130,48 @@ namespace MobileGL::MG_Backend::DirectGLES {
         // drawBufferCount <= 1, i.e. for everything but a framebuffer that actually
         // enables several draw buffers, so the ordinary single-target shader is untouched.
         String BroadcastLegacyFragColor(String glslCode, GLenum shaderType, Uint drawBufferCount);
+        // SPIRV-Cross emits `#extension GL_EXT_texture_buffer : require` for every buffer-texture
+        // sampler when it targets ESSL below 320, and offers no way to ask for the OES spelling.
+        // On a driver that advertises only GL_OES_texture_buffer that directive is a compile
+        // error, so the name is retargeted in the emitted source. A no-op on every other tier:
+        // ES 3.2 needs no directive at all and an EXT driver already has the right one.
+        String RetargetTextureBufferExtension(String glslCode,
+                                              MG_External::GLESCapabilities::TextureBufferTier tier);
         String RemoveLayoutBinding(const String& glslCode);
+        // Prefix of the writeonly half a read+write image uniform is split into (see
+        // SplitReadWriteImageUniforms); the suffix is the image's own name.
+        constexpr const char* IMAGE_WRITE_ALIAS_PREFIX = "mg_imageWrite_";
+        // ESSL refuses an image variable that carries a format qualifier other than r32f /
+        // r32i / r32ui unless it also carries `readonly` or `writeonly` (GLSL ES 3.10 4.9 /
+        // 3.20 4.10; glslang enforces it verbatim in ParseHelper.cpp's layoutObjectCheck).
+        // SPIRV-Cross emits NEITHER for an image the shader both reads and writes: it
+        // speculatively decorates every storage image NonWritable+NonReadable
+        // (fixup_image_load_store_access), then OpImageRead clears NonReadable and
+        // OpImageWrite clears NonWritable, and to_qualifiers_glsl only prints `readonly`
+        // from NonWritable and `writeonly` from NonReadable. Desktop GLSL is happy with the
+        // bare declaration, so the frontend raises no error and the illegal ESSL only shows
+        // up as a device compile failure - and then as a silently no-op draw.
+        //
+        // Restores a legal declaration:
+        //  * loaded only            -> add `readonly`
+        //  * stored only            -> add `writeonly`
+        //  * both                   -> emit TWO declarations on the same binding and of the
+        //                              same type, `readonly <name>` and `writeonly
+        //                              <IMAGE_WRITE_ALIAS_PREFIX><name>`, and point every
+        //                              imageStore at the second one. Several image variables
+        //                              may share an image unit as long as they have the same
+        //                              type and format, which is exactly what the pair is.
+        //
+        // Budget note: the split DOUBLES the image-uniform count of the stage it fires in, so
+        // a driver advertising a tight GL_MAX_{FRAGMENT,VERTEX,...}_IMAGE_UNIFORMS can turn a
+        // shader that used to compile into a link failure. ES only guarantees 4 fragment image
+        // uniforms, so a shader with more than half the limit in read+write images is the case
+        // to watch.
+        //
+        // Runs on the transpiled ESSL, so it must see the bindings the frontend units were
+        // already rewritten to and must run before those bindings are stripped - see the call
+        // site in Managers.cpp.
+        String SplitReadWriteImageUniforms(const String& glslCode);
         // Prefix of the per-sampler float uniform that carries GL_TEXTURE_LOD_BIAS into
         // the shader (see EmulateTextureLodBias); the suffix is the sampler's own name.
         constexpr const char* LOD_BIAS_UNIFORM_PREFIX = "mg_lodBias_";
@@ -142,7 +183,12 @@ namespace MobileGL::MG_Backend::DirectGLES {
         // the bound texture's (or sampler object's) value into it; a shader whose samplers
         // all have a zero bias is therefore unaffected. Returns the source unchanged when
         // there is nothing to rewrite.
-        String EmulateTextureLodBias(const String& glslCode);
+        //
+        // avoidExplicitLodBias leaves lookups that already carry an explicit LOD untouched,
+        // so their constant level stays constant; only the implicit-LOD forms take the bias.
+        // Off by default and only ever set on ANGLE + llvmpipe, where injecting the uniform
+        // into a constant LOD crashes the driver (MOBILEGL_AVOID_EXPLICIT_LOD_BIAS).
+        String EmulateTextureLodBias(const String& glslCode, Bool avoidExplicitLodBias = false);
     } // namespace PrgramImpl
 
     namespace Utils {

MobileGL/MG_Backend/DirectVulkan/BackendObject_DirectVulkan.cpp

@@ -768,14 +768,48 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // the Uint32 attribute masks the draw path passes around are both bounded by MAX_VERTEX_ATTRIBS.
         m_dynamicParameters.MaxVertexAttribs = std::min(
             m_vulkanCaps.MaxVertexAttribs, static_cast<Int>(MG_State::GLState::VertexArrayObject::MAX_VERTEX_ATTRIBS));
-        m_dynamicParameters.MaxComputeShaderStorageBlocks = m_vulkanCaps.MaxComputeShaderStorageBlocks;
-        m_dynamicParameters.MaxCombinedShaderStorageBlocks = m_vulkanCaps.MaxCombinedShaderStorageBlocks;
-        m_dynamicParameters.MaxComputeUniformBlocks = m_vulkanCaps.MaxComputeUniformBlocks;
+        // Vulkan descriptor limits are not GL limits, and a GL application reads an advertised
+        // limit as an amount it may actually USE. Adreno answers the per-stage/per-set descriptor
+        // queries at descriptor-indexing scale - the same driver whose
+        // GL_MAX_SHADER_STORAGE_BLOCK_SIZE is clamped from 2147483647 further down - so
+        // KHR-GL44.multi_bind.dispatch_bind_buffers_base read GL_MAX_COMPUTE_UNIFORM_BLOCKS,
+        // created that many buffers and spliced that many UBO declarations into a single compute
+        // shader: ~14 s of allocation, then death on std::bad_alloc. Its sibling
+        // dispatch_bind_buffers_range hard-codes 4 buffers and passes, which is the clean
+        // discriminator. Every ceiling below is far above what any desktop driver advertises for
+        // these (84-96 for the binding families) and far below a descriptor-indexing count, so it
+        // can only lower a limit that was never usable in the first place. The zero floor is not
+        // decoration: a driver reporting UINT32_MAX used to arrive here as -1.
+        const auto clampLimit = [](const char* name, Int reported, Int ceiling) {
+            const Int clamped = std::min(std::max(reported, 0), ceiling);
+            if (clamped != reported) {
+                MGLOG_I("DirectVulkan: clamped %s from %d to %d", name, reported, clamped);
+            }
+            return clamped;
+        };
+        // GL 4.6 required minimums, for the record: MAX_COMPUTE_UNIFORM_BLOCKS 12,
+        // MAX_COMPUTE/COMBINED_SHADER_STORAGE_BLOCKS 8, MAX_SHADER_STORAGE_BUFFER_BINDINGS 8,
+        // MAX_UNIFORM_BUFFER_BINDINGS 84, MAX_TEXTURE_BUFFER_SIZE 65536.
+        constexpr Int kMaxAdvertisedBufferBlocks = 256;
+        constexpr Int kMaxAdvertisedTextureBufferSize = 1 << 27; // texels; what desktop GL reports
+        m_dynamicParameters.MaxComputeShaderStorageBlocks =
+            clampLimit("GL_MAX_COMPUTE_SHADER_STORAGE_BLOCKS", m_vulkanCaps.MaxComputeShaderStorageBlocks,
+                       kMaxAdvertisedBufferBlocks);
+        m_dynamicParameters.MaxCombinedShaderStorageBlocks =
+            clampLimit("GL_MAX_COMBINED_SHADER_STORAGE_BLOCKS", m_vulkanCaps.MaxCombinedShaderStorageBlocks,
+                       kMaxAdvertisedBufferBlocks);
+        m_dynamicParameters.MaxComputeUniformBlocks =
+            clampLimit("GL_MAX_COMPUTE_UNIFORM_BLOCKS", m_vulkanCaps.MaxComputeUniformBlocks,
+                       kMaxAdvertisedBufferBlocks);
         m_dynamicParameters.MaxComputeWorkGroupInvocations = m_vulkanCaps.MaxComputeWorkGroupInvocations;
-        m_dynamicParameters.MaxShaderStorageBufferBindings = m_vulkanCaps.MaxShaderStorageBufferBindings;
-        m_dynamicParameters.MaxTextureBufferSize = m_vulkanCaps.MaxTextureBufferSize;
+        m_dynamicParameters.MaxShaderStorageBufferBindings =
+            clampLimit("GL_MAX_SHADER_STORAGE_BUFFER_BINDINGS", m_vulkanCaps.MaxShaderStorageBufferBindings,
+                       kMaxAdvertisedBufferBlocks);
+        m_dynamicParameters.MaxTextureBufferSize = clampLimit(
+            "GL_MAX_TEXTURE_BUFFER_SIZE", m_vulkanCaps.MaxTextureBufferSize, kMaxAdvertisedTextureBufferSize);
         m_dynamicParameters.TextureBufferOffsetAlignment = m_vulkanCaps.TextureBufferOffsetAlignment;
-        m_dynamicParameters.MaxUniformBufferBindings = m_vulkanCaps.MaxUniformBufferBindings;
+        m_dynamicParameters.MaxUniformBufferBindings = clampLimit(
+            "GL_MAX_UNIFORM_BUFFER_BINDINGS", m_vulkanCaps.MaxUniformBufferBindings, kMaxAdvertisedBufferBlocks);
         m_dynamicParameters.MaxUniformBlockSize = m_vulkanCaps.MaxUniformBlockSize;
         m_dynamicParameters.MaxImageUnits = std::max(std::min(m_vulkanCaps.MaxImageUnits, maxSupportedTextureUnits), 0);
         m_dynamicParameters.MaxCombinedImageUniforms = std::max(m_vulkanCaps.MaxCombinedImageUniforms, 0);

MobileGL/MG_Backend/DirectVulkan/DirectVulkan.cpp

@@ -801,9 +801,18 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // already recorded the new binding on the program - which is what reseeds this cache
         // whenever it is rebuilt. Writing the entry here as well keeps an ALREADY-BUILT cache
         // (the common case: the very next draw reads it) from having to be thrown away.
-        auto& cache = GetProgramResourceCache(*programObject);
+        //
+        // Resolve the index BEFORE taking the reference, and bounds-check the way the
+        // sibling getter does. GetShaderStorageBlockIndex re-enters GetProgramResourceCache,
+        // which indexes g_programResourceCaches and can therefore insert - and that map is
+        // open-addressed, so a rehash MOVES its entries and a reference taken before the
+        // call is left dangling. Binding a program's storage block
+        // while another program's entry was still absent from the cache was a reproducible
+        // segfault (ProgramPipelineScenario's two storage-block cases, in one process).
         const GLuint blockIndex = GetShaderStorageBlockIndex(*programObject, storageBlockName);
         if (blockIndex == GL_INVALID_INDEX) return;
+        auto& cache = GetProgramResourceCache(*programObject);
+        if (blockIndex >= cache.storageBlocks.size()) return;
         cache.storageBlocks[blockIndex].binding = storageBlockBinding;
     }
     void ReadPixels(GLint x, GLint y, GLsizei width, GLsizei height, GLenum format, GLenum type, void* pixels) {
@@ -966,6 +975,29 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         if (drawcount <= 0) {
             return;
         }
+
+        // With no element-array buffer bound, every indices[i] is a client pointer into a
+        // separate CPU allocation, not an offset into one shared buffer. The batched payload
+        // below cannot express that: it carries ONE index-buffer view for the whole batch and
+        // turns each pointer into a firstIndex relative to it. Replay the sub-draws through
+        // the single-draw entry point instead - it snapshots each client range into its own
+        // transient slice, which is exactly what the unrolled draws this must match do.
+        // (The batch used to be built this way; the shared-view rewrite that added
+        // MultiDrawIndexedCmd left the client-memory shape addressing a view whose byte
+        // offset is a hardcoded 0, so UploadAndBindIndexBuffer saw a null client pointer,
+        // declined the whole batch and painted nothing.)
+        const auto& vao = *MG_State::pGLContext->GetBoundVertexArray();
+        if (vao.GetIndexBufferBindingSlot().GetBoundObject() == nullptr) {
+            for (GLsizei i = 0; i < drawcount; ++i) {
+                if (count[i] <= 0) {
+                    continue;
+                }
+                DrawElementsBaseVertex(mode, count[i], type, indices[i],
+                                       basevertex != nullptr ? basevertex[i] : 0);
+            }
+            return;
+        }
+
         MultiDrawIndexedCmd payload{};
         payload.mode = mode;
         payload.indexBufferView.indexType = type;

MobileGL/MG_Backend/DirectVulkan/Renderer/BufferArena.cpp

@@ -111,6 +111,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         }
 
         if (buffer.IsValid()) {
+            // Outgrown, not dead: every BufferSlice handed out from this frame's arena so far
+            // still names it, and those slices stay in service until the frame slot is rewound
+            // (VkBufferResource::transientSlice, the converted-vertex-stream cache, the draw
+            // memos). The release therefore has to survive every mid-frame reclaim and land on
+            // the next ResetFrame of this slot - see VkBufferManager::CollectAllDeferredReleases.
             m_deferredReleases[frameIndex].push_back(std::move(buffer));
         }
 

MobileGL/MG_Backend/DirectVulkan/Renderer/PipelineFactory.cpp

@@ -252,6 +252,19 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         }
 
         VkPipeline pipeline = CreatePipeline(payload);
+        // A failed creation must never be memoized. Caching VK_NULL_HANDLE served the null back for
+        // the rest of the process, so one transient driver rejection turned every later draw with
+        // the same state into a vkCmdBindPipeline(VK_NULL_HANDLE) - the SIGSEGV behind 9 of the 15
+        // CTS process deaths. Retrying costs one failed vkCreateGraphicsPipelines per draw, which
+        // is the correct price for a broken pipeline and is bounded by the draw itself being
+        // skipped.
+        if (pipeline == VK_NULL_HANDLE) {
+            MGLOG_I("PipelineFactory::GetOrCreatePipeline: creation failed for hash=0x%llx "
+                    "programHash=0x%llx; not caching the failure",
+                    static_cast<unsigned long long>(hash),
+                    static_cast<unsigned long long>(payload.programHash));
+            return VK_NULL_HANDLE;
+        }
         m_cache.emplace(hash, PipelineCacheEntry{pipeline, payload.programHash, payload.renderPass,
                                                  m_frameCounter});
         return pipeline;
@@ -507,6 +520,35 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             MGLOG_F("PipelineFactory::CreatePipeline vertex input: bindingCount=%u attributeCount=%u",
                     payload.vertexInputState->vertexBindingDescriptionCount,
                     payload.vertexInputState->vertexAttributeDescriptionCount);
+            // The driver's own answer is VK_ERROR_UNKNOWN, i.e. no information at all, so the only
+            // way to work out WHICH shader it choked on (the open sampler-array-in-struct
+            // investigation) is to name the modules. MGLOG_I, not _D/_E: this must survive in the
+            // INFO-level builds that CTS actually runs against.
+            if (payload.stageSpirvDigests) {
+                for (SizeT i = 0; i < payload.stageSpirvDigests->size(); ++i) {
+                    const auto& digest = (*payload.stageSpirvDigests)[i];
+                    MGLOG_I("PipelineFactory::CreatePipeline spirv[%zu]: stage=0x%x words=%u bytes=%zu "
+                            "hash=0x%llx",
+                            i, digest.stage, digest.wordCount,
+                            static_cast<SizeT>(digest.wordCount) * sizeof(Uint32),
+                            static_cast<unsigned long long>(digest.hash));
+                }
+            } else {
+                MGLOG_I("PipelineFactory::CreatePipeline: no SPIR-V digests attached to the payload");
+            }
+            if (payload.stages) {
+                for (SizeT i = 0; i < payload.stages->size(); ++i) {
+                    const auto& stage = (*payload.stages)[i];
+                    // VkShaderModule is a non-dispatchable handle: a pointer on 64-bit but a
+                    // plain uint64_t on 32-bit ABIs, where a cast to const void* is ill-formed
+                    // (broke the armeabi-v7a build). Print it as the 64-bit value it is.
+                    MGLOG_I("PipelineFactory::CreatePipeline stage[%zu]: stage=0x%x module=0x%llx entry=%s "
+                            "specialization=%d",
+                            i, static_cast<Uint32>(stage.stage),
+                            static_cast<unsigned long long>(reinterpret_cast<Uint64>(stage.module)),
+                            stage.pName ? stage.pName : "(null)", stage.pSpecializationInfo ? 1 : 0);
+                }
+            }
             for (Uint32 i = 0; i < payload.colorAttachmentCount; ++i) {
                 const auto& attachment = payload.colorBlendAttachments[i];
                 MGLOG_F("PipelineFactory::CreatePipeline colorAttachment[%u]: blend=%d colorWriteMask=0x%x srcColor=%d dstColor=%d colorOp=%d srcAlpha=%d dstAlpha=%d alphaOp=%d",

MobileGL/MG_Backend/DirectVulkan/Renderer/PipelineFactory.h

@@ -14,6 +14,16 @@
 #include <Includes.h>
 
 namespace MobileGL::MG_Backend::DirectVulkan {
+    // Enough of a fingerprint to identify the exact module the driver rejected without keeping the
+    // SPIR-V alive for every program in the cache: a driver that answers VK_ERROR_UNKNOWN tells us
+    // nothing, so the log has to carry the shader's identity itself. Diagnostic only - never part
+    // of any pipeline or program hash.
+    struct ShaderStageSpirvDigest {
+        Uint32 stage = 0; // VkShaderStageFlagBits
+        Uint32 wordCount = 0;
+        Uint64 hash = 0;
+    };
+
     class PipelineFactory {
     public:
         using HashType = Uint64;
@@ -62,6 +72,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             Array<VkPipelineColorBlendAttachmentState, kMaxColorAttachments> colorBlendAttachments{};
             const Vector<VkPipelineShaderStageCreateInfo>* stages = nullptr;
             const VkPipelineVertexInputStateCreateInfo* vertexInputState = nullptr;
+            // Diagnostic only; may be null. Read solely from the pipeline-creation failure path.
+            const Vector<ShaderStageSpirvDigest>* stageSpirvDigests = nullptr;
         };
 
         explicit PipelineFactory(VkDevice device, const VulkanRendererConfig& config);

MobileGL/MG_Backend/DirectVulkan/Renderer/ProgramFactory.cpp

@@ -12,7 +12,10 @@
 #include "MG_Util/ShaderTranspiler/ShaderCompiler.h"
 #include "MG_Util/ShaderTranspiler/SpvcSession.h"
 #include "MG_Util/ShaderTranspiler/Types.h"
+#include <algorithm>
 #include <cstring>
+#include <map>
+#include <utility>
 #include <spirv-tools/libspirv.h>
 #include <spirv-tools/optimizer.hpp>
 #include <source/opt/build_module.h>
@@ -372,6 +375,20 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
             spv_diagnostic diagnostic = nullptr;
             const spv_result_t result = spvValidateWithOptions(context, options, &binary, &diagnostic);
+            if (result != SPV_SUCCESS) {
+                // MGLOG_I, not E: at the INFO compile level of the CI/test lanes that arm
+                // the validation switch, MGLOG_E is compiled out (Log.h orders
+                // DEBUG < WARN < ERROR < INFO) and the VUID would never reach a log. The
+                // latch is what a test harness asserts on.
+                MG_Util::ShaderTranspiler::ShaderCompiler::NoteSpirvValidationFailure();
+                MGLOG_I(
+                    "ProgramFactory::ValidateTransformedSpirv: validation failed for stage=%d program=%u result=%d index=%zu msg=%s",
+                    static_cast<Int>(shaderStage),
+                    programExternalIndex,
+                    static_cast<Int>(result),
+                    diagnostic != nullptr ? diagnostic->position.index : 0,
+                    diagnostic != nullptr && diagnostic->error != nullptr ? diagnostic->error : "<null>");
+            }
             MOBILEGL_ASSERT(
                 result == SPV_SUCCESS,
                 "ProgramFactory::ValidateTransformedSpirv: validation failed for stage=%d program=%u result=%d line=%zu column=%zu index=%zu msg=%s",
@@ -923,6 +940,189 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             ProgramFactory::CompileOptionFlags m_transformFlags;
         };
 
+        // gl_FragCoord back into GL's window space, for default-framebuffer draws only.
+        //
+        // Vulkan's gl_FragCoord.y is the framebuffer ROW being written - not a value the
+        // viewport rect can move independently of placement. The default framebuffer's image is
+        // stored display-side-up and the vertex stage compensates by negating gl_Position.y, so
+        // for every default-FBO draw the framebuffer row of a fragment is exactly
+        // `height - y_GL` (the viewport terms cancel: yf_VK = H - yf_GL for any viewport rect).
+        // A shader that reads gl_FragCoord therefore sees a flipped Y, and once the viewport
+        // rect started being converted to the stored orientation it also sees a Y that is
+        // OUTSIDE the range GL promises - a 32-pixel-tall viewport at GL y=0 reports 224..255 on
+        // a 256-tall surface. GL CTS shader_image_load_store writes imageStore(image,
+        // ivec2(gl_FragCoord.xy)) into an image exactly the size of that viewport, so every
+        // store fell outside the image and the test read back zeroes.
+        //
+        // The rewrite redirects every read of the builtin to a Private copy initialised once at
+        // entry, which is exact for all access forms (whole-vector loads, `.y` access chains,
+        // OpCopyMemory) and leaves the builtin itself - and its decorations - untouched.
+        class GlFragCoordYFlipPass final : public spvtools::opt::Pass {
+        public:
+            const char* name() const override { return "mobilegl-fragcoord-y-flip"; }
+            explicit GlFragCoordYFlipPass(Uint32 framebufferHeight) : m_framebufferHeight(framebufferHeight) {}
+
+            Status Process() override {
+                using namespace spvtools::opt;
+                if (m_framebufferHeight == 0) return Status::SuccessWithoutChange;
+
+                Instruction* entryPoint = nullptr;
+                for (auto& candidate : get_module()->entry_points()) {
+                    if (candidate.NumInOperands() >= 2 &&
+                        static_cast<spv::ExecutionModel>(candidate.GetSingleWordInOperand(0)) ==
+                            spv::ExecutionModel::Fragment) {
+                        entryPoint = &candidate;
+                        break;
+                    }
+                }
+                if (!entryPoint) return Status::SuccessWithoutChange;
+
+                const Uint32 builtinVarId = FindFragCoordVariable();
+                if (builtinVarId == 0) return Status::SuccessWithoutChange;
+
+                Instruction* builtinVar = context()->get_def_use_mgr()->GetDef(builtinVarId);
+                if (!builtinVar || builtinVar->opcode() != spv::Op::OpVariable) return Status::SuccessWithoutChange;
+
+                // The builtin is `Input vec4`; take the vector and component types from its own
+                // pointer type rather than assuming float32x4, so a module that spells it
+                // differently declines instead of miscompiling.
+                Instruction* inputPtrType = context()->get_def_use_mgr()->GetDef(builtinVar->type_id());
+                if (!inputPtrType || inputPtrType->opcode() != spv::Op::OpTypePointer) {
+                    return Status::SuccessWithoutChange;
+                }
+                const Uint32 vectorTypeId = inputPtrType->GetSingleWordInOperand(1);
+                Instruction* vectorType = context()->get_def_use_mgr()->GetDef(vectorTypeId);
+                if (!vectorType || vectorType->opcode() != spv::Op::OpTypeVector ||
+                    vectorType->GetSingleWordInOperand(1) != 4) {
+                    return Status::SuccessWithoutChange;
+                }
+                const Uint32 floatTypeId = vectorType->GetSingleWordInOperand(0);
+                auto* floatType = context()->get_type_mgr()->GetType(floatTypeId);
+                if (!floatType || !floatType->AsFloat() || floatType->AsFloat()->width() != 32) {
+                    return Status::SuccessWithoutChange;
+                }
+
+                const auto heightBits = std::bit_cast<Uint32>(static_cast<float>(m_framebufferHeight));
+                const auto* heightConst = context()->get_constant_mgr()->GetConstant(floatType, {heightBits});
+                auto* heightInst = context()->get_constant_mgr()->GetDefiningInstruction(heightConst);
+                if (!heightInst) return Status::SuccessWithoutChange;
+
+                auto* function = context()->GetFunction(entryPoint->GetSingleWordInOperand(1));
+                if (!function || function->begin() == function->end()) return Status::SuccessWithoutChange;
+
+                const Uint32 privatePtrTypeId =
+                    context()->get_type_mgr()->FindPointerToType(vectorTypeId, spv::StorageClass::Private);
+                if (privatePtrTypeId == 0) return Status::SuccessWithoutChange;
+
+                const Uint32 copyVarId = context()->TakeNextId();
+                if (copyVarId == 0) return Status::SuccessWithoutChange;
+                auto copyVar = std::make_unique<Instruction>(
+                    context(), spv::Op::OpVariable, privatePtrTypeId, copyVarId,
+                    std::initializer_list<Operand>{
+                        {SPV_OPERAND_TYPE_STORAGE_CLASS, {static_cast<Uint32>(spv::StorageClass::Private)}}});
+                context()->AddGlobalValue(std::move(copyVar));
+
+                // Redirect the reads BEFORE emitting the initialiser, so the initialiser's own
+                // load of the builtin is not rewritten into a load of the (still empty) copy.
+                if (!RedirectReads(builtinVarId, copyVarId)) return Status::SuccessWithoutChange;
+
+                auto& entryBlock = *function->begin();
+                auto insertPoint = entryBlock.begin();
+                while (insertPoint != entryBlock.end() && insertPoint->opcode() == spv::Op::OpVariable) {
+                    ++insertPoint;
+                }
+                if (insertPoint == entryBlock.end()) return Status::SuccessWithoutChange;
+
+                InstructionBuilder builder(context(), &*insertPoint,
+                                           IRContext::kAnalysisDefUse | IRContext::kAnalysisInstrToBlockMapping);
+                auto* raw = builder.AddLoad(vectorTypeId, builtinVarId);
+                if (!raw) return Status::SuccessWithoutChange;
+                auto* x = builder.AddCompositeExtract(floatTypeId, raw->result_id(), {0});
+                auto* y = builder.AddCompositeExtract(floatTypeId, raw->result_id(), {1});
+                auto* z = builder.AddCompositeExtract(floatTypeId, raw->result_id(), {2});
+                auto* w = builder.AddCompositeExtract(floatTypeId, raw->result_id(), {3});
+                if (!x || !y || !z || !w) return Status::SuccessWithoutChange;
+                auto* flippedY =
+                    builder.AddBinaryOp(floatTypeId, spv::Op::OpFSub, heightInst->result_id(), y->result_id());
+                if (!flippedY) return Status::SuccessWithoutChange;
+                auto* corrected = builder.AddCompositeConstruct(
+                    vectorTypeId, {x->result_id(), flippedY->result_id(), z->result_id(), w->result_id()});
+                if (!corrected) return Status::SuccessWithoutChange;
+                if (!builder.AddStore(copyVarId, corrected->result_id())) return Status::SuccessWithoutChange;
+
+                // SPIR-V 1.4 widened the entry-point interface to every global the entry point
+                // statically uses, Private included; earlier versions accept Input/Output only,
+                // so listing it there would be invalid.
+                if (get_module()->version() >= 0x00010400u) {
+                    entryPoint->AddOperand({SPV_OPERAND_TYPE_ID, {copyVarId}});
+                    context()->AnalyzeUses(entryPoint);
+                }
+
+                context()->InvalidateAnalysesExceptFor(spvtools::opt::IRContext::kAnalysisDefUse |
+                                                       spvtools::opt::IRContext::kAnalysisInstrToBlockMapping);
+                return Status::SuccessWithChange;
+            }
+
+        private:
+            Uint32 FindFragCoordVariable() const {
+                for (const auto& annotation : get_module()->annotations()) {
+                    if (annotation.opcode() != spv::Op::OpDecorate) continue;
+                    if (annotation.NumInOperands() < 3) continue;
+                    if (static_cast<spv::Decoration>(annotation.GetSingleWordInOperand(1)) !=
+                        spv::Decoration::BuiltIn) {
+                        continue;
+                    }
+                    if (static_cast<spv::BuiltIn>(annotation.GetSingleWordInOperand(2)) != spv::BuiltIn::FragCoord) {
+                        continue;
+                    }
+                    return annotation.GetSingleWordInOperand(0);
+                }
+                return 0;
+            }
+
+            // Every instruction that reads through the builtin's POINTER gets the copy instead.
+            // Decorations, names and the entry-point interface keep naming the builtin.
+            Bool RedirectReads(Uint32 builtinVarId, Uint32 copyVarId) {
+                using namespace spvtools::opt;
+                Bool ok = true;
+                Vector<Instruction*> users;
+                context()->get_def_use_mgr()->ForEachUser(builtinVarId, [&](Instruction* user) {
+                    switch (user->opcode()) {
+                    case spv::Op::OpLoad:
+                    case spv::Op::OpAccessChain:
+                    case spv::Op::OpInBoundsAccessChain:
+                    case spv::Op::OpPtrAccessChain:
+                    case spv::Op::OpInBoundsPtrAccessChain:
+                    case spv::Op::OpCopyMemory:
+                    case spv::Op::OpCopyMemorySized:
+                        users.push_back(user);
+                        break;
+                    case spv::Op::OpStore:
+                        // gl_FragCoord is read-only; a store through it means this is not the
+                        // module we think it is.
+                        ok = false;
+                        break;
+                    default:
+                        break;
+                    }
+                });
+                if (!ok) return false;
+                for (Instruction* user : users) {
+                    for (Uint32 i = 0; i < user->NumInOperands(); ++i) {
+                        auto& operand = user->GetInOperand(i);
+                        if (operand.type == SPV_OPERAND_TYPE_ID && !operand.words.empty() &&
+                            operand.words[0] == builtinVarId) {
+                            operand.words[0] = copyVarId;
+                        }
+                    }
+                    context()->AnalyzeUses(user);
+                }
+                return true;
+            }
+
+            Uint32 m_framebufferHeight = 0;
+        };
+
         // Decorates the module's captured varyings for VK_EXT_transform_feedback:
         // user outputs get XfbBuffer/XfbStride/Offset directly; a captured
         // gl_Position (a gl_PerVertex member) is mirrored into a dedicated output
@@ -934,6 +1134,15 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 std::string name;
                 Uint32 bufferIndex = 0;
                 Uint32 offsetBytes = 0;
+                // Set when the capture names a member of an output interface block
+                // ("Block.member"): the decoration target is then the block's struct TYPE,
+                // decorated per member, not the variable. `name` keeps the GL spelling and
+                // is useless for the id lookup, so the instance name is carried separately.
+                std::string blockInstanceName;
+                std::string blockName;
+                Int blockMemberIndex = -1;
+                Int blockMemberElement = -1; // array element of that member, -1 = the whole member
+                Uint32 byteSize = 0;
             };
             const char* name() const override { return "mobilegl-xfb-capture-decorate"; }
             XfbCaptureDecoratePass(Vector<CapturedVarying> varyings, Vector<Uint32> strides)
@@ -965,6 +1174,33 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                     decorationManager->AddDecorationVal(targetId, static_cast<Uint32>(spv::Decoration::Offset),
                                                         offsetBytes);
                 };
+                // SPIR-V puts XfbBuffer/XfbStride/Offset on the struct MEMBER when the
+                // captured varying lives in an interface block (SPIR-V 1.6 §3.20 lists all
+                // three as member-decoratable); Offset in particular is illegal on the block
+                // variable once the type is decorated Block.
+                const auto decorateMemberForXfb = [&](Uint32 structTypeId, Uint32 memberIndex, Uint32 bufferIndex,
+                                                      Uint32 offsetBytes) {
+                    const Uint32 stride = bufferIndex < m_strides.size() ? m_strides[bufferIndex] : 0;
+                    decorationManager->AddMemberDecoration(structTypeId, memberIndex,
+                                                           static_cast<Uint32>(spv::Decoration::XfbBuffer),
+                                                           bufferIndex);
+                    decorationManager->AddMemberDecoration(structTypeId, memberIndex,
+                                                           static_cast<Uint32>(spv::Decoration::XfbStride), stride);
+                    decorationManager->AddMemberDecoration(structTypeId, memberIndex,
+                                                           static_cast<Uint32>(spv::Decoration::Offset), offsetBytes);
+                };
+
+                // A member array captured element by element ("Block.attrib[0]" .. "[15]")
+                // is one SPIR-V member, so its captures collapse into a single decoration
+                // placed at the first element's offset - the rest follow from the member's
+                // own layout. Collected first so the group is complete before it decorates.
+                struct MemberGroup {
+                    Uint32 bufferIndex = 0;
+                    Uint32 minOffset = 0;
+                    Uint32 elementBytes = 0;
+                    Vector<Uint32> offsets;
+                };
+                std::map<std::pair<Uint32, Uint32>, MemberGroup> memberGroups;
 
                 Bool modified = false;
                 Bool needsPositionMirror = false;
@@ -977,6 +1213,41 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                         positionOffset = varying.offsetBytes;
                         continue;
                     }
+                    if (varying.blockMemberIndex >= 0) {
+                        // glslang names the block's instance variable and its struct type
+                        // separately; an anonymous instance leaves only the type named, so
+                        // both spellings are tried before giving up.
+                        Uint32 structTypeId = 0;
+                        if (const auto it = idsByName.find(varying.blockInstanceName); it != idsByName.end()) {
+                            structTypeId = BlockStructTypeOf(it->second);
+                        }
+                        if (structTypeId == 0) {
+                            if (const auto it = idsByName.find(varying.blockName); it != idsByName.end()) {
+                                const spvtools::opt::Instruction* def = context()->get_def_use_mgr()->GetDef(it->second);
+                                if (def != nullptr && def->opcode() == spv::Op::OpTypeStruct) {
+                                    structTypeId = it->second;
+                                } else if (def != nullptr && def->opcode() == spv::Op::OpVariable) {
+                                    structTypeId = BlockStructTypeOf(it->second);
+                                }
+                            }
+                        }
+                        if (structTypeId == 0) {
+                            MGLOG_E("XfbCaptureDecoratePass: no SPIR-V interface block '%s' (instance '%s') for "
+                                    "capture '%s'",
+                                    varying.blockName.c_str(), varying.blockInstanceName.c_str(),
+                                    varying.name.c_str());
+                            continue;
+                        }
+                        auto& group =
+                            memberGroups[{structTypeId, static_cast<Uint32>(varying.blockMemberIndex)}];
+                        if (group.offsets.empty() || varying.offsetBytes < group.minOffset) {
+                            group.minOffset = varying.offsetBytes;
+                        }
+                        group.bufferIndex = varying.bufferIndex;
+                        group.elementBytes = varying.byteSize;
+                        group.offsets.push_back(varying.offsetBytes);
+                        continue;
+                    }
                     const auto idIt = idsByName.find(varying.name);
                     if (idIt == idsByName.end()) {
                         MGLOG_E("XfbCaptureDecoratePass: no SPIR-V variable named '%s'", varying.name.c_str());
@@ -986,6 +1257,25 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                     modified = true;
                 }
 
+                for (auto& [key, group] : memberGroups) {
+                    // The single Offset can only stand for the whole group when the group's
+                    // captures are a gap-free ascending run - that is what SPIR-V lays the
+                    // member's elements out as. Anything else still gets a best-effort
+                    // decoration, but say so, because the capture layout will not match GL.
+                    std::sort(group.offsets.begin(), group.offsets.end());
+                    for (SizeT i = 1; i < group.offsets.size(); ++i) {
+                        if (group.elementBytes == 0 ||
+                            group.offsets[i] != group.offsets[i - 1] + group.elementBytes) {
+                            MGLOG_I("XfbCaptureDecoratePass: block member %u of type %%%u is captured with a "
+                                    "non-contiguous element set; the capture layout will differ from GL's",
+                                    key.second, key.first);
+                            break;
+                        }
+                    }
+                    decorateMemberForXfb(key.first, key.second, group.bufferIndex, group.minOffset);
+                    modified = true;
+                }
+
                 if (needsPositionMirror) {
                     modified |= MirrorPositionForCapture(entryFunctionId, *entryPoint, positionBufferIndex,
                                                          positionOffset, decorateForXfb);
@@ -1007,6 +1297,27 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             }
 
         private:
+            // The struct type an interface-block variable points at, peeling an array of
+            // block instances on the way. 0 when the id is not a block variable at all.
+            Uint32 BlockStructTypeOf(Uint32 variableId) {
+                auto* defUse = context()->get_def_use_mgr();
+                const spvtools::opt::Instruction* variable = defUse->GetDef(variableId);
+                if (variable == nullptr || variable->opcode() != spv::Op::OpVariable) return 0;
+                const spvtools::opt::Instruction* pointer = defUse->GetDef(variable->type_id());
+                if (pointer == nullptr || pointer->opcode() != spv::Op::OpTypePointer) return 0;
+                Uint32 pointeeId = pointer->GetSingleWordInOperand(1);
+                for (const spvtools::opt::Instruction* pointee = defUse->GetDef(pointeeId); pointee != nullptr;
+                     pointee = defUse->GetDef(pointeeId)) {
+                    if (pointee->opcode() == spv::Op::OpTypeStruct) return pointeeId;
+                    if (pointee->opcode() != spv::Op::OpTypeArray &&
+                        pointee->opcode() != spv::Op::OpTypeRuntimeArray) {
+                        return 0;
+                    }
+                    pointeeId = pointee->GetSingleWordInOperand(0);
+                }
+                return 0;
+            }
+
             template <typename DecorateFn>
             Bool MirrorPositionForCapture(Uint32 entryFunctionId, spvtools::opt::Instruction& entryPoint,
                                           Uint32 bufferIndex, Uint32 offsetBytes, const DecorateFn& decorateForXfb) {
@@ -1263,8 +1574,10 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             }
             spvtools::Optimizer optimizer(SPV_ENV_VULKAN_1_3);
             spvtools::OptimizerOptions options;
-            // Matches the position-fix pass: this build of spirv-tools asserts rather than
-            // reporting, so validation stays off in the shipping path.
+            // Always off: the optimizer's input validator conflates "input invalid" with
+            // "transform failed", and this call site fails open. Validating lanes check the
+            // FINAL module via ValidateTransformedSpirv, which latches instead of rerouting
+            // control flow.
             options.set_run_validator(false);
             optimizer.SetMessageConsumer([](spv_message_level_t, const char*, const spv_position_t&,
                                             const char* message) {
@@ -1285,6 +1598,35 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             return spvtools::Optimizer::PassToken(MakeUnique<GlToVulkanPositionFixPass>(transformFlags));
         }
 
+        Bool TransformSpirvForFragCoordYFlip(const Vector<Uint>& input, Vector<Uint>& output,
+                                             Uint32 framebufferHeight) {
+            if (input.empty()) {
+                output.clear();
+                return true;
+            }
+            if (framebufferHeight == 0) {
+                output = input;
+                return true;
+            }
+
+            spvtools::Optimizer optimizer(SPV_ENV_VULKAN_1_3);
+            spvtools::OptimizerOptions options;
+            options.set_run_validator(false); // see TransformSpirvForExplicitLod0Sampling
+            optimizer.SetMessageConsumer([](spv_message_level_t, const char*, const spv_position_t&,
+                                            const char* message) {
+                MGLOG_E("Vulkan: fragcoord y-flip pass: %s", message != nullptr ? message : "");
+            });
+            optimizer.RegisterPass(
+                spvtools::Optimizer::PassToken(MakeUnique<GlFragCoordYFlipPass>(framebufferHeight)));
+
+            const Bool success = optimizer.Run(input.data(), input.size(), &output, options);
+            if (!success) {
+                MGLOG_E("Vulkan: failed to run the gl_FragCoord y-flip pass; keeping the original module");
+                output = input;
+            }
+            return success;
+        }
+
         Bool TransformSpirvForXfbCapture(const Vector<Uint>& input, Vector<Uint>& output,
                                          const MG_State::GLState::ProgramObject& program) {
             if (input.empty()) {
@@ -1294,7 +1636,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             Vector<XfbCaptureDecoratePass::CapturedVarying> varyings;
             varyings.reserve(program.GetTransformFeedbackVaryingCount());
             for (const auto& varying : program.GetTransformFeedbackVaryings()) {
-                varyings.push_back({varying.name, varying.bufferIndex, varying.offsetBytes});
+                varyings.push_back({varying.name, varying.bufferIndex, varying.offsetBytes,
+                                    varying.blockInstanceName, varying.blockName, varying.blockMemberIndex,
+                                    varying.blockMemberElement, varying.byteSize});
             }
             Vector<Uint32> strides;
             strides.reserve(program.GetTransformFeedbackBufferCount());
@@ -1304,7 +1648,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
             spvtools::Optimizer optimizer(SPV_ENV_VULKAN_1_3);
             spvtools::OptimizerOptions options;
-            options.set_run_validator(false);
+            options.set_run_validator(false); // see TransformSpirvForExplicitLod0Sampling
             optimizer.SetMessageConsumer([](spv_message_level_t, const char*, const spv_position_t&,
                                             const char* message) {
                 MGLOG_E("Vulkan: xfb capture pass: %s", message != nullptr ? message : "");
@@ -1334,7 +1678,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
             spvtools::Optimizer optimizer(SPV_ENV_VULKAN_1_3);
             spvtools::OptimizerOptions options;
-            options.set_run_validator(false);
+            options.set_run_validator(false); // see TransformSpirvForExplicitLod0Sampling
+            optimizer.SetMessageConsumer([](spv_message_level_t, const char*, const spv_position_t&,
+                                            const char* message) {
+                MGLOG_E("Vulkan: position fix pass: %s", message != nullptr ? message : "");
+            });
             optimizer.RegisterPass(CreateGlToVulkanPositionFixPass(transformFlags));
 
             const Bool success = optimizer.Run(input.data(), input.size(), &output, options);
@@ -1486,11 +1834,31 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 for (auto* binding : bindings) {
                     MOBILEGL_ASSERT(binding != nullptr, "ProgramFactory: null descriptor binding reflection record");
                     const auto kind = ReflectDescriptorTypeToBindingKind(binding->descriptor_type);
-                    // UBO instance arrays (uniform Block {...} b[N];) occupy one binding with
-                    // descriptorCount = N; other descriptor arrays stay unsupported and must
-                    // fail program creation cleanly rather than continue with corrupt state.
-                    if (binding->count != 1 && kind != ProgramFactory::DescriptorBindingKind::UniformBufferDynamic) {
-                        MGLOG_E("ProgramFactory: descriptor arrays are unsupported for this descriptor "
+                    // A descriptor ARRAY occupies one binding with descriptorCount = N, and is
+                    // supported for exactly the kinds that have a per-element resolve path in
+                    // UniformManager::BindProgramUniformBuffers: UBO instance arrays
+                    // (uniform Block {...} b[N];), storage-block instance arrays, image uniform
+                    // arrays, and combined-image-sampler arrays (uniform sampler2D s[N];).
+                    // Anything else - a uniform TEXEL buffer array is the one remaining kind -
+                    // must fail program creation cleanly rather than continue with corrupt state.
+                    //
+                    // Getting listed here is not cosmetic: a kind that is rejected leaves
+                    // GetOrCreateProgram's MOBILEGL_ASSERT(remapOk) as the only complaint, and
+                    // that assert compiles out above DEBUG - so a release build SILENTLY kept
+                    // glslang's per-stage auto-mapped binding numbers, skipping the cross-stage
+                    // unification and the set->0 normalisation this function exists to do. A
+                    // program with an image array plus any second descriptor got aliased
+                    // bindings out of that, and a DEBUG build trapped on the same program.
+                    // Which is also why the message below is MGLOG_I: MGLOG_E is compiled out
+                    // of an INFO build, so a refusal that only said MGLOG_E said nothing at all
+                    // in the builds that ship.
+                    const Bool arraySupportedForKind =
+                        kind == ProgramFactory::DescriptorBindingKind::UniformBufferDynamic ||
+                        kind == ProgramFactory::DescriptorBindingKind::StorageBuffer ||
+                        kind == ProgramFactory::DescriptorBindingKind::StorageImage ||
+                        kind == ProgramFactory::DescriptorBindingKind::CombinedImageSampler;
+                    if (binding->count != 1 && !arraySupportedForKind) {
+                        MGLOG_I("ProgramFactory: descriptor arrays are unsupported for this descriptor "
                                 "kind (name='%s' count=%u type=%d)",
                                 binding->name ? binding->name : "<null>", binding->count,
                                 static_cast<Int>(binding->descriptor_type));
@@ -1752,6 +2120,12 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             XXHASH_VERIFY(XXH64_update(m_hashState, spv.data(), spv.size() * sizeof(Uint)));
         }
         XXHASH_VERIFY(XXH64_update(m_hashState, &flags, sizeof(CompileOptionFlags)));
+        // Only FragCoordYFlip variants bake the height in, so mixing it unconditionally would
+        // re-key every program in the cache on a resize for no reason.
+        if (flags & CompileOptionBit::FragCoordYFlip) {
+            XXHASH_VERIFY(XXH64_update(m_hashState, &m_defaultFramebufferHeight,
+                                        sizeof(m_defaultFramebufferHeight)));
+        }
 
         // Include UBO block bindings in hash so different binding configurations produce different entries
         const Uint32 blockCount = static_cast<Uint32>(program.GetActiveUniformBlocksCount());
@@ -2017,6 +2391,78 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         }
     }
 
+    // How many descriptors to declare for an ARRAY of opaque uniforms (samplers, images) at one
+    // binding. A returned count is always DECLARED in the descriptor set layout; `outDeclined`
+    // says whether the binding can also be RESOLVED at draw time, or whether the program has to
+    // be refused instead.
+    //
+    // Those are deliberately two different things. The layout must keep describing what the
+    // shader declares even for a binding MobileGL cannot resolve: a descriptor the shader reads
+    // and the layout omits is not a missing draw, it is an undefined descriptor access, and
+    // lavapipe segfaults on it inside pipeline creation - in a JIT worker thread, before any
+    // draw runs, which is why removing the binding produced a flaky crash rather than a clean
+    // refusal. Declining is done by refusing the draw (VkProgramObject::declinedDescriptors),
+    // not by shrinking the layout.
+    //
+    // Two separate things have to hold, and neither is checkable from the SPIR-V alone:
+    //
+    //  * the count has to fit a VkDescriptorSetLayoutBinding this device will accept, and fit
+    //    the Uint16 it is stored in (65536 would narrow to 0) and the scratch the bind path
+    //    reserves from it;
+    //  * the frontend reflection has to have RESERVED that many consecutive uniform locations
+    //    for this uniform, because the per-element resolve paths address element k as
+    //    baseLocation + k. SPIRV-Reflect's `count` is the FLATTENED element count, while GL
+    //    locations follow the OUTER dimension only (ProgramObject::GetUniformArraySizeByTIndex
+    //    answers TType::getOuterArraySize()). For a one-dimensional array the two agree; for
+    //    `uniform sampler2D g[2][3]` SPIR-V says 6 where the reflection reserved 2, and
+    //    elements 2..5 would silently resolve onto whichever uniform got the next locations.
+    //
+    // Asking the reflection whether baseLocation and baseLocation + count - 1 are slots of the
+    // SAME uniform tests exactly that precondition, without this code having to model how
+    // glslang chooses to lay an array of arrays out.
+    //
+    // That is NOT on its own enough to start supporting the shape, though, and this check must
+    // not be relaxed alone: the binding-qualifier unit seeding in ProgramLinkTask looks an
+    // opaque uniform up by its name minus a trailing "[0]", so `goku[0][0]` misses the `goku`
+    // key and every element of an array of arrays seeds texture unit 0. Resolving those elements
+    // would then paint silently-wrong pixels with no diagnostic at all - strictly worse than
+    // declining. The decline goes away together with the seeding fix, not before it.
+    static Uint32 DescriptorCountForOpaqueUniformArray(const MG_State::GLState::ProgramObject& program,
+                                                       const String& uniformName, Uint32 binding, Int baseLocation,
+                                                       Uint32 reflectedCount, Uint32 maxBindings,
+                                                       const char* kindLabel, Bool& outDeclined) {
+        const Uint32 count = std::max<Uint32>(1u, reflectedCount);
+        if (count == 1) {
+            return 1u;
+        }
+        if (count > maxBindings) {
+            // Nothing legal to declare: the count would not fit a VkDescriptorSetLayoutBinding
+            // this device accepts, and it would narrow badly into the Uint16 that carries it
+            // (65536 becomes 0). Unlike the extent case below, this one CANNOT keep the layout
+            // consistent with the shader, so refusing the draw does not fully protect it - the
+            // driver still JITs a shader indexing past the declared count. Declaring as many as
+            // the device allows keeps vkCreateDescriptorSetLayout succeeding and the program
+            // inert; a device whose binding cap is smaller than a shader's array is not a
+            // configuration MobileGL can serve at all. Needs a >maxBindings-element array to
+            // reach (256 on desktop, ~16 on mobile).
+            MGLOG_I("ProgramFactory::ReflectLayout: %s array '%s' at binding %u has %u elements, past the %u "
+                    "this device can describe - declining the program",
+                    kindLabel, uniformName.c_str(), binding, count, maxBindings);
+            outDeclined = true;
+            return maxBindings;
+        }
+        if (baseLocation < 0 ||
+            !program.UniformLocationsAliasSameUniform(baseLocation, baseLocation + static_cast<Int>(count - 1u))) {
+            MGLOG_I("ProgramFactory::ReflectLayout: %s array '%s' at binding %u spans %u descriptors but the "
+                    "reflection reserved fewer uniform locations for it (base=%d) - a multi-dimensional array "
+                    "is the usual cause, and MobileGL declines it rather than resolve elements onto a "
+                    "neighbouring uniform",
+                    kindLabel, uniformName.c_str(), binding, count, baseLocation);
+            outDeclined = true;
+        }
+        return count;
+    }
+
     void ProgramFactory::ReflectLayout(const MG_State::GLState::ProgramObject& program,
                                        const Vector<Vector<Uint>>& spirv, VkProgramObject& entry) const {
         // Initialize layout vectors
@@ -2034,6 +2480,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         entry.dynamicBindings.clear();
         entry.bindingDescriptorCounts.assign(m_maxBindings, 1);
         entry.arrayedUniformBlockIndicesByBinding.clear();
+        entry.declinedDescriptors = false;
 
         // Use SpvcSession (Reflection mode) to reflect all SPIR-V modules in a single pass per module
         for (const auto& module : spirv) {
@@ -2224,11 +2671,58 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                     }
                     entry.storageBlockNameByBinding[binding] = uniformName;
                     entry.storageBlockIndexByBinding[binding] = static_cast<Int>(blockIndex);
+
+                    // A block INSTANCE array is ONE Vulkan binding carrying `count`
+                    // descriptors, while GL assigns its elements consecutive binding points
+                    // starting at the declared one (GL 4.6 core 7.8). Recording only element 0 -
+                    // which is all this used to do - left the layout claiming descriptorCount 1,
+                    // so every element past the first read a descriptor nobody wrote and
+                    // `b[1].data.length()` answered from an unconstrained buffer instead of its
+                    // own bound range (KHR-GL43.shader_storage_buffer_object.-
+                    // advanced-unsizedArrayLength-*).
+                    //
+                    // Bounds-checked like every other array kind. The EXTENT rule differs - a
+                    // block array's elements take consecutive GL binding points rather than
+                    // consecutive uniform locations, so DescriptorCountForOpaqueUniformArray's
+                    // location test does not apply here - but the size rule is identical: this
+                    // count goes straight into a VkDescriptorSetLayoutBinding and is narrowed to
+                    // a Uint16 on the way, where 65536 would silently become 0.
+                    const Uint32 storageArrayCount = std::max<Uint32>(1u, sampler->count);
+                    if (storageArrayCount > m_maxBindings) {
+                        MGLOG_I("ProgramFactory::ReflectLayout: storage block array '%s' at binding %u has %u "
+                                "elements, past the %u this device can describe - declining the program",
+                                uniformName.c_str(), binding, storageArrayCount, m_maxBindings);
+                        entry.declinedDescriptors = true;
+                        entry.bindingDescriptorCounts[binding] = static_cast<Uint16>(m_maxBindings);
+                        continue;
+                    }
+                    entry.bindingDescriptorCounts[binding] = static_cast<Uint16>(storageArrayCount);
                     continue;
                 }
 
                 const Int location = program.GetUniformLocation(uniformName);
                 if (location < 0) {
+                    // A uniform with no location is ordinarily one GL never made active, and
+                    // dropping it is routine. An ARRAY reaching here is not routine: it is the
+                    // multi-dimensional case. `uniform sampler2D g[2][3]` arrives from
+                    // SPIRV-Reflect as one binding of 6 descriptors named "g", while the frontend
+                    // reflection keys an array of arrays by its full "[0]"-terminated spelling
+                    // ("g[0][0]"), so no base location resolves and the per-element paths have
+                    // nothing to count from. Declining is the honest answer - but it has to SAY
+                    // so at a level that survives a release build, because dropping the binding
+                    // leaves the shader reading a descriptor the layout never declared.
+                    if (sampler->count > 1) {
+                        MGLOG_I("ProgramFactory::ReflectLayout: declining '%s' at binding %u - a %u-element "
+                                "descriptor array with no frontend uniform location (a multi-dimensional array "
+                                "of samplers or images is the known cause)",
+                                uniformName.c_str(), binding, sampler->count);
+                        entry.declinedDescriptors = true;
+                        // Declared, not resolved - see DescriptorCountForOpaqueUniformArray for
+                        // why the layout keeps describing a binding the draw path will refuse.
+                        entry.bindingDescriptorCounts[binding] =
+                            static_cast<Uint16>(std::min<Uint32>(sampler->count, m_maxBindings));
+                        continue;
+                    }
                     entry.bindingKinds[binding] = DescriptorBindingKind::None;
                     continue;
                 }
@@ -2236,6 +2730,24 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 const GLenum uniformType = program.GetUniformType(static_cast<Uint>(location));
 
                 if (descriptorKind == DescriptorBindingKind::StorageImage) {
+                    // An ARRAY of image uniforms is ONE binding carrying `count` descriptors,
+                    // and the layout has to say so. Leaving it at the default 1 declared
+                    // `uniform image2D g_image[4]` as a single-descriptor binding while the
+                    // shader indexed descriptors 1..3 of it - an out-of-bounds descriptor
+                    // access that lavapipe SIGSEGVs inside the JIT-ed shader thread rather than
+                    // reporting (KHR-GL42.shader_image_load_store.advanced-sso-simple). Unlike
+                    // a storage BLOCK array, whose elements take consecutive GL binding points
+                    // from the declared one, each element of an image array carries its own
+                    // independently assigned image unit - see ResolveStorageImageDescriptor.
+                    // Bounds- and extent-checked like the UBO array path above; see
+                    // DescriptorCountForOpaqueUniformArray for what "declined" costs and why
+                    // the reflection's reserved extent - not SPIRV-Reflect's flattened count -
+                    // is what the per-element resolve can actually address.
+                    const Uint32 imageArrayCount =
+                        DescriptorCountForOpaqueUniformArray(program, uniformName, binding, location, sampler->count,
+                                                             m_maxBindings, "image", entry.declinedDescriptors);
+                    entry.bindingDescriptorCounts[binding] = static_cast<Uint16>(imageArrayCount);
+
                     const VkFormat reflectedFormat =
                         ConvertSpirvImageFormatToVkFormat(sampler->image.image_format);
                     VkFormat& existingFormat = entry.storageImageFormatByBinding[binding];
@@ -2266,6 +2778,20 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                                 "ProgramFactory::ReflectLayout: failed to resolve texture target for '%s'",
                                 uniformName.c_str());
                 if (descriptorKind == DescriptorBindingKind::CombinedImageSampler) {
+                    // An ARRAY of sampler uniforms is ONE binding carrying `count` descriptors,
+                    // exactly like the image array above, and for the same reason: GLSL 4.20
+                    // gives `layout(binding = 1) uniform sampler2D goku[4]` one declaration
+                    // spanning texture units 1..4, each element with its own glUniform1i-assigned
+                    // unit. Leaving descriptorCount at 1 declared a single-descriptor binding
+                    // while the shader indexed descriptors 1..3 of it, and the bind path wrote
+                    // only element 0 - so elements 1..N read a descriptor nobody had written
+                    // (KHR-GL42.shading_language_420pack.binding_sampler_array; lavapipe faults
+                    // inside the JIT-ed shader rather than reporting).
+                    const Uint32 samplerArrayCount =
+                        DescriptorCountForOpaqueUniformArray(program, uniformName, binding, location, sampler->count,
+                                                             m_maxBindings, "sampler", entry.declinedDescriptors);
+                    entry.bindingDescriptorCounts[binding] = static_cast<Uint16>(samplerArrayCount);
+
                     const SamplerNumericDomain numericDomain = UniformTypeToSamplerNumericDomain(uniformType);
                     MOBILEGL_ASSERT(numericDomain != SamplerNumericDomain::Unknown,
                                     "ProgramFactory::ReflectLayout: failed to resolve sampler numeric domain "
@@ -2360,14 +2886,36 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         }
     }
 
+    void ProgramFactory::SetDefaultFramebufferHeight(Uint32 height) {
+        if (m_defaultFramebufferHeight == height) {
+            return;
+        }
+        m_defaultFramebufferHeight = height;
+        // Both memos key on (program, flags) alone, so neither can tell the two heights apart:
+        // drop the lookup memo, and bump the structure epoch so every caller holding a
+        // VkProgramObject* re-runs GetOrCreateProgram and lands on the new hash. The cached
+        // entries themselves stay - they are keyed by a hash that now includes the old height,
+        // so they can only be reached again if that height comes back, and the frame-boundary
+        // sweep retires them otherwise.
+        m_lastLookup = {};
+        ++m_cacheStructureEpoch;
+    }
+
     const ProgramFactory::VkProgramObject& ProgramFactory::GetOrCreateProgram(
         const MG_State::GLState::ProgramObject& program, CompileOptionFlags flags) {
         // Hashing the full SPIR-V of every stage is far too expensive to repeat per draw;
         // reuse the program's memoized hash while its backend state version is unchanged.
+        // The memo keys on the flags word, which ComputeHash is no longer a pure function of:
+        // a FragCoordYFlip variant also depends on the baked default-framebuffer height, so
+        // that height rides in the free high half of the key. Flags occupy the low bits, and a
+        // height cannot exceed the 16 bits a swapchain extent fits in.
+        const Uint memoKey = (flags & CompileOptionBit::FragCoordYFlip)
+                                 ? (flags.GetRaw() | (m_defaultFramebufferHeight << 16))
+                                 : flags.GetRaw();
         HashType hash = 0;
-        if (!program.GetBackendHashMemo(flags.GetRaw(), hash)) {
+        if (!program.GetBackendHashMemo(memoKey, hash)) {
             hash = ComputeHash(program, flags);
-            program.SetBackendHashMemo(flags.GetRaw(), hash);
+            program.SetBackendHashMemo(memoKey, hash);
         }
         auto it = m_cache.find(hash);
         if (it != m_cache.end()) {
@@ -2420,6 +2968,14 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 }
             }
 
+            if ((flags & ProgramFactory::CompileOptionBit::FragCoordYFlip) && shaders[i] &&
+                shaders[i]->GetShaderStage() == ShaderStage::Fragment) {
+                Vector<Uint> fragCoordSpirv;
+                if (TransformSpirvForFragCoordYFlip(moduleSpirvs[i], fragCoordSpirv, m_defaultFramebufferHeight)) {
+                    moduleSpirvs[i] = Move(fragCoordSpirv);
+                }
+            }
+
             // Vulkan's SPIR-V environment has no rectangle image dimension, so a
             // GL_TEXTURE_RECTANGLE lookup has to become the 2D one the texture is really
             // stored as - which addresses [0,1] where the application addressed texels.
@@ -2525,6 +3081,12 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
 #if MOBILEGL_LOG_ACTIVE_LEVEL <= MOBILEGL_LOG_LEVEL_DEBUG
             ValidateTransformedSpirv(moduleSpv, shaders[i]->GetShaderStage(), program.GetExternalIndex());
+#else
+            // Final module the driver receives; also checked in the INFO-level CI/test
+            // lanes, where the DEBUG gate above is compiled out.
+            if (MG_Util::ShaderTranspiler::ShaderCompiler::SpirvValidationEnabled()) {
+                ValidateTransformedSpirv(moduleSpv, shaders[i]->GetShaderStage(), program.GetExternalIndex());
+            }
 #endif
 
             VkShaderModuleCreateInfo smci{VK_STRUCTURE_TYPE_SHADER_MODULE_CREATE_INFO};
@@ -2542,6 +3104,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
             entry.modules.push_back(module);
             entry.stages.push_back(stage);
+            entry.stageSpirvDigests.push_back(ShaderStageSpirvDigest{
+                static_cast<Uint32>(stage.stage), static_cast<Uint32>(moduleSpv.size()),
+                XXH64(moduleSpv.data(), moduleSpv.size() * sizeof(Uint), 0)});
         }
 
         // Reflect and create layout as part of the program object
@@ -2551,6 +3116,20 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         ReflectVertexInputs(shaders, moduleSpirvs, entry);
         ReflectFragmentOutputs(shaders, moduleSpirvs, entry);
         ReflectLayout(program, moduleSpirvs, entry);
+        // A failed remap means the modules kept glslang's per-stage auto-mapped binding numbers -
+        // no cross-stage unification, no set->0 normalisation - so the bindings this layout
+        // describes are not the bindings the shader reads. That has to stop the program from
+        // drawing, and until now nothing did: the MOBILEGL_ASSERT above compiles out of every
+        // build past DEBUG, and RemapDescriptorBindingsForVulkan's own refusal message said so at
+        // a level an INFO build also drops. Declining is the mechanism that already exists for
+        // "the layout and the shader disagree", so route it through that. Set AFTER ReflectLayout,
+        // which clears the flag.
+        if (!remapOk) {
+            MGLOG_I("ProgramFactory::GetOrCreateProgram: declining program %u - its descriptor bindings could not "
+                    "be remapped, so the layout does not describe what the shader reads",
+                    program.GetExternalIndex());
+            entry.declinedDescriptors = true;
+        }
 
         return entry;
     }

MobileGL/MG_Backend/DirectVulkan/Renderer/ProgramFactory.h

@@ -9,6 +9,7 @@
 #pragma once
 
 #include "../VkIncludes.h"
+#include "PipelineFactory.h"
 #include "MG_State/GLState/ProgramState/ProgramObject.h"
 #include "MG_State/GLState/ProgramState/ShaderObject.h"
 #include "MG_State/GLState/TextureState/TextureEnum.h"
@@ -52,6 +53,13 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             // recorded while GL transform feedback is active, so plain draws keep the
             // undecorated variant.
             XfbCapture = 1 << 6,
+            // Rewrites the fragment stage's gl_FragCoord reads to GL's bottom-left window
+            // origin. Vulkan's gl_FragCoord.y IS the framebuffer row being written, and the
+            // default framebuffer's image is stored in display (top-left) order, so a shader
+            // that reads gl_FragCoord there sees `height - y_GL`. Set together with
+            // PositionYFlip (the two are the same fact about the same draws) except under a
+            // quarter turn, which this renderer does not convert rectangles for either.
+            FragCoordYFlip = 1 << 7,
         };
         using CompileOptionFlags = Flags<CompileOptionBit>;
         using HashType = Uint64;
@@ -62,6 +70,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             HashType hash = 0;
             Vector<VkPipelineShaderStageCreateInfo> stages;
             Vector<VkShaderModule> modules;
+            // Parallel to stages; identifies the exact module bytes handed to the driver when a
+            // pipeline creation fails. Sixteen bytes per stage instead of keeping the SPIR-V.
+            Vector<ShaderStageSpirvDigest> stageSpirvDigests;
 
             // Layout data (previously in separate VkProgramLayout)
             VkDescriptorSetLayout descriptorSetLayout = VK_NULL_HANDLE;
@@ -75,8 +86,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             Vector<Uint32> activeBindings;
             Vector<Uint32> dynamicBindings;
             Vector<Int> uniformBlockIndexByBinding;
-            // Descriptor count per binding (1 except for UBO instance arrays, which occupy one
-            // binding with descriptorCount = N).
+            // Descriptor count per binding (1 except for a descriptor ARRAY - a UBO or storage
+            // block instance array, an image uniform array or a sampler uniform array - each of
+            // which occupies one binding with descriptorCount = N).
             Vector<Uint16> bindingDescriptorCounts;
             // Per-element GL uniform block indices for arrayed UBO bindings (count > 1);
             // element 0 of a non-arrayed binding stays in uniformBlockIndexByBinding.
@@ -92,6 +104,19 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             // Set once during ReflectLayout so the per-draw path can skip the whole
             // storage-image preparation for the overwhelming majority of programs.
             Bool hasStorageImages = false;
+            // Something about this program's descriptors could not be resolved - an opaque
+            // uniform array whose elements have no addressable uniform locations (the
+            // multi-dimensional case), or a binding remap that failed outright. The binding
+            // STAYS DECLARED in the descriptor set layout; declining is done here, by refusing
+            // every draw, and BindProgramUniformBuffers returns false so the draw setup skips
+            // the draw exactly as it does for any other bind failure.
+            //
+            // Keeping the layout intact is the load-bearing half. Shrinking it instead - which
+            // is what the first cut of this did - leaves the shader reading a descriptor the
+            // layout never declared, and lavapipe segfaults on that inside PIPELINE CREATION,
+            // in a JIT worker thread, before any draw runs where a refusal could help. The
+            // reason was logged once at MGLOG_I when the descriptor was declined.
+            Bool declinedDescriptors = false;
             Int globalUboBinding = -1;
             Uint32 activeVertexInputLocationMask = 0;
             Array<GLenum, kMaxVertexInputLocations> vertexInputTypes{};
@@ -118,6 +143,14 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 hash = other.hash;
                 stages = std::move(other.stages);
                 modules = std::move(other.modules);
+                // Must travel with `modules`: these digests name the SPIR-V those exact
+                // shader modules were built from, and the pipeline-failure diagnostics
+                // print the two together. Leaving it behind used to merely lose the
+                // digests on a rehash; now that the cache is a robin-hood table, insertion
+                // SWAPS two entries, and a field that no move touches stays behind in the
+                // slot - pairing one program's modules with another program's digests, so
+                // a pipeline failure would be reported against the wrong SPIR-V.
+                stageSpirvDigests = std::move(other.stageSpirvDigests);
                 descriptorSetLayout = other.descriptorSetLayout;
                 pipelineLayout = other.pipelineLayout;
                 bindingKinds = std::move(other.bindingKinds);
@@ -136,6 +169,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 storageBlockNameByBinding = std::move(other.storageBlockNameByBinding);
                 storageBlockIndexByBinding = std::move(other.storageBlockIndexByBinding);
                 hasStorageImages = other.hasStorageImages;
+                declinedDescriptors = other.declinedDescriptors;
                 globalUboBinding = other.globalUboBinding;
                 activeVertexInputLocationMask = other.activeVertexInputLocationMask;
                 vertexInputTypes = other.vertexInputTypes;
@@ -150,6 +184,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 other.descriptorSetLayout = VK_NULL_HANDLE;
                 other.pipelineLayout = VK_NULL_HANDLE;
                 other.hasStorageImages = false;
+                other.declinedDescriptors = false;
                 other.globalUboBinding = -1;
                 other.activeVertexInputLocationMask = 0;
                 other.activeFragmentOutputLocationMask = 0;
@@ -167,6 +202,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 hash = other.hash;
                 stages = std::move(other.stages);
                 modules = std::move(other.modules);
+                stageSpirvDigests = std::move(other.stageSpirvDigests); // travels with `modules` - see the move ctor
                 descriptorSetLayout = other.descriptorSetLayout;
                 pipelineLayout = other.pipelineLayout;
                 bindingKinds = std::move(other.bindingKinds);
@@ -185,6 +221,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 storageBlockNameByBinding = std::move(other.storageBlockNameByBinding);
                 storageBlockIndexByBinding = std::move(other.storageBlockIndexByBinding);
                 hasStorageImages = other.hasStorageImages;
+                declinedDescriptors = other.declinedDescriptors;
                 globalUboBinding = other.globalUboBinding;
                 activeVertexInputLocationMask = other.activeVertexInputLocationMask;
                 vertexInputTypes = other.vertexInputTypes;
@@ -199,6 +236,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 other.descriptorSetLayout = VK_NULL_HANDLE;
                 other.pipelineLayout = VK_NULL_HANDLE;
                 other.hasStorageImages = false;
+                other.declinedDescriptors = false;
                 other.globalUboBinding = -1;
                 other.activeVertexInputLocationMask = 0;
                 other.activeFragmentOutputLocationMask = 0;
@@ -233,6 +271,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 }
                 modules.clear();
                 stages.clear();
+                stageSpirvDigests.clear(); // the modules they describe are gone
             }
         };
 
@@ -263,6 +302,17 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         const VkProgramObject& GetOrCreateProgram(
             const MG_State::GLState::ProgramObject& program, CompileOptionFlags flags);
 
+        // The default framebuffer's current image height, baked as a literal into every
+        // FragCoordYFlip variant (there is no push-constant or specialization channel here, and
+        // adding one for a value that changes only on swapchain recreation would cost the draw
+        // path more than a recompile costs a resize). It is therefore part of those variants'
+        // identity: ComputeHash mixes it in when the bit is set, so a height change re-keys them
+        // and leaves every other program's hash untouched. Setting a NEW height also bumps the
+        // cache-structure epoch, because a caller holding a memoised VkProgramObject* would
+        // otherwise keep using a module compiled against the old height.
+        void SetDefaultFramebufferHeight(Uint32 height);
+        Uint32 GetDefaultFramebufferHeight() const { return m_defaultFramebufferHeight; }
+
         // Bumped whenever m_cache's STRUCTURE changes (any insert or erase): the cache is
         // an open-addressing map holding entries by value, so both moves existing entries.
         // A caller that memoised a VkProgramObject* may keep dereferencing it only while
@@ -320,6 +370,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // True only when the logical device enabled both
         // shaderStorageImageReadWithoutFormat and shaderStorageImageWriteWithoutFormat.
         Bool m_unformattedFloatStorageImagesEnabled = false;
+        // See SetDefaultFramebufferHeight. 0 means "not known yet"; the FragCoordYFlip bit is
+        // never set before the swapchain exists, so no variant can be compiled against it.
+        Uint32 m_defaultFramebufferHeight = 0;
         mutable ProgramLookupCache m_lastLookup;
         // Monotonic frame-boundary counter (bumped in OnFrameBoundary) for cache aging.
         Uint64 m_frameCounter = 0;

MobileGL/MG_Backend/DirectVulkan/Renderer/UniformManager.cpp

@@ -68,6 +68,29 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         }
     }
 
+    // Uniform location of ELEMENT `element` of the opaque-uniform array at `baseLocation`, or
+    // -1 when the reflection did not reserve that element. DoReflection hands out one location
+    // per array element, so the element's location is the base plus its index - bounded by the
+    // array's real extent so a descriptorCount that outran the reflection cannot walk onto the
+    // next uniform. Element 0 is the ordinary non-array case and costs nothing extra.
+    static Int ResolveDescriptorElementLocation(const MG_State::GLState::ProgramObject& program, Int baseLocation,
+                                                Uint32 element) {
+        if (baseLocation < 0 || element == 0) {
+            return baseLocation;
+        }
+        const Int location = baseLocation + static_cast<Int>(element);
+        return program.UniformLocationsAliasSameUniform(baseLocation, location) ? location : -1;
+    }
+
+    // descriptorCount this binding declares in the descriptor set layout (1 for everything that
+    // is not an array). Kept in one place because the layout, the scratch reservation and the
+    // per-element write loops must all agree on it.
+    static Uint32 BindingDescriptorCount(const ProgramFactory::VkProgramObject& programObj, Uint32 binding) {
+        return binding < programObj.bindingDescriptorCounts.size()
+                   ? std::max<Uint32>(1u, programObj.bindingDescriptorCounts[binding])
+                   : 1u;
+    }
+
     static Int ResolveSamplerUnitIndex(const MG_State::GLState::ProgramObject& program, Int location, Uint32 binding) {
         MOBILEGL_ASSERT(location >= -1, "ResolveSamplerUnitIndex: invalid sampler location for binding %u", binding);
         if (location < 0) {
@@ -268,28 +291,44 @@ namespace MobileGL::MG_Backend::DirectVulkan {
     Bool UniformManager::ResolveSamplerDescriptor(VkCommandBuffer commandBuffer,
                                                             const MG_State::GLState::ProgramObject& program,
                                                             const ProgramFactory::VkProgramObject& programObj,
-                                                            Uint32 binding, VkDescriptorImageInfo& outImageInfo,
+                                                            Uint32 binding, Uint32 element,
+                                                            VkDescriptorImageInfo& outImageInfo,
                                                             Bool trustUnchangedHint) const {
         MOBILEGL_ASSERT(m_textureManager != nullptr, "ResolveSamplerDescriptor: texture manager is null");
         MOBILEGL_ASSERT(m_samplerManager != nullptr, "ResolveSamplerDescriptor: sampler manager is null");
+        // The whole-descriptor memo below is keyed by binding alone, so it describes a binding
+        // that carries exactly one descriptor. An arrayed binding's elements would overwrite
+        // each other in it (see SamplerResolveMemo::info); they re-resolve instead.
+        const Bool descriptorMemoUsable = BindingDescriptorCount(programObj, binding) == 1u;
         // The caller proved every input of this binding's resolution unchanged since the
         // last full resolve (which also filled the cache), so the whole chain below -
         // texture/sampler resolution, completeness probe, sync, layout handling, sampler
         // and view lookups - would recompute the identical descriptor.
-        if (trustUnchangedHint && binding < m_samplerResolveMemo.size() &&
+        if (trustUnchangedHint && descriptorMemoUsable && binding < m_samplerResolveMemo.size() &&
             m_samplerResolveMemo[binding].infoValid) {
             outImageInfo = m_samplerResolveMemo[binding].info;
             return true;
         }
         MOBILEGL_ASSERT(binding < programObj.samplerNameByBinding.size(),
                         "ResolveSamplerDescriptor: sampler binding %u name lookup out of range", binding);
+        // Per ELEMENT, and resolved BEFORE anything is looked up through it: GLSL 4.20 gives every
+        // element of `uniform sampler2D goku[4]` its own texture unit (consecutive from the
+        // declared binding, but glUniform1i may scatter them afterwards), so the unit - and with
+        // it the bound texture, the unit's sampler override and the fallback decision - is the
+        // element's, not the binding's. An element past the array's reserved extent has no unit
+        // at all, and must not fall back to resolving unit 0's texture.
+        const Int location =
+            ResolveDescriptorElementLocation(program, programObj.samplerUniformLocationByBinding[binding], element);
+        if (location < 0 && element > 0) {
+            MGLOG_D("ResolveSamplerDescriptor: binding %u element %u is past the end of its sampler array", binding,
+                    element);
+            return false;
+        }
+        const Int unit = ResolveSamplerUnitIndex(program, location, binding);
         // Raw-pointer resolve to skip the SharedPtr atomic refcount churn: the bound texture stays
         // alive through the draw via GL binding state. Only the fallback path needs a SharedPtr to
         // keep the fallback texture alive for the rest of this call.
-        MG_State::GLState::ITextureObject* texture = ResolveSamplerTextureRaw(program, programObj, binding);
-
-        const Int location = programObj.samplerUniformLocationByBinding[binding];
-        const Int unit = ResolveSamplerUnitIndex(program, location, binding);
+        MG_State::GLState::ITextureObject* texture = ResolveSamplerTextureRaw(program, programObj, binding, element);
         auto& textureUnit = MG_State::pGLContext->GetTextureUnitObject(unit);
         const auto& samplerOverride = textureUnit.GetSamplerObject();
         const auto preferredTarget = programObj.samplerTextureTargetByBinding[binding];
@@ -459,9 +498,21 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         if (outImageInfo.sampler == VK_NULL_HANDLE) {
             return false;
         }
+        // Only for a binding that carries a single descriptor - an array's elements would
+        // publish each other's descriptors here, and the next hinted draw would hand element
+        // N-1's texture to element 0.
         if (binding < m_samplerResolveMemo.size()) {
-            m_samplerResolveMemo[binding].info = outImageInfo;
-            m_samplerResolveMemo[binding].infoValid = true;
+            if (descriptorMemoUsable) {
+                m_samplerResolveMemo[binding].info = outImageInfo;
+                m_samplerResolveMemo[binding].infoValid = true;
+            } else {
+                // An arrayed binding publishes nothing here, and clears what a previous program
+                // published at this index. Not strictly required - the hint's proof obligations
+                // are program-scoped and the entry is reset every frame - but leaving another
+                // program's descriptor sitting in a slot this one never refreshes is the kind of
+                // thing the next reader has to re-derive is safe.
+                m_samplerResolveMemo[binding].infoValid = false;
+            }
             NoteSamplerResolveMemoTouched(binding);
         }
         return true;
@@ -500,42 +551,58 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
     Bool UniformManager::ProgramSamplesOnlySingleLevelTextures(
         const MG_State::GLState::ProgramObject& program, const ProgramFactory::VkProgramObject& programObj) {
+        // A declined program never draws (see VkProgramObject::declinedDescriptors), and its
+        // declined binding has no resolvable uniform location - so there is nothing to prove
+        // about the textures it would have sampled.
+        if (programObj.declinedDescriptors) {
+            return false;
+        }
         Bool sawSampler = false;
         for (Uint32 binding = 0; binding < programObj.bindingKinds.size(); ++binding) {
             if (programObj.bindingKinds[binding] != ProgramFactory::DescriptorBindingKind::CombinedImageSampler) {
                 continue;
             }
-            const auto* texture = ResolveSamplerTextureRaw(program, programObj, binding);
-            if (texture == nullptr) return false;
-            const auto& levelRange = texture->GetLevelRange();
-            if (levelRange.x() != levelRange.y()) return false;
+            // The rewrite this gates is program-wide, so EVERY sampler the program can read has
+            // to qualify - including every element of a sampler array, each of which reaches a
+            // different texture through its own unit.
+            const Uint32 descriptorCount = BindingDescriptorCount(programObj, binding);
+            for (Uint32 element = 0; element < descriptorCount; ++element) {
+                // The element's own location first, exactly as ResolveSamplerDescriptor resolves
+                // it - an element with no location would otherwise be judged on unit 0's texture.
+                const Int location = ResolveDescriptorElementLocation(
+                    program, programObj.samplerUniformLocationByBinding[binding], element);
+                if (location < 0 && element > 0) return false;
+                const auto* texture = ResolveSamplerTextureRaw(program, programObj, binding, element);
+                if (texture == nullptr) return false;
+                const auto& levelRange = texture->GetLevelRange();
+                if (levelRange.x() != levelRange.y()) return false;
 
-            // An explicit-LOD sample is a single filtered tap, so it also gives up anisotropic
-            // filtering - which a single-level view can still have. Resolve the sampler exactly
-            // the way ResolveSamplerDescriptor does and bail if anisotropy would apply.
-            const Int location = programObj.samplerUniformLocationByBinding[binding];
-            const Int unit = ResolveSamplerUnitIndex(program, location, binding);
-            const auto& samplerOverride = MG_State::pGLContext->GetTextureUnitObject(unit).GetSamplerObject();
-            const auto* effectiveSampler =
-                samplerOverride ? samplerOverride.get() : texture->GetSamplerObject().get();
-            if (effectiveSampler == nullptr) return false;
-            if (effectiveSampler->GetMaxAnisotropy() > 1.0f &&
-                effectiveSampler->GetMinFilter() == SamplerFilterMode::Linear &&
-                effectiveSampler->GetMagFilter() == SamplerFilterMode::Linear) {
-                return false;
-            }
+                // An explicit-LOD sample is a single filtered tap, so it also gives up anisotropic
+                // filtering - which a single-level view can still have. Resolve the sampler exactly
+                // the way ResolveSamplerDescriptor does and bail if anisotropy would apply.
+                const Int unit = ResolveSamplerUnitIndex(program, location, binding);
+                const auto& samplerOverride = MG_State::pGLContext->GetTextureUnitObject(unit).GetSamplerObject();
+                const auto* effectiveSampler =
+                    samplerOverride ? samplerOverride.get() : texture->GetSamplerObject().get();
+                if (effectiveSampler == nullptr) return false;
+                if (effectiveSampler->GetMaxAnisotropy() > 1.0f &&
+                    effectiveSampler->GetMinFilter() == SamplerFilterMode::Linear &&
+                    effectiveSampler->GetMagFilter() == SamplerFilterMode::Linear) {
+                    return false;
+                }
 
-            // An explicit LOD 0 makes lambda exactly 0, which is the magnification side of the
-            // min/mag decision. That only matches the implicit form when lambda could not have been
-            // positive anyway (the LOD clamp already pins it at or below 0), or when the two
-            // filters are the same and the choice cannot be observed.
-            const Float effectiveMaxLod = effectiveSampler->GetMipmapMode() == SamplerMipmapMode::None
-                                              ? 0.0f
-                                              : effectiveSampler->GetMaxLod();
-            if (effectiveMaxLod > 0.0f && effectiveSampler->GetMinFilter() != effectiveSampler->GetMagFilter()) {
-                return false;
+                // An explicit LOD 0 makes lambda exactly 0, which is the magnification side of the
+                // min/mag decision. That only matches the implicit form when lambda could not have been
+                // positive anyway (the LOD clamp already pins it at or below 0), or when the two
+                // filters are the same and the choice cannot be observed.
+                const Float effectiveMaxLod = effectiveSampler->GetMipmapMode() == SamplerMipmapMode::None
+                                                  ? 0.0f
+                                                  : effectiveSampler->GetMaxLod();
+                if (effectiveMaxLod > 0.0f && effectiveSampler->GetMinFilter() != effectiveSampler->GetMagFilter()) {
+                    return false;
+                }
+                sawSampler = true;
             }
-            sawSampler = true;
         }
         return sawSampler;
     }
@@ -568,14 +635,15 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
     MG_State::GLState::ITextureObject* UniformManager::ResolveSamplerTextureRaw(
         const MG_State::GLState::ProgramObject& program, const ProgramFactory::VkProgramObject& programObj,
-        Uint32 binding) {
+        Uint32 binding, Uint32 element) {
         MOBILEGL_ASSERT(MG_State::pGLContext != nullptr, "ResolveSamplerTextureRaw: GL context is null");
         MOBILEGL_ASSERT(binding < programObj.samplerUniformLocationByBinding.size(),
                         "ResolveSamplerTextureRaw: sampler location binding %u out of range", binding);
         MOBILEGL_ASSERT(binding < programObj.samplerTextureTargetByBinding.size(),
                         "ResolveSamplerTextureRaw: sampler target binding %u out of range", binding);
 
-        const Int location = programObj.samplerUniformLocationByBinding[binding];
+        const Int location =
+            ResolveDescriptorElementLocation(program, programObj.samplerUniformLocationByBinding[binding], element);
         const Int unit = ResolveSamplerUnitIndex(program, location, binding);
 
         auto& textureUnit = MG_State::pGLContext->GetTextureUnitObject(unit);
@@ -677,7 +745,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
     Bool UniformManager::ResolveStorageBufferDescriptor(const MG_State::GLState::ProgramObject& program,
                                                         const ProgramFactory::VkProgramObject& programObj,
-                                                        Uint32 binding,
+                                                        Uint32 binding, Uint32 element,
                                                         VkDescriptorBufferInfo& outBufferInfo) const {
         outBufferInfo = {};
         MOBILEGL_ASSERT(m_bufferManager != nullptr, "ResolveStorageBufferDescriptor: buffer manager is null");
@@ -688,8 +756,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         const Int blockIndex = programObj.storageBlockIndexByBinding[binding];
         MOBILEGL_ASSERT(blockIndex >= 0, "ResolveStorageBufferDescriptor: no SSBO block mapped to binding %u",
                         binding);
+        // A block instance array declares one block whose elements take consecutive GL binding
+        // points from the declared one (GL 4.6 core 7.8), and the reflection collapses the whole
+        // array to that one block - so the element index IS the offset from its binding.
         const GLuint frontendBinding =
-            GetShaderStorageBlockBinding(program, static_cast<GLuint>(blockIndex));
+            GetShaderStorageBlockBinding(program, static_cast<GLuint>(blockIndex)) + element;
         const Uint32 bindingPointCount =
             static_cast<Uint32>(MG_State::pGLContext->GetBufferBindingPointCount(BufferTarget::ShaderStorage));
         MOBILEGL_ASSERT(frontendBinding < bindingPointCount,
@@ -742,7 +813,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
     Bool UniformManager::ResolveStorageImageDescriptor(VkCommandBuffer commandBuffer,
                                                        const MG_State::GLState::ProgramObject& program,
                                                        const ProgramFactory::VkProgramObject& programObj,
-                                                       Uint32 binding,
+                                                       Uint32 binding, Uint32 element,
                                                        VkDescriptorImageInfo& outImageInfo) const {
         outImageInfo = {};
         MOBILEGL_ASSERT(m_textureManager != nullptr, "ResolveStorageImageDescriptor: texture manager is null");
@@ -750,11 +821,24 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         MOBILEGL_ASSERT(binding < programObj.samplerUniformLocationByBinding.size(),
                         "ResolveStorageImageDescriptor: binding %u out of range", binding);
 
-        const Int location = programObj.samplerUniformLocationByBinding[binding];
-        if (location < 0) {
+        const Int baseLocation = programObj.samplerUniformLocationByBinding[binding];
+        if (baseLocation < 0) {
             MGLOG_E("ResolveStorageImageDescriptor: storage image binding %u has no uniform location", binding);
             return false;
         }
+        // Per ELEMENT, and this is where an image array differs from a storage-block array: GL
+        // gives every element of `uniform image2D g_image[4]` its own glUniform1i-assigned image
+        // unit, and the four units need not be consecutive or even ordered (the conformance case
+        // uses 0, 2, 4, 6). DoReflection reserves one uniform location per array element, so the
+        // element's location is the base plus its index - checked against the array's real
+        // extent so a descriptorCount that outran the reflection cannot walk onto the next
+        // uniform.
+        const Int location = baseLocation + static_cast<Int>(element);
+        if (!program.UniformLocationsAliasSameUniform(baseLocation, location)) {
+            MGLOG_E("ResolveStorageImageDescriptor: binding %u element %u is past the end of its image array",
+                    binding, element);
+            return false;
+        }
         const Int imageUnit = program.GetUniformSamplerOrImageUnitIndex(static_cast<Uint>(location));
         if (imageUnit < 0 || imageUnit >= MG_State::GLState::TextureState::MAX_TEXTURE_IMAGE_UNITS) {
             MGLOG_E("ResolveStorageImageDescriptor: image unit %d out of range for binding %u",
@@ -844,7 +928,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
     Bool UniformManager::ResolveSampledBinding(const MG_State::GLState::ProgramObject& program,
                                                const ProgramFactory::VkProgramObject& programObj,
-                                               Uint32 binding,
+                                               Uint32 binding, Uint32 element,
                                                MG_State::GLState::ITextureObject*& outTexture,
                                                const MG_State::GLState::SamplerObject*& outSampler) const {
         // Open-coded ResolveSamplerTextureRaw so the unit is resolved once for both the
@@ -855,7 +939,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                         "ResolveSampledBinding: sampler location binding %u out of range", binding);
         MOBILEGL_ASSERT(binding < programObj.samplerTextureTargetByBinding.size(),
                         "ResolveSampledBinding: sampler target binding %u out of range", binding);
-        const Int location = programObj.samplerUniformLocationByBinding[binding];
+        const Int location =
+            ResolveDescriptorElementLocation(program, programObj.samplerUniformLocationByBinding[binding], element);
+        if (location < 0 && element > 0) {
+            return false;
+        }
         const Int unit = ResolveSamplerUnitIndex(program, location, binding);
         auto& textureUnit = MG_State::pGLContext->GetTextureUnitObject(unit);
         const TextureTarget preferredTarget = programObj.samplerTextureTargetByBinding[binding];
@@ -891,6 +979,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         if (outBindingRecords != nullptr) {
             outBindingRecords->clear();
         }
+        // Nothing to prepare for a program the bind path is going to refuse; its declined
+        // binding has no uniform location to resolve a texture through either.
+        if (programObj.declinedDescriptors) {
+            return true;
+        }
 
         const Uint32 bindingCount =
             std::min<Uint32>(m_maxBindings, static_cast<Uint32>(programObj.bindingKinds.size()));
@@ -899,19 +992,27 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 continue;
             }
 
-            MG_State::GLState::ITextureObject* texture = nullptr;
-            const MG_State::GLState::SamplerObject* sampler = nullptr;
-            if (!ResolveSampledBinding(program, programObj, binding, texture, sampler)) {
-                continue;
-            }
-            if (outBindingRecords != nullptr) {
-                outBindingRecords->push_back({texture != nullptr ? texture->GetLifetimeId() : 0,
-                                              sampler != nullptr ? sampler->GetLifetimeId() : 0});
-            }
+            // Every ELEMENT of a sampler array reaches its own texture through its own unit,
+            // so every element has to be in the sampled set: this walk is what gets those
+            // textures synced and transitioned to a sampled layout BEFORE the render pass
+            // opens, and a missed element would first be touched by the descriptor resolve
+            // inside an active pass.
+            const Uint32 descriptorCount = BindingDescriptorCount(programObj, binding);
+            for (Uint32 element = 0; element < descriptorCount; ++element) {
+                MG_State::GLState::ITextureObject* texture = nullptr;
+                const MG_State::GLState::SamplerObject* sampler = nullptr;
+                if (!ResolveSampledBinding(program, programObj, binding, element, texture, sampler)) {
+                    continue;
+                }
+                if (outBindingRecords != nullptr) {
+                    outBindingRecords->push_back({texture != nullptr ? texture->GetLifetimeId() : 0,
+                                                  sampler != nullptr ? sampler->GetLifetimeId() : 0});
+                }
 
-            auto found = std::find(outTextures.begin(), outTextures.end(), texture);
-            if (found == outTextures.end()) {
-                outTextures.push_back(texture);
+                auto found = std::find(outTextures.begin(), outTextures.end(), texture);
+                if (found == outTextures.end()) {
+                    outTextures.push_back(texture);
+                }
             }
         }
         return true;
@@ -920,6 +1021,10 @@ namespace MobileGL::MG_Backend::DirectVulkan {
     Bool UniformManager::SampledBindingsUnchanged(const MG_State::GLState::ProgramObject& program,
                                                   const ProgramFactory::VkProgramObject& programObj,
                                                   const Vector<SampledBindingRecord>& previousRecords) const {
+        // A declined program takes the full path every time and is refused there.
+        if (programObj.declinedDescriptors) {
+            return false;
+        }
         SizeT recordIndex = 0;
         // Iterate only the bindings this program declares (ascending), exactly like
         // BindProgramUniformBuffers: this runs per draw whenever the texture bind
@@ -932,18 +1037,24 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             if (programObj.bindingKinds[binding] != ProgramFactory::DescriptorBindingKind::CombinedImageSampler) {
                 continue;
             }
-            MG_State::GLState::ITextureObject* texture = nullptr;
-            const MG_State::GLState::SamplerObject* sampler = nullptr;
-            if (!ResolveSampledBinding(program, programObj, binding, texture, sampler)) {
-                continue;
-            }
-            if (recordIndex >= previousRecords.size()) {
-                return false;
-            }
-            const SampledBindingRecord& record = previousRecords[recordIndex++];
-            if (record.textureLifetimeId != (texture != nullptr ? texture->GetLifetimeId() : 0) ||
-                record.samplerLifetimeId != (sampler != nullptr ? sampler->GetLifetimeId() : 0)) {
-                return false;
+            // Element-for-element, in the same order CollectSampledTextures recorded them -
+            // the two walks have to visit the identical descriptor sequence or the positional
+            // comparison below drifts.
+            const Uint32 descriptorCount = BindingDescriptorCount(programObj, binding);
+            for (Uint32 element = 0; element < descriptorCount; ++element) {
+                MG_State::GLState::ITextureObject* texture = nullptr;
+                const MG_State::GLState::SamplerObject* sampler = nullptr;
+                if (!ResolveSampledBinding(program, programObj, binding, element, texture, sampler)) {
+                    continue;
+                }
+                if (recordIndex >= previousRecords.size()) {
+                    return false;
+                }
+                const SampledBindingRecord& record = previousRecords[recordIndex++];
+                if (record.textureLifetimeId != (texture != nullptr ? texture->GetLifetimeId() : 0) ||
+                    record.samplerLifetimeId != (sampler != nullptr ? sampler->GetLifetimeId() : 0)) {
+                    return false;
+                }
             }
         }
         return recordIndex == previousRecords.size();
@@ -956,6 +1067,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         outTextures.clear();
         MOBILEGL_ASSERT(MG_State::pGLContext != nullptr,
                         "CollectStorageImageTextures: GL context is null");
+        // Same as the sampled walk: a declined program is refused at bind time, and its declined
+        // binding has no uniform location to reach an image unit through.
+        if (programObj.declinedDescriptors) {
+            return true;
+        }
 
         const Uint32 bindingCount =
             std::min<Uint32>(m_maxBindings, static_cast<Uint32>(programObj.bindingKinds.size()));
@@ -968,26 +1084,40 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 return false;
             }
 
-            const Int location = programObj.samplerUniformLocationByBinding[binding];
-            if (location < 0) {
+            const Int baseLocation = programObj.samplerUniformLocationByBinding[binding];
+            if (baseLocation < 0) {
                 MGLOG_E("CollectStorageImageTextures: binding %u has no image uniform location", binding);
                 return false;
             }
-            const Int imageUnit = program.GetUniformSamplerOrImageUnitIndex(static_cast<Uint>(location));
-            if (imageUnit < 0 || imageUnit >= MG_State::GLState::TextureState::MAX_TEXTURE_IMAGE_UNITS) {
-                MGLOG_E("CollectStorageImageTextures: image unit %d is invalid for binding %u",
-                        imageUnit, binding);
-                return false;
-            }
+            // Per ELEMENT, for the same reason the sampled walk above is: an image ARRAY is one
+            // binding whose elements each carry their own image unit, so each reaches its own
+            // texture. This walk is what puts those textures into the pre-pass sync and layout
+            // transition; collecting only element 0 left elements 1..N to be first touched by
+            // the descriptor resolve, which happens with a render pass already open.
+            const Uint32 descriptorCount = BindingDescriptorCount(programObj, binding);
+            for (Uint32 element = 0; element < descriptorCount; ++element) {
+                const Int location = ResolveDescriptorElementLocation(program, baseLocation, element);
+                if (location < 0) {
+                    MGLOG_E("CollectStorageImageTextures: binding %u element %u is past the end of its image array",
+                            binding, element);
+                    return false;
+                }
+                const Int imageUnit = program.GetUniformSamplerOrImageUnitIndex(static_cast<Uint>(location));
+                if (imageUnit < 0 || imageUnit >= MG_State::GLState::TextureState::MAX_TEXTURE_IMAGE_UNITS) {
+                    MGLOG_E("CollectStorageImageTextures: image unit %d is invalid for binding %u element %u",
+                            imageUnit, binding, element);
+                    return false;
+                }
 
-            auto* texture = MG_State::pGLContext->GetImageTextureBinding(imageUnit).Texture.get();
-            if (texture == nullptr) {
-                MGLOG_E("CollectStorageImageTextures: image unit %d is unbound for binding %u",
-                        imageUnit, binding);
-                return false;
-            }
-            if (std::find(outTextures.begin(), outTextures.end(), texture) == outTextures.end()) {
-                outTextures.push_back(texture);
+                auto* texture = MG_State::pGLContext->GetImageTextureBinding(imageUnit).Texture.get();
+                if (texture == nullptr) {
+                    MGLOG_E("CollectStorageImageTextures: image unit %d is unbound for binding %u element %u",
+                            imageUnit, binding, element);
+                    return false;
+                }
+                if (std::find(outTextures.begin(), outTextures.end(), texture) == outTextures.end()) {
+                    outTextures.push_back(texture);
+                }
             }
         }
         return true;
@@ -1358,6 +1488,17 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                                                              VkPipelineBindPoint bindPoint,
                                                              const SamplerBindingOverride* samplerBindingOverride,
                                                              Bool samplerDescriptorsUnchangedHint) {
+        // This program has a descriptor MobileGL could not resolve (see
+        // VkProgramObject::declinedDescriptors). Refusing here is the whole of the decline: the
+        // binding is still declared in the layout, so the pipeline is consistent with the shader
+        // and creating it is safe - what must not happen is the draw, because the descriptor
+        // behind that binding can never be written. The draw setup skips the draw on a false
+        // return. ReflectLayout already said why, once, at MGLOG_I.
+        if (programObj.declinedDescriptors) {
+            MGLOG_D("UniformDescriptorBinder::BindProgramUniformBuffers: refusing a program whose descriptor layout "
+                    "was declined at reflection");
+            return false;
+        }
         auto& frame = m_frames[frameIndex];
         if (frame.descriptorPools.empty()) {
             MGLOG_E("UniformDescriptorBinder::BindProgramUniformBuffers failed: frame descriptor pools are invalid");
@@ -1416,13 +1557,27 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         dynamicOffsets.clear();
         // Arrayed UBO bindings contribute extra buffer infos and dynamic offsets; reserve for
         // the worst case so the pBufferInfo pointers taken below never dangle on reallocation.
+        // Arrayed SSBO bindings contribute extra buffer infos too (but no dynamic offsets).
         Uint32 uboArrayExtra = 0;
         for (const auto& arrayEntry : programObj.arrayedUniformBlockIndicesByBinding) {
             uboArrayExtra += static_cast<Uint32>(arrayEntry.second.size()) - 1u;
         }
+        // Surplus descriptors over "one per binding", summed across EVERY arrayed binding
+        // whatever its kind - storage blocks, image arrays and sampler arrays all land here.
+        // One number for all of them because each container below is bounded by the same total.
+        Uint32 arrayDescriptorExtra = 0;
+        for (const Uint16 count : programObj.bindingDescriptorCounts) {
+            if (count > 1) arrayDescriptorExtra += static_cast<Uint32>(count) - 1u;
+        }
         writes.reserve(m_maxBindings);
-        bufferInfos.reserve(m_maxBindings + uboArrayExtra);
-        imageInfos.reserve(m_maxBindings);
+        bufferInfos.reserve(m_maxBindings + uboArrayExtra + arrayDescriptorExtra);
+        // Every binding pushes at most descriptorCount image infos, so bindings + surplus is the
+        // worst case. Reserving only m_maxBindings here was exact while every binding pushed
+        // exactly one - and reallocates under an image or sampler array, dangling every
+        // pImageInfo already recorded in `writes` before vkUpdateDescriptorSets reads them. That
+        // is reachable wherever m_maxBindings is small (it clamps to ~16 on Adreno and Mali),
+        // which is exactly where a 7-element CTS sampler array does not fit the slack.
+        imageInfos.reserve(m_maxBindings + arrayDescriptorExtra);
         texelBufferViews.reserve(m_maxBindings);
         dynamicOffsets.reserve(programObj.dynamicBindings.size() + uboArrayExtra);
 
@@ -1450,10 +1605,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             write.descriptorCount = 1;
 
             if (kind == ProgramFactory::DescriptorBindingKind::UniformBufferDynamic) {
-                const Uint32 descriptorCount =
-                    binding < programObj.bindingDescriptorCounts.size()
-                        ? std::max<Uint32>(1, programObj.bindingDescriptorCounts[binding])
-                        : 1u;
+                const Uint32 descriptorCount = BindingDescriptorCount(programObj, binding);
                 dynamicUboDescriptorCount += descriptorCount;
                 fastRebindUboBinding = binding;
                 const SizeT firstBufferInfoIndex = bufferInfos.size();
@@ -1493,59 +1645,109 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 write.pTexelBufferView = &texelBufferViews.back();
                 writes.push_back(write);
             } else if (kind == ProgramFactory::DescriptorBindingKind::StorageBuffer) {
-                VkDescriptorBufferInfo bufferInfo{};
-                if (!ResolveStorageBufferDescriptor(program, programObj, binding, bufferInfo)) {
-                    MGLOG_E(
-                        "UniformDescriptorBinder::BindProgramUniformBuffers failed: storage buffer binding %u has no valid descriptor",
-                        binding);
-                    return false;
+                // One write per binding, but `descriptorCount` buffer infos: a GLSL block
+                // instance array occupies a single binding whose elements each come from their
+                // own GL binding point.
+                const Uint32 descriptorCount = BindingDescriptorCount(programObj, binding);
+                const SizeT firstBufferInfoIndex = bufferInfos.size();
+                for (Uint32 element = 0; element < descriptorCount; ++element) {
+                    VkDescriptorBufferInfo bufferInfo{};
+                    if (!ResolveStorageBufferDescriptor(program, programObj, binding, element, bufferInfo)) {
+                        MGLOG_E(
+                            "UniformDescriptorBinder::BindProgramUniformBuffers failed: storage buffer binding %u "
+                            "element %u has no valid descriptor",
+                            binding, element);
+                        return false;
+                    }
+                    bufferInfos.push_back(bufferInfo);
                 }
 
-                bufferInfos.push_back(bufferInfo);
                 fastRebindKindsEligible = false;
                 write.descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_BUFFER;
-                write.pBufferInfo = &bufferInfos.back();
+                write.descriptorCount = descriptorCount;
+                write.pBufferInfo = &bufferInfos[firstBufferInfoIndex];
                 writes.push_back(write);
             } else if (kind == ProgramFactory::DescriptorBindingKind::StorageImage) {
-                VkDescriptorImageInfo imageInfo{};
-                if (!ResolveStorageImageDescriptor(commandBuffer, program, programObj, binding, imageInfo)) {
-                    MGLOG_E(
-                        "UniformDescriptorBinder::BindProgramUniformBuffers failed: storage image binding %u has no valid descriptor",
-                        binding);
-                    return false;
+                // One write per binding, but `descriptorCount` image infos: an ARRAY of image
+                // uniforms is a single binding whose elements each carry their own image unit.
+                // Writing only element 0 - which is all this used to do - left elements 1..N
+                // never written at all, and a shader that indexes them reads an undefined
+                // descriptor (lavapipe faults inside the shader; a real driver is free to do
+                // anything).
+                const Uint32 descriptorCount = BindingDescriptorCount(programObj, binding);
+                const SizeT firstImageInfoIndex = imageInfos.size();
+                for (Uint32 element = 0; element < descriptorCount; ++element) {
+                    VkDescriptorImageInfo imageInfo{};
+                    if (!ResolveStorageImageDescriptor(commandBuffer, program, programObj, binding, element,
+                                                       imageInfo)) {
+                        MGLOG_E(
+                            "UniformDescriptorBinder::BindProgramUniformBuffers failed: storage image binding %u "
+                            "element %u has no valid descriptor",
+                            binding, element);
+                        return false;
+                    }
+                    imageInfos.push_back(imageInfo);
                 }
-                imageInfos.push_back(imageInfo);
                 fastRebindKindsEligible = false;
                 write.descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_IMAGE;
-                write.pImageInfo = &imageInfos.back();
+                write.descriptorCount = descriptorCount;
+                write.pImageInfo = &imageInfos[firstImageInfoIndex];
                 writes.push_back(write);
             } else {
-                VkDescriptorImageInfo imageInfo{};
-                Bool hasImage = false;
-                if (samplerBindingOverride != nullptr &&
-                    samplerBindingOverride->binding == binding &&
-                    samplerBindingOverride->texture != nullptr &&
-                    samplerBindingOverride->sampler != nullptr) {
-                    hasImage = ResolveSamplerDescriptorOverride(*samplerBindingOverride, imageInfo);
-                } else {
-                    hasImage = ResolveSamplerDescriptor(commandBuffer, program, programObj, binding, imageInfo,
-                                                        samplerDescriptorsUnchangedHint);
+                // One write per binding, but `descriptorCount` image infos: a sampler ARRAY is a
+                // single binding whose elements each carry their own texture unit. Writing only
+                // element 0 - which is all this used to do - left elements 1..N never written,
+                // so a shader indexing them sampled a descriptor nobody had filled in
+                // (KHR-GL42.shading_language_420pack.binding_sampler_array).
+                const Uint32 descriptorCount = BindingDescriptorCount(programObj, binding);
+                // Overrides come only from MobileGL's own blit and depth-mipmap programs, whose
+                // samplers are scalars; the override replaces THE descriptor at its binding, so
+                // there is no element for it to mean on an arrayed one.
+                const Bool overrideThisBinding = samplerBindingOverride != nullptr &&
+                                                 samplerBindingOverride->binding == binding &&
+                                                 samplerBindingOverride->texture != nullptr &&
+                                                 samplerBindingOverride->sampler != nullptr;
+                MOBILEGL_ASSERT(
+                    !overrideThisBinding || descriptorCount == 1,
+                    "BindProgramUniformBuffers: sampler override targets arrayed binding %u (%u descriptors)",
+                    binding, descriptorCount);
+                const SizeT firstImageInfoIndex = imageInfos.size();
+                for (Uint32 element = 0; element < descriptorCount; ++element) {
+                    VkDescriptorImageInfo imageInfo{};
+                    Bool hasImage = false;
+                    if (overrideThisBinding && element == 0) {
+                        hasImage = ResolveSamplerDescriptorOverride(*samplerBindingOverride, imageInfo);
+                    } else {
+                        hasImage = ResolveSamplerDescriptor(commandBuffer, program, programObj, binding, element,
+                                                            imageInfo, samplerDescriptorsUnchangedHint);
+                    }
+                    if (!hasImage) {
+                        MGLOG_E(
+                            "UniformDescriptorBinder::BindProgramUniformBuffers failed: sampler binding %u element %u "
+                            "has no valid texture descriptor",
+                            binding, element);
+                        return false;
+                    }
+                    if (imageInfo.sampler == VK_NULL_HANDLE || imageInfo.imageView == VK_NULL_HANDLE) {
+                        MGLOG_E(
+                            "UniformDescriptorBinder::BindProgramUniformBuffers failed: sampler binding %u element %u "
+                            "has null sampler or imageView",
+                            binding, element);
+                        return false;
+                    }
+                    imageInfos.push_back(imageInfo);
                 }
-                if (!hasImage) {
-                    MGLOG_E(
-                        "UniformDescriptorBinder::BindProgramUniformBuffers failed: sampler binding %u has no valid texture descriptor",
-                        binding);
-                    return false;
+                if (descriptorCount > 1) {
+                    // The dynamic-offset-only rebind replays a whole descriptor set on the
+                    // strength of the sampler hint alone, and its eligibility probe was written
+                    // for bindings that carry one descriptor each. An arrayed sampler binding
+                    // also bypasses the per-binding descriptor memo, so there is nothing for it
+                    // to win here either.
+                    fastRebindKindsEligible = false;
                 }
-                if (imageInfo.sampler == VK_NULL_HANDLE || imageInfo.imageView == VK_NULL_HANDLE) {
-                    MGLOG_E(
-                        "UniformDescriptorBinder::BindProgramUniformBuffers failed: sampler binding %u has null sampler or imageView",
-                        binding);
-                    return false;
-                }
-                imageInfos.push_back(imageInfo);
                 write.descriptorType = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER;
-                write.pImageInfo = &imageInfos.back();
+                write.descriptorCount = descriptorCount;
+                write.pImageInfo = &imageInfos[firstImageInfoIndex];
                 writes.push_back(write);
             }
         }

MobileGL/MG_Backend/DirectVulkan/Renderer/UniformManager.h

@@ -53,10 +53,13 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // caches - a live layout's entry must never be purged (its sets would be
         // unreachable pool slots), so there is deliberately no age-based sweep here.
         void OnDescriptorSetLayoutDestroyed(VkDescriptorSetLayout descriptorSetLayout);
-        // One record per visited CombinedImageSampler binding (post fallback substitution,
-        // in binding order): the resolved texture and effective sampler, as never-reused
-        // lifetime ids so a freed-and-reallocated object at the same heap address can only
-        // MISS a comparison, never false-hit it (same ABA rule as SamplerResolveMemo).
+        // One record per visited CombinedImageSampler DESCRIPTOR (post fallback substitution,
+        // in binding order, and within a binding in array-element order): the resolved texture
+        // and effective sampler, as never-reused lifetime ids so a freed-and-reallocated object
+        // at the same heap address can only MISS a comparison, never false-hit it (same ABA
+        // rule as SamplerResolveMemo). An arrayed binding contributes one record per element -
+        // element granularity is required, or swapping the textures of two elements of the same
+        // array would leave the record list identical and the fast path would keep a stale set.
         struct SampledBindingRecord {
             Uint64 textureLifetimeId = 0;
             Uint64 samplerLifetimeId = 0;
@@ -143,8 +146,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // texture after the fallback substitution (may still be null when no fallback
         // exists), effective sampler = unit override else the texture's own sampler.
         // False = the binding is skipped (unbound with a non-2D fallback target).
+        // `element` indexes a sampler array inside the binding; see ResolveSamplerDescriptor.
         Bool ResolveSampledBinding(const MG_State::GLState::ProgramObject& program,
-                                   const ProgramFactory::VkProgramObject& programObj, Uint32 binding,
+                                   const ProgramFactory::VkProgramObject& programObj, Uint32 binding, Uint32 element,
                                    MG_State::GLState::ITextureObject*& outTexture,
                                    const MG_State::GLState::SamplerObject*& outSampler) const;
         // Raw-pointer variant for the per-draw sampled-texture walk (CollectSampledTextures):
@@ -152,27 +156,37 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // only need the pointer skip the SharedPtr copy's atomic refcount churn.
         static MG_State::GLState::ITextureObject* ResolveSamplerTextureRaw(
             const MG_State::GLState::ProgramObject& program,
-            const ProgramFactory::VkProgramObject& programObj, Uint32 binding);
+            const ProgramFactory::VkProgramObject& programObj, Uint32 binding, Uint32 element);
         SharedPtr<MG_State::GLState::ITextureObject> GetFallbackTexture(TextureTarget target) const;
+        // `element` indexes a sampler ARRAY inside one binding; each element carries its own
+        // independently assigned GL texture unit, so it selects the texture, the sampler
+        // override and the fallback separately from its neighbours.
+        //
         // trustUnchangedHint: reuse this binding's cached VkDescriptorImageInfo outright
         // (see BindProgramUniformBuffers' samplerDescriptorsUnchangedHint for the proof
-        // obligations the caller carries).
+        // obligations the caller carries). The cache is keyed by binding alone, so it is
+        // used ONLY for single-descriptor bindings - see m_samplerResolveMemo.
         Bool ResolveSamplerDescriptor(VkCommandBuffer commandBuffer, const MG_State::GLState::ProgramObject& program,
                                       const ProgramFactory::VkProgramObject& programObj, Uint32 binding,
-                                      VkDescriptorImageInfo& outImageInfo,
+                                      Uint32 element, VkDescriptorImageInfo& outImageInfo,
                                       Bool trustUnchangedHint = false) const;
         Bool ResolveSamplerDescriptorOverride(const SamplerBindingOverride& samplerBindingOverride,
                                               VkDescriptorImageInfo& outImageInfo) const;
         Bool ResolveTexelBufferDescriptor(const MG_State::GLState::ProgramObject& program,
                                           const ProgramFactory::VkProgramObject& programObj, Uint32 binding,
                                           Uint32 frameIndex, VkBufferView& outBufferView);
+        // `element` indexes a block INSTANCE array's descriptors; it is 0 for every ordinary
+        // block. Each element resolves through its own GL storage block, and so its own GL
+        // binding point, buffer and glBindBufferRange window.
         Bool ResolveStorageBufferDescriptor(const MG_State::GLState::ProgramObject& program,
                                             const ProgramFactory::VkProgramObject& programObj, Uint32 binding,
-                                            VkDescriptorBufferInfo& outBufferInfo) const;
+                                            Uint32 element, VkDescriptorBufferInfo& outBufferInfo) const;
+        // `element` indexes an image ARRAY inside one binding; each element carries its own
+        // independently assigned GL image unit.
         Bool ResolveStorageImageDescriptor(VkCommandBuffer commandBuffer,
                                            const MG_State::GLState::ProgramObject& program,
                                            const ProgramFactory::VkProgramObject& programObj, Uint32 binding,
-                                           VkDescriptorImageInfo& outImageInfo) const;
+                                           Uint32 element, VkDescriptorImageInfo& outImageInfo) const;
         // Result of resolving a UBO binding: either a zero-copy direct bind to the app's resident
         // VkBuffer (the GLES backend's approach - no per-draw copy) or the CPU payload to upload.
         struct UboBindResult {
@@ -341,6 +355,14 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             // proves every resolve input unchanged; cleared with the per-frame reset
             // (the cached VkSampler outlives a frame only via a fresh resolve, which
             // also re-stamps it against VkSamplerManager's frame-boundary sweep).
+            //
+            // This one field is keyed by binding but describes ONE descriptor, so it is
+            // written and read only for single-descriptor bindings. A sampler ARRAY's
+            // elements share the binding and would overwrite each other here - the last
+            // element resolved would then be handed to element 0 on the next hinted draw.
+            // Every other field above is self-validating (each compares its full key
+            // before reuse, and the view-format entry is a pure function of format and
+            // numeric domain), so an arrayed binding may keep using those.
             VkDescriptorImageInfo info{};
             Bool infoValid = false;
         };

MobileGL/MG_Backend/DirectVulkan/Renderer/VertexInputStateFactory.h

@@ -111,10 +111,12 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
         const VulkanRendererConfig& m_config;
         VkPhysicalDevice m_physicalDevice = VK_NULL_HANDLE;
-        // Values are heap-allocated: FastSTL::unordered_map is open-addressing,
-        // so INSERT invalidates references to stored values. The draw path (and
-        // the VAOs' state-pointer memos) hold entry pointers across inserts;
-        // only the unique_ptr cell moves, never the pointee.
+        // Values are heap-allocated: UnorderedMap is open-addressing, so INSERT
+        // invalidates references to stored values - and so does ERASE, which shifts
+        // the rest of the probe cluster into the hole and therefore moves entries
+        // other than the erased one. The draw path (and the VAOs' state-pointer
+        // memos) hold entry pointers across both; only the unique_ptr cell moves,
+        // never the pointee.
         UnorderedMap<HashType, UniquePtr<BackendVertexInputState>> m_cache;
         // Monotonic frame-boundary counter (bumped in OnFrameBoundary) for cache aging.
         Uint64 m_frameBoundaryCounter = 0;

MobileGL/MG_Backend/DirectVulkan/Renderer/VkBufferManager.cpp

@@ -161,12 +161,23 @@ namespace MobileGL::MG_Backend::DirectVulkan {
     }
 
     void VkBufferManager::CollectAllDeferredReleases() {
+        // Per-resource releases only. Every one of them was deferred behind a BumpSliceEpoch,
+        // so no memo can still name the handle, and the caller has proved the GPU is idle.
+        //
+        // The transient arena's releases are deliberately NOT collected here. A buffer lands
+        // there when the arena outgrows it mid-frame (BufferArena::EnsureCapacity), and at
+        // that moment every slice already handed out from this frame's arena still names it -
+        // VkBufferResource::transientSlice above all, which AcquireStreamedSlice keeps
+        // serving for the whole frame serial on the strength of transientFrameSerial alone.
+        // Nothing bumps the slice epoch for those other resources, so freeing the buffer
+        // here left the streamed memo handing a destroyed VkBuffer to vkCmdBindIndexBuffer
+        // (llvmpipe then faulted inside the draw; the Create/Flywheel indirect retrace died
+        // exactly this way). Mid-frame drains do not advance m_frameSerial, so they must not
+        // free arena storage either: the arena's own ResetFrame/BeginFrame is the point where
+        // the slot's slices stop being reachable, and that is where these releases land.
         for (Uint32 frameIndex = 0; frameIndex < m_deferredBufferReleases.size(); ++frameIndex) {
             CollectDeferredReleases(frameIndex);
         }
-        for (Uint32 frameIndex = 0; frameIndex < m_transientUploadArena.GetFrameCount(); ++frameIndex) {
-            m_transientUploadArena.CollectDeferredReleases(frameIndex);
-        }
     }
 
     void VkBufferManager::NotifyDeviceIdle() {

MobileGL/MG_Backend/DirectVulkan/Renderer/VkBufferManager.h

@@ -102,10 +102,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // Recreate all per-frame transient arenas
         Bool RecreateTransientArenas(Uint32 frameCount);
         void BeginFrame(Uint32 frameIndex);
-        // Drains every frame slot's deferred buffer/resource releases (and the
-        // transient arena's parked superseded blocks). Only valid when the
-        // caller has proven every queue submission complete; used by the
-        // present-less frame-boundary drain.
+        // Drains every frame slot's deferred buffer/resource releases. Only valid when
+        // the caller has proven every queue submission complete; used by the present-less
+        // frame-boundary drain. Deliberately does NOT touch the transient arena's parked
+        // superseded blocks: those are still named by this frame's slices (see the
+        // definition), and only a frame rewind retires them.
         void CollectAllDeferredReleases();
         // All previously submitted GPU work has completed (vkDeviceWaitIdle).
         void NotifyDeviceIdle();

MobileGL/MG_Backend/DirectVulkan/Renderer/VkRenderPassManager.h

@@ -101,6 +101,42 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             std::swap(layers, that.layers);
             std::swap(lastUsedFrame, that.lastUsedFrame);
         }
+        // Move ASSIGNMENT, not just construction. The move constructor above and the
+        // destructor below each independently suppress the implicit one, which left the
+        // type move-constructible but not move-assignable - and therefore not swappable,
+        // which std::swap(pair&, pair&) requires. That was invisible while UnorderedMap
+        // only ever move-CONSTRUCTED an element into a fresh slot. ska::flat_hash_map
+        // probes robin-hood: inserting swaps the entry being placed against the one
+        // already sitting in the slot whenever it has travelled further from its desired
+        // position, so the mapped type has to be swappable or the table fails to
+        // instantiate at all.
+        //
+        // SWAP SEMANTICS, exactly like the move constructor: this does not release the
+        // destination's handles, it parks them in `that`, which destroys them when it
+        // dies. That is correct for the only caller - std::swap, whose temporary expires
+        // immediately - and it is what keeps the three-move sequence from destroying a
+        // live render pass. It is NOT correct for a hand-written `a = std::move(b)` where
+        // `a` held live handles and `b` outlives the statement: those handles would then
+        // survive until `b` dies. There is no such caller; add a destroy-then-steal
+        // assignment before writing one.
+        RenderPassEntry& operator=(RenderPassEntry&& that) noexcept {
+            if (this != &that) {
+                std::swap(hash, that.hash);
+                std::swap(renderPass, that.renderPass);
+                std::swap(framebuffer, that.framebuffer);
+                std::swap(compatibilityHash, that.compatibilityHash);
+                std::swap(pendingClearAttachments, that.pendingClearAttachments);
+                std::swap(trackedAttachmentLayouts, that.trackedAttachmentLayouts);
+                std::swap(attachmentCount, that.attachmentCount);
+                std::swap(colorAttachmentCount, that.colorAttachmentCount);
+                std::swap(hasDepthStencilAttachment, that.hasDepthStencilAttachment);
+                std::swap(sampleCount, that.sampleCount);
+                std::swap(extent, that.extent);
+                std::swap(layers, that.layers);
+                std::swap(lastUsedFrame, that.lastUsedFrame);
+            }
+            return *this;
+        }
         RenderPassEntry(
             Uint64 hash,
             VkRenderPass renderpass,
@@ -315,26 +351,30 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             Uint64 deferredAtFrame = 0;
         };
 
-        // Node-based std::unordered_map, deliberately not FastSTL's open-addressing UnorderedMap:
+        // Node-based std::unordered_map, deliberately NOT the open-addressing UnorderedMap:
         // callers cache a RenderbufferResource* - or a bare &resource->layout - and then make further
         // calls that touch this map. BlitFramebuffer is the one that bit: it resolves the source and
         // destination colour bindings (ResolveColorBlitBinding caches &rbResource->layout), then
-        // materializes the source's pending clear, which looks that same resource up again. FastSTL's
-        // operator[] runs its load-factor check before find_key and reallocates the whole bucket array
-        // when occupancy crosses it, so even a plain lookup relocates every element; erase only
-        // tombstones and never decrements the occupancy, so the doubling keeps firing. After a
-        // relocation the cached pointer names freed storage still holding the pre-clear
-        // VK_IMAGE_LAYOUT_UNDEFINED, and BlitFramebuffer bails out at "source image layout is
-        // undefined", silently dropping the blit - renderbuffers_storage_multisample read back zero
-        // instead of the clear colour on exactly the iterations that grew the table.
+        // materializes the source's pending clear, which looks that same resource up again. Growing
+        // an open-addressed table relocates every element, so the cached pointer went on to name
+        // freed storage still holding the pre-clear VK_IMAGE_LAYOUT_UNDEFINED; BlitFramebuffer bailed
+        // out at "source image layout is undefined", silently dropping the blit -
+        // renderbuffers_storage_multisample read back zero instead of the clear colour on exactly the
+        // iterations that grew the table.
         //
         // Reordering the materialize ahead of the resolves - the fix ReadPixels got - does not cover
         // this: the destination resolve still runs after the source pointer is taken. The depth blit,
         // GetOrCreateRenderPass's depthRenderbufferResource and ReadDepthStencilPixels cache the same
         // kind of pointer, so the invariant belongs in the container rather than in a per-call-site
-        // ordering rule. m_textureResources is node-based for the same reason. This buys stability
-        // across rehash and insert only - erase still invalidates the erased element, which is safe
-        // here because a renderbuffer that is an FBO attachment is held alive by that attachment.
+        // ordering rule. m_textureResources is node-based for the same reason.
+        //
+        // The case for keeping this node-based got STRONGER with ska::flat_hash_map, so do not read
+        // the paragraph above as merely historical: ska erases by shifting the rest of the probe
+        // cluster backwards into the hole, so erasing one renderbuffer relocates OTHER renderbuffers'
+        // entries - a cached pointer can now be invalidated by a key it has nothing to do with, which
+        // no call-site ordering rule can defend against. (What did change: ska's operator[] returns on
+        // a hit before it runs its grow check, so a plain lookup of a PRESENT key no longer relocates.
+        // That narrows the insert hazard; it does not touch the erase one.)
         std::unordered_map<MG_State::GLState::RenderbufferObject*, RenderbufferResource> m_renderbufferResources;
         UnorderedMap<MG_State::GLState::RenderbufferObject*, PendingRenderbufferClear> m_pendingRenderbufferClears;
         Vector<DeferredRenderbufferRelease> m_deferredRenderbufferReleases;

MobileGL/MG_Backend/DirectVulkan/Renderer/VulkanRenderer.h

@@ -211,10 +211,15 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                                     GLsizei height, GLenum format, GLenum type, void* pixels);
         // Copy-and-repack core shared by depth-stencil ReadPixels and GetTexImage;
         // expects command recording to be active and any render pass already ended.
+        //
+        // `defaultFramebufferOrientation` is set only when the source is the swapchain's
+        // depth/stencil image, which this renderer stores display-side-up: the copy rect then
+        // has to be mapped out of GL's bottom-origin space and the copied rows re-oriented on
+        // the way back, exactly as the colour ReadPixels path does.
         void ReadDepthStencilImageToClient(VkImage image, VkFormat vkFormat, VkImageLayout* trackedLayout,
                                            VkImageAspectFlags imageAspect, Uint32 mipLevel, Uint32 baseArrayLayer,
                                            GLint x, GLint y, GLsizei width, GLsizei height, GLenum format, GLenum type,
-                                           void* pixels);
+                                           void* pixels, Bool defaultFramebufferOrientation = false);
         // Same-extent depth blit between images of different depth formats: host
         // round-trip with a per-texel re-encode (see BlitNamedFramebuffer).
         Bool BlitDepthAcrossFormats(FrameContext::FrameData& frame, VkImage srcImage, VkFormat srcFormat,
@@ -363,6 +368,31 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             Uint32 samplerBinding = 0;
         };
 
+        // A single-sample staging image for multisample-resolve blits that also have to change
+        // orientation. vkCmdResolveImage cannot flip (it takes one offset per side, not the
+        // invertible pair vkCmdBlitImage takes), so a resolve into or out of the default
+        // framebuffer used to land the mirrored band. Resolving here first and then blitting from
+        // here separates the two operations, and each one then does only what it can express.
+        //
+        // Pooled rather than created per blit: the CTS runs hundreds of these back to back, and
+        // create-destroy per call would both cost allocations and, worse, need per-call deferred
+        // destruction to outlive the recording. It grows to the largest extent asked for and is
+        // reused; format changes recreate it.
+        struct MultisampleResolveScratchImage {
+            VkImage image = VK_NULL_HANDLE;
+            VmaAllocation allocation = VK_NULL_HANDLE;
+            VkFormat format = VK_FORMAT_UNDEFINED;
+            VkExtent2D extent = {0, 0};
+            VkImageLayout layout = VK_IMAGE_LAYOUT_UNDEFINED;
+        };
+        MultisampleResolveScratchImage m_msResolveScratch;
+        // Returns a scratch image at least `extent` in size with exactly `format`, transitioned to
+        // TRANSFER_DST and ready to be resolved into. Null image on failure (the caller then falls
+        // back to the direct resolve).
+        Bool AcquireMultisampleResolveScratchImage(VkCommandBuffer commandBuffer, VkFormat format,
+                                                   VkExtent2D extent);
+        void DestroyMultisampleResolveScratchImage();
+
         struct DeferredDepthMipmapCleanup {
             Vector<VkImageView> imageViews;
             Vector<VkFramebuffer> framebuffers;
@@ -445,15 +475,13 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         void* m_platformDisplay = nullptr;
         void* m_platformLibrary = nullptr;
         void* m_platformCloseDisplay = nullptr;
-        // Some real ICDs (e.g. NVIDIA's proprietary Linux driver) don't implement
-        // VK_EXT_headless_surface at all. Detected once in CreateInstance() from the
-        // enumerated instance extensions; when false, CreateSurface() falls back to a
-        // hidden Xlib window instead of vkCreateHeadlessSurfaceEXT.
+        // Whether the loader exposes VK_EXT_headless_surface, detected once in
+        // CreateInstance() from the enumerated instance extensions. On desktop an
+        // offscreen surface REQUIRES it: false is a clean, loud bring-up failure, never
+        // a substituted window. (Android is the one exception and has its own path -
+        // no Mali/Adreno driver seen so far exposes the extension, so a windowless
+        // context is given an AImageReader ANativeWindow that is never displayed.)
         Bool m_headlessSurfaceSupported = true;
-        // Set when CreateSurface() had to create its own Xlib window for the fallback
-        // above (rather than being handed one by the caller), so Shutdown() knows it
-        // owns that window and must destroy it.
-        Bool m_ownsFallbackXlibWindow = false;
         // Android has the same shortfall: no Mali/Adreno driver seen so far exposes
         // VK_EXT_headless_surface, so a windowless (EGL pbuffer) context gets an
         // AImageReader's ANativeWindow to hand the WSI instead. Nothing is ever
@@ -1120,6 +1148,18 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         Bool MaterializePendingClearForRenderbuffer(
             VkCommandBuffer commandBuffer,
             const SharedPtr<MG_State::GLState::RenderbufferObject>& renderbuffer);
+        // The default framebuffer's twin of the two above. It cannot go through
+        // MaterializePendingClearForTexture: the default FBO's colour attachment is a
+        // placeholder texture object, and syncing THAT would clear a texture image nobody
+        // presents instead of the acquired swapchain image.
+        Bool MaterializePendingClearForDefaultFramebuffer(VkCommandBuffer commandBuffer,
+                                                          MG_State::GLState::FramebufferObject& fbo,
+                                                          FramebufferAttachmentType attachmentType);
+        // Its depth/stencil half: a different image (the swapchain's depth/stencil twin), a
+        // different clear command and per-aspect masking.
+        Bool MaterializePendingDepthStencilClearForDefaultFramebuffer(
+            VkCommandBuffer commandBuffer, const MG_State::GLState::FramebufferAttachmentObject& attachment,
+            const ClearAttachmentPayload& payload);
         VkPipeline GetOrCreateBlitPipeline(const RenderPassEntry& renderPassEntry);
         Bool GenerateDepthMipmapWithShader(FrameContext::FrameData& frame,
                                            MG_State::GLState::ITextureObject& texture,

MobileGL/MG_Benchmark/CMakeLists.txt

@@ -43,3 +43,4 @@ set_tests_properties(SanityBench PROPERTIES LABELS benchmark)
 add_subdirectory(Program)
 add_subdirectory(Buffer)
 add_subdirectory(Driver)
+add_subdirectory(Container)

MobileGL/MG_Benchmark/Container/CMakeLists.txt

@@ -0,0 +1,20 @@
+cmake_minimum_required(VERSION 3.24)
+
+add_executable(
+        UnorderedMapBench
+        UnorderedMapBench.cpp
+)
+
+target_include_directories(UnorderedMapBench PRIVATE
+        ${MGL_ROOT}/include
+        ${MGL_ROOT}/MobileGL
+)
+
+target_link_libraries(
+        UnorderedMapBench PRIVATE
+        benchmark::benchmark
+        ${LINK_LIBRARIES}
+)
+
+add_test(NAME UnorderedMapBench COMMAND UnorderedMapBench --benchmark_counters_tabular=true)
+set_tests_properties(UnorderedMapBench PROPERTIES LABELS benchmark)

MobileGL/MG_Benchmark/Container/UnorderedMapBench.cpp

@@ -0,0 +1,248 @@
+// MobileGL - MobileGL/MG_Benchmark/Container/UnorderedMapBench.cpp
+// Copyright (c) 2025-2026 MobileGL-Dev
+// Licensed under the GNU Lesser General Public License v3.0:
+//   https://www.gnu.org/licenses/gpl-3.0.txt
+//   https://www.gnu.org/licenses/lgpl-3.0.txt
+// SPDX-License-Identifier: LGPL-3.0-only
+// End of Source File Header
+//
+// The standing performance observatory for MobileGL::UnorderedMap.
+//
+// This benchmarks the ALIAS, never a concrete table, so whatever UnorderedMap
+// names today is what gets measured - swap the container in MG_Util/Types.h and
+// re-run this same binary to get a directly comparable set of numbers. That is
+// the point of it: the container sits on per-draw paths, so a change to it needs
+// evidence, and the evidence should be produced the same way every time.
+//
+// The workloads are the shapes the tree actually exercises, not generic hash-map
+// microbenchmarks. Four key shapes, because they stress a hash function very
+// differently:
+//   * SEQUENTIAL   dense small integers - GL object names from the index generator
+//                  (buffer/texture/framebuffer/sampler registries).
+//   * POINTER      real heap addresses - StateBackendObjectRegistry keys on
+//                  StateObject*. These are aligned, so their low bits are the
+//                  least random part of the key; a table that indexes on raw low
+//                  bits clusters badly here and one that mixes first does not.
+//                  Taken from the real allocator rather than a synthetic stride,
+//                  which would flatter whichever table mixes its bits.
+//   * DIGEST       already well-mixed 64-bit values - the XXH64 pipeline,
+//                  vertex-input-state and program memos.
+//   * NAME         short strings - uniform/attribute name to location maps.
+//
+// Sizes sweep from 8 upward because the per-draw memos are usually SMALL; a table
+// that only wins at 4096 entries has not won anything that matters here.
+//
+// Run:  build-linux/MobileGL/MG_Benchmark/Container/UnorderedMapBench
+//   or: ctest -R UnorderedMapBench     (label: benchmark)
+
+#include <cstdint>
+#include <memory>
+#include <random>
+#include <string>
+#include <vector>
+#include <benchmark/benchmark.h>
+
+#include "MG_Util/Types.h"
+
+using namespace MobileGL;
+
+namespace {
+
+    constexpr Int64 kMinSize = 8;
+    constexpr Int64 kMaxSize = 4096;
+
+    // Keep the real allocations alive for the whole process: the POINTER shape is
+    // only honest if the keys are addresses the allocator actually handed out, and
+    // they have to stay unique (a freed address can be handed out twice).
+    std::vector<std::unique_ptr<char[]>>& PointerKeyStorage() {
+        static std::vector<std::unique_ptr<char[]>> storage;
+        return storage;
+    }
+
+    Vector<Uint64> SequentialKeys(SizeT n) {
+        Vector<Uint64> keys;
+        keys.reserve(n);
+        for (SizeT i = 0; i < n; ++i) keys.push_back(static_cast<Uint64>(i) + 1);
+        return keys;
+    }
+
+    Vector<Uint64> PointerKeys(SizeT n) {
+        auto& storage = PointerKeyStorage();
+        Vector<Uint64> keys;
+        keys.reserve(n);
+        std::mt19937_64 rng(0xBEEF);
+        std::vector<std::unique_ptr<char[]>> churn;
+        for (SizeT i = 0; i < n; ++i) {
+            // State objects are not all one size, and the allocator sees other
+            // traffic between them - a single uniform stride is not what this
+            // registry ever sees.
+            const SizeT sz = 96 + (rng() % 192);
+            auto p = std::make_unique<char[]>(sz);
+            keys.push_back(reinterpret_cast<Uint64>(p.get()));
+            storage.push_back(std::move(p));
+            if ((rng() & 3) == 0) churn.push_back(std::make_unique<char[]>(32 + (rng() % 128)));
+        }
+        return keys;
+    }
+
+    Vector<Uint64> DigestKeys(SizeT n) {
+        Vector<Uint64> keys;
+        keys.reserve(n);
+        std::mt19937_64 rng(0xC0FFEE);
+        for (SizeT i = 0; i < n; ++i) keys.push_back(rng());
+        return keys;
+    }
+
+    Vector<String> NameKeys(SizeT n) {
+        static const char* kPrefixes[] = {"u_", "a_", "mc_", "iris_", "gl_", "v_"};
+        Vector<String> keys;
+        keys.reserve(n);
+        for (SizeT i = 0; i < n; ++i) {
+            keys.push_back(String(kPrefixes[i % 6]) + "Uniform" + std::to_string(i) + "_xyz");
+        }
+        return keys;
+    }
+
+    // Key sets are built once per size and shared: generating them inside the timed
+    // loop would measure the generator (and, for POINTER, the allocator) instead of
+    // the table.
+    template <typename KeyVec, KeyVec (*Make)(SizeT)>
+    const KeyVec& CachedKeys(SizeT n) {
+        static UnorderedMap<SizeT, KeyVec> cache;
+        auto it = cache.find(n);
+        if (it != cache.end()) return it->second;
+        return cache.emplace(n, Make(n)).first->second;
+    }
+
+    template <typename Key>
+    UnorderedMap<Key, Uint64> Populated(const Vector<Key>& keys) {
+        UnorderedMap<Key, Uint64> map;
+        for (SizeT i = 0; i < keys.size(); ++i) map[keys[i]] = i;
+        return map;
+    }
+
+    // ---- the workloads ----------------------------------------------------
+
+    // The dominant per-draw operation by a wide margin: a populated cache that is
+    // read far more often than it is written.
+    template <typename KeyVec, KeyVec (*Make)(SizeT)>
+    void LookupHit(benchmark::State& state) {
+        const auto& keys = CachedKeys<KeyVec, Make>(static_cast<SizeT>(state.range(0)));
+        auto map = Populated(keys);
+        for (auto _ : state) {
+            for (const auto& k : keys) {
+                auto it = map.find(k);
+                benchmark::DoNotOptimize(it->second);
+            }
+        }
+        state.SetItemsProcessed(state.iterations() * static_cast<Int64>(keys.size()));
+    }
+
+    // "Is this resource cached yet?" answered NO - the probe length on a miss is a
+    // different cost from a hit, and resource caches ask this constantly.
+    template <typename KeyVec, KeyVec (*Make)(SizeT)>
+    void LookupMiss(benchmark::State& state) {
+        const SizeT n = static_cast<SizeT>(state.range(0));
+        const auto& keys = CachedKeys<KeyVec, Make>(n);
+        auto map = Populated(keys);
+        const KeyVec absent = Make(n); // same shape, never inserted
+        for (auto _ : state) {
+            for (const auto& k : absent) {
+                benchmark::DoNotOptimize(map.find(k) != map.end());
+            }
+        }
+        state.SetItemsProcessed(state.iterations() * static_cast<Int64>(absent.size()));
+    }
+
+    // Building a cache from empty, rehashes included.
+    template <typename KeyVec, KeyVec (*Make)(SizeT)>
+    void InsertGrow(benchmark::State& state) {
+        const auto& keys = CachedKeys<KeyVec, Make>(static_cast<SizeT>(state.range(0)));
+        for (auto _ : state) {
+            UnorderedMap<typename KeyVec::value_type, Uint64> map;
+            for (SizeT i = 0; i < keys.size(); ++i) map[keys[i]] = i;
+            benchmark::DoNotOptimize(map.size());
+        }
+        state.SetItemsProcessed(state.iterations() * static_cast<Int64>(keys.size()));
+    }
+
+    // Cache eviction and refill: erase half by key, put them back. This is the
+    // aged-out-entry sweep the pipeline and vertex-input caches do.
+    template <typename KeyVec, KeyVec (*Make)(SizeT)>
+    void EraseChurn(benchmark::State& state) {
+        const auto& keys = CachedKeys<KeyVec, Make>(static_cast<SizeT>(state.range(0)));
+        for (auto _ : state) {
+            state.PauseTiming();
+            auto map = Populated(keys);
+            state.ResumeTiming();
+            for (SizeT i = 0; i < keys.size(); i += 2) benchmark::DoNotOptimize(map.erase(keys[i]));
+            for (SizeT i = 0; i < keys.size(); i += 2) map[keys[i]] = i;
+            benchmark::DoNotOptimize(map.size());
+        }
+        state.SetItemsProcessed(state.iterations() * static_cast<Int64>(keys.size()));
+    }
+
+    // Mass eviction: erase-while-iterating across the whole table. This is the loop
+    // shape that a container's erase()-return contract can get wrong, and the one
+    // that fed garbage handles to vkDestroyPipeline when it was wrong before.
+    template <typename KeyVec, KeyVec (*Make)(SizeT)>
+    void EraseSweep(benchmark::State& state) {
+        const auto& keys = CachedKeys<KeyVec, Make>(static_cast<SizeT>(state.range(0)));
+        for (auto _ : state) {
+            state.PauseTiming();
+            auto map = Populated(keys);
+            state.ResumeTiming();
+            for (auto it = map.begin(); it != map.end();) it = map.erase(it);
+            benchmark::DoNotOptimize(map.size());
+        }
+        state.SetItemsProcessed(state.iterations() * static_cast<Int64>(keys.size()));
+    }
+
+    // Whole-table walks: the per-frame sweeps that age entries out, and the
+    // teardown loops that destroy every Vulkan object a cache owns.
+    template <typename KeyVec, KeyVec (*Make)(SizeT)>
+    void Iterate(benchmark::State& state) {
+        const auto& keys = CachedKeys<KeyVec, Make>(static_cast<SizeT>(state.range(0)));
+        auto map = Populated(keys);
+        for (auto _ : state) {
+            Uint64 acc = 0;
+            for (const auto& entry : map) acc += entry.second;
+            benchmark::DoNotOptimize(acc);
+        }
+        state.SetItemsProcessed(state.iterations() * static_cast<Int64>(keys.size()));
+    }
+
+} // namespace
+
+#define MGL_MAP_BENCH(WORKLOAD, SHAPE, VEC, MAKER)                                                 \
+    BENCHMARK_TEMPLATE(WORKLOAD, VEC, MAKER)                                                       \
+        ->Name(#WORKLOAD "/" #SHAPE)                                                               \
+        ->RangeMultiplier(8)                                                                       \
+        ->Range(kMinSize, kMaxSize)
+
+MGL_MAP_BENCH(LookupHit, sequential, Vector<Uint64>, SequentialKeys);
+MGL_MAP_BENCH(LookupHit, pointer, Vector<Uint64>, PointerKeys);
+MGL_MAP_BENCH(LookupHit, digest, Vector<Uint64>, DigestKeys);
+MGL_MAP_BENCH(LookupHit, name, Vector<String>, NameKeys);
+
+MGL_MAP_BENCH(LookupMiss, sequential, Vector<Uint64>, SequentialKeys);
+MGL_MAP_BENCH(LookupMiss, pointer, Vector<Uint64>, PointerKeys);
+MGL_MAP_BENCH(LookupMiss, digest, Vector<Uint64>, DigestKeys);
+MGL_MAP_BENCH(LookupMiss, name, Vector<String>, NameKeys);
+
+MGL_MAP_BENCH(InsertGrow, sequential, Vector<Uint64>, SequentialKeys);
+MGL_MAP_BENCH(InsertGrow, pointer, Vector<Uint64>, PointerKeys);
+MGL_MAP_BENCH(InsertGrow, digest, Vector<Uint64>, DigestKeys);
+MGL_MAP_BENCH(InsertGrow, name, Vector<String>, NameKeys);
+
+MGL_MAP_BENCH(EraseChurn, sequential, Vector<Uint64>, SequentialKeys);
+MGL_MAP_BENCH(EraseChurn, digest, Vector<Uint64>, DigestKeys);
+MGL_MAP_BENCH(EraseChurn, name, Vector<String>, NameKeys);
+
+MGL_MAP_BENCH(EraseSweep, sequential, Vector<Uint64>, SequentialKeys);
+MGL_MAP_BENCH(EraseSweep, digest, Vector<Uint64>, DigestKeys);
+
+MGL_MAP_BENCH(Iterate, sequential, Vector<Uint64>, SequentialKeys);
+MGL_MAP_BENCH(Iterate, digest, Vector<Uint64>, DigestKeys);
+
+BENCHMARK_MAIN();

MobileGL/MG_Impl/GLImpl/Buffer/GL_Buffer.cpp

@@ -1491,8 +1491,8 @@ namespace MobileGL::MG_Impl::GLImpl {
     // offset and size, which is also how glBindBuffersRange spells "reset this element"
     // (a NULL buffers array, or a zero entry inside one).
     static Bool ValidateBufferRangeOffsetAndSize(GLenum target, GLintptr offset, GLsizeiptr size,
-                                                 const char* funcName) {
-        if (size <= 0) {
+                                                 const char* funcName, Bool hasBuffer = true) {
+        if (hasBuffer && size <= 0) {
             MG_State::pGLContext->RecordError(
                 ErrorCode::InvalidValue,
                 MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", funcName,
@@ -1527,16 +1527,27 @@ namespace MobileGL::MG_Impl::GLImpl {
                 return false;
             }
         }
-        // A transform feedback capture binding is addressed in 32-bit components, so BOTH the
-        // offset and the size must be multiples of 4.
-        if (target == GL_TRANSFORM_FEEDBACK_BUFFER && ((offset % 4) != 0 || (size % 4) != 0)) {
+        // GL 4.6 core 6.1.1 constrains the OFFSET to a multiple of four for both
+        // TRANSFORM_FEEDBACK_BUFFER and ATOMIC_COUNTER_BUFFER (the atomic-counter one has no
+        // queryable alignment pname, which is why it was missing here), and the SIZE only for
+        // transform feedback, whose capture is written in whole 32-bit components. Extending the
+        // size rule to atomic counters as well breaks a legal bind: the conformance suite splits
+        // MAX_ATOMIC_COUNTER_BUFFER_SIZE evenly across the binding points and that quotient is
+        // not required to land on four.
+        if ((target == GL_TRANSFORM_FEEDBACK_BUFFER || target == GL_ATOMIC_COUNTER_BUFFER) && (offset % 4) != 0) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidValue,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", funcName,
+                                             std::format("offset ({}) must be a multiple of 4 for {}.", offset,
+                                                         MG_Util::ConvertGLEnumToString(target))));
+            return false;
+        }
+        if (target == GL_TRANSFORM_FEEDBACK_BUFFER && hasBuffer && (size % 4) != 0) {
             MG_State::pGLContext->RecordError(
                 ErrorCode::InvalidValue,
                 MakeUnique<GenericErrorInfo>(
                     "MG_Impl/GLImpl", funcName,
-                    std::format("offset ({}) and size ({}) must both be multiples of 4 for "
-                                "GL_TRANSFORM_FEEDBACK_BUFFER.",
-                                offset, size)));
+                    std::format("size ({}) must be a multiple of 4 for GL_TRANSFORM_FEEDBACK_BUFFER.", size)));
             return false;
         }
         return true;
@@ -1548,7 +1559,12 @@ namespace MobileGL::MG_Impl::GLImpl {
         BufferTarget bufferTarget = MG_Util::ConvertGLEnumToBufferTarget(target);
         if (!BufferImpl::ValidateBufferBindingPointTarget(bufferTarget)) return;
         if (!BufferImpl::ValidateBufferBindingPointIndex(bufferTarget, index)) return;
-        if (buffer != 0 && !ValidateBufferRangeOffsetAndSize(target, offset, size, __func__)) return;
+        // The target's alignment rules are a property of the BINDING POINT, not of the buffer,
+        // so they apply even when buffer is zero - which is exactly how
+        // KHR-GL43.shader_storage_buffer_object.negative-api-bind probes the SSBO alignment
+        // (glBindBufferRange(SHADER_STORAGE_BUFFER, 0, 0, alignment - 1, 0)). Only the size
+        // rules need a buffer, since buffer 0 detaches the binding point and ignores size.
+        if (!ValidateBufferRangeOffsetAndSize(target, offset, size, __func__, /*hasBuffer: */ buffer != 0)) return;
         if (bufferTarget == BufferTarget::TransformFeedback && MG_State::pGLContext->IsTransformFeedbackActive()) {
             MG_State::pGLContext->RecordError(
                 ErrorCode::InvalidOperation,
@@ -1732,10 +1748,30 @@ namespace MobileGL::MG_Impl::GLImpl {
         return BufferImpl::ValidateBufferBindingPointRange(bufferTarget, first, count, funcName);
     }
 
+    // ARB_multi_bind states the equivalence to a loop of single binds "except that ... buffers
+    // will not be created if they do not exist": glBindBuffer instantiates a name glGenBuffers
+    // merely reserved, glBindBuffers* must refuse it and raise INVALID_OPERATION instead
+    // (KHR-GL44.multi_bind.errors_bind_buffers).
+    //
+    // Deliberately PER ELEMENT, not all-or-nothing: the equivalence the extension defines is a
+    // loop, so a bad entry costs its own binding point and nothing else. Rejecting the whole
+    // call instead cost multi_bind.functional_bind_buffers_base its bindings.
+    static Bool IsExistingBufferForMultiBind(GLuint buffer, GLsizei index, const char* funcName) {
+        if (buffer == 0 || MG_State::pGLContext->ValidateBufferObject(buffer)) return true;
+        MG_State::pGLContext->RecordError(
+            ErrorCode::InvalidOperation,
+            MakeUnique<GenericErrorInfo>(
+                "MG_Impl/GLImpl", funcName,
+                std::format("buffers[{}] ({}) is not the name of an existing buffer object.", index, buffer)));
+        return false;
+    }
+
     void BindBuffersBase(GLenum target, GLuint first, GLsizei count, const GLuint* buffers) {
         if (!ValidateMultiBindBufferRange(target, first, count, __func__)) return;
         for (GLsizei i = 0; i < count; ++i) {
-            BindBufferBase_State(target, first + i, buffers ? buffers[i] : 0);
+            const GLuint buffer = buffers ? buffers[i] : 0;
+            if (!IsExistingBufferForMultiBind(buffer, i, __func__)) continue;
+            BindBufferBase_State(target, first + i, buffer);
         }
     }
 
@@ -1749,6 +1785,7 @@ namespace MobileGL::MG_Impl::GLImpl {
                           const GLsizeiptr* sizes) {
         if (!ValidateMultiBindBufferRange(target, first, count, __func__)) return;
         for (GLsizei i = 0; i < count; ++i) {
+            if (buffers && !IsExistingBufferForMultiBind(buffers[i], i, __func__)) continue;
             if (!buffers || buffers[i] == 0) {
                 BindBufferBase_State(target, first + i, 0);
             } else {

MobileGL/MG_Impl/GLImpl/Drawing/GL_Drawing.cpp

@@ -14,8 +14,8 @@
 #include "../Getter/GL_Getter.h"
 
 namespace MobileGL::MG_Impl::GLImpl {
-    static Bool ValidateCurrentProgramForExecution(const char* functionName) {
-        const auto& currentProgram = MG_State::pGLContext->GetProgramForDraw();
+    static Bool ValidateProgramForExecution(const SharedPtr<MG_State::GLState::ProgramObject>& currentProgram,
+                                            const char* functionName) {
         if (!currentProgram) {
             MG_State::pGLContext->RecordError(
                 ErrorCode::InvalidOperation,
@@ -34,10 +34,17 @@ namespace MobileGL::MG_Impl::GLImpl {
         return true;
     }
 
+    static Bool ValidateCurrentProgramForExecution(const char* functionName) {
+        return ValidateProgramForExecution(MG_State::pGLContext->GetProgramForDraw(), functionName);
+    }
+
+    // A dispatch resolves its program through the DISPATCH accessor: with a pipeline bound
+    // that is the pipeline's compute stage program, not the graphics composite a draw would
+    // build - which no longer contains a compute stage to find at all.
     static Bool ValidateCurrentProgramForCompute(const char* functionName) {
-        if (!ValidateCurrentProgramForExecution(functionName)) return false;
+        const auto& currentProgram = MG_State::pGLContext->GetProgramForDispatch();
+        if (!ValidateProgramForExecution(currentProgram, functionName)) return false;
 
-        const auto& currentProgram = MG_State::pGLContext->GetProgramForDraw();
         if (currentProgram->GetShaderIndexByStage(ShaderStage::Compute) < 0) {
             MG_State::pGLContext->RecordError(
                 ErrorCode::InvalidOperation,
@@ -493,15 +500,12 @@ namespace MobileGL::MG_Impl::GLImpl {
     }
 
     void DispatchComputeIndirect(GLintptr indirect) {
-        auto dispatchComputeIndirect = MG_Backend::gBackendFunctionsTable.GL.DispatchComputeIndirect;
-        if (!dispatchComputeIndirect) {
-            MG_State::pGLContext->RecordError(
-                ErrorCode::InvalidOperation,
-                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", __func__,
-                                             "Backend does not support indirect compute dispatch."));
-            return;
-        }
-        if (!ValidateCurrentProgramForCompute(__func__)) return;
+        // Argument and binding validation runs FIRST. Both are properties of the call and of GL
+        // state, so a context whose backend cannot dispatch at all must still report the
+        // argument error the spec names rather than masking every one of them with
+        // "unsupported" - which is what put GL_INVALID_OPERATION where
+        // KHR-GL43.compute_shader.api-indirect expects GL_INVALID_VALUE.
+        //
         // GL 4.6 core 19: `indirect` is a byte offset into GL_DISPATCH_INDIRECT_BUFFER -
         // negative or misaligned is INVALID_VALUE, nothing bound is INVALID_OPERATION.
         if (indirect < 0 || (indirect % 4) != 0) {
@@ -520,6 +524,29 @@ namespace MobileGL::MG_Impl::GLImpl {
                                              "No buffer is bound to GL_DISPATCH_INDIRECT_BUFFER."));
             return;
         }
+        // ...and the same INVALID_OPERATION covers "the command would source data beyond the end
+        // of the bound buffer object" (GL 4.6 core 19): the dispatch reads three uints starting
+        // at `indirect`.
+        constexpr SizeT kDispatchIndirectCommandSize = 3 * sizeof(Uint32);
+        if (static_cast<SizeT>(indirect) + kDispatchIndirectCommandSize > indirectBuffer->GetSize()) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidOperation,
+                MakeUnique<GenericErrorInfo>(
+                    "MG_Impl/GLImpl", __func__,
+                    std::format("indirect ({}) + 12 bytes runs past the end of the {}-byte buffer bound to "
+                                "GL_DISPATCH_INDIRECT_BUFFER.",
+                                indirect, indirectBuffer->GetSize())));
+            return;
+        }
+        auto dispatchComputeIndirect = MG_Backend::gBackendFunctionsTable.GL.DispatchComputeIndirect;
+        if (!dispatchComputeIndirect) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidOperation,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", __func__,
+                                             "Backend does not support indirect compute dispatch."));
+            return;
+        }
+        if (!ValidateCurrentProgramForCompute(__func__)) return;
         dispatchComputeIndirect(indirect);
     }
 
@@ -580,8 +607,80 @@ namespace MobileGL::MG_Impl::GLImpl {
         MultiDrawArraysIndirect_Backend(mode, indirect, drawcount, stride);
     }
 
+    // ARB_indirect_parameters / GL 4.6 core 10.4: `drawcount` is a byte offset into the buffer
+    // bound to PARAMETER_BUFFER and holds one uint draw count. Three errors have to be raised
+    // before the call reaches a backend, and none of them was
+    // (KHR-GL46.indirect_parameters_tests.MultiDraw{Arrays,Elements}IndirectCount):
+    //   * drawcount not a multiple of four                                  INVALID_VALUE
+    //   * nothing bound to PARAMETER_BUFFER, or the uint at `drawcount`
+    //     lies past its end                                                 INVALID_OPERATION
+    //   * maxdrawcount commands from `indirect` run past the end of the
+    //     buffer bound to DRAW_INDIRECT_BUFFER                              INVALID_OPERATION
+    static Bool ValidateIndirectCountDraw(GLintptr indirect, GLintptr drawcount, GLsizei maxdrawcount,
+                                          GLsizei stride, SizeT commandSize, const char* funcName) {
+        if (drawcount < 0 || (drawcount % 4) != 0) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidValue,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", funcName,
+                                             "drawcount must be non-negative and a multiple of four."));
+            return false;
+        }
+        const auto& parameterBuffer =
+            MG_State::pGLContext->GetBufferBindingSlot(BufferTarget::Parameter).GetBoundObject();
+        if (!parameterBuffer ||
+            static_cast<SizeT>(drawcount) + sizeof(Uint32) > parameterBuffer->GetSize()) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidOperation,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", funcName,
+                                             "No buffer is bound to GL_PARAMETER_BUFFER, or drawcount runs past "
+                                             "the end of the one that is."));
+            return false;
+        }
+        if (maxdrawcount < 0 || stride < 0 || indirect < 0) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidValue,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", funcName,
+                                             "indirect, maxdrawcount and stride must all be non-negative."));
+            return false;
+        }
+        const SizeT effectiveStride = stride != 0 ? static_cast<SizeT>(stride) : commandSize;
+        const auto& indirectBuffer =
+            MG_State::pGLContext->GetBufferBindingSlot(BufferTarget::DrawIndirect).GetBoundObject();
+        // A zero maxdrawcount sources nothing, so it cannot run past anything.
+        const SizeT requiredBytes =
+            maxdrawcount == 0 ? 0
+                              : static_cast<SizeT>(indirect) +
+                                    static_cast<SizeT>(maxdrawcount - 1) * effectiveStride + commandSize;
+        if (!indirectBuffer || requiredBytes > indirectBuffer->GetSize()) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidOperation,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", funcName,
+                                             "maxdrawcount commands would be sourced from beyond the end of the "
+                                             "buffer bound to GL_DRAW_INDIRECT_BUFFER."));
+            return false;
+        }
+        return true;
+    }
+
     void MultiDrawElementsIndirectCount(GLenum mode, GLenum type, const void* indirect, GLintptr drawcount,
                                         GLsizei maxdrawcount, GLsizei stride) {
+        // Argument validation before the backend-availability check: see DispatchComputeIndirect.
+        // DrawElementsIndirectCommand: count, instanceCount, firstIndex, baseVertex, baseInstance.
+        if (!ValidateIndirectCountDraw(reinterpret_cast<GLintptr>(indirect), drawcount, maxdrawcount, stride,
+                                       5 * sizeof(Uint32), __func__)) {
+            return;
+        }
+        // The only two draw entry points that were missing this. Every backend draw path
+        // dereferences GetProgramForDraw() unconditionally, so "no current program" has to be
+        // stopped here or it is a null dereference rather than the INVALID_OPERATION the spec
+        // asks for - reachable through a bound pipeline that supplies no graphics stage.
+        //
+        // AFTER the argument checks, unlike the sibling draw entry points, and deliberately:
+        // the argument rules here are properties of the call rather than of GL state, and
+        // NegativeApiErrorsTest.IndirectParameterDrawsCheckBothBuffers pins the INVALID_VALUE
+        // they produce for a call made with no program bound. Same precedence decision, and
+        // the same reason, as DispatchComputeIndirect above.
+        if (!ValidateCurrentProgramForExecution(__func__)) return;
         auto multiDrawElementsIndirectCount = MG_Backend::gBackendFunctionsTable.GL.MultiDrawElementsIndirectCount;
         if (!multiDrawElementsIndirectCount) {
             MG_State::pGLContext->RecordError(
@@ -595,6 +694,14 @@ namespace MobileGL::MG_Impl::GLImpl {
 
     void MultiDrawArraysIndirectCount(GLenum mode, const void* indirect, GLintptr drawcount,
                                       GLsizei maxdrawcount, GLsizei stride) {
+        // Argument validation before the backend-availability check: see DispatchComputeIndirect.
+        // DrawArraysIndirectCommand: count, instanceCount, first, baseInstance.
+        if (!ValidateIndirectCountDraw(reinterpret_cast<GLintptr>(indirect), drawcount, maxdrawcount, stride,
+                                       4 * sizeof(Uint32), __func__)) {
+            return;
+        }
+        // See MultiDrawElementsIndirectCount, including why this one goes last.
+        if (!ValidateCurrentProgramForExecution(__func__)) return;
         auto multiDrawArraysIndirectCount = MG_Backend::gBackendFunctionsTable.GL.MultiDrawArraysIndirectCount;
         if (!multiDrawArraysIndirectCount) {
             MG_State::pGLContext->RecordError(

MobileGL/MG_Impl/GLImpl/Getter/GL_Getter.cpp

@@ -722,10 +722,14 @@ namespace MobileGL::MG_Impl::GLImpl {
                     *data = 0;
                     return;
                 }
+                // GL 4.6 core table 23.4/23.5: *_BUFFER_SIZE reports the size glBindBufferRange
+                // was ASKED for, verbatim. It is not clamped to the buffer's storage, and it does
+                // not follow the buffer when a later glBufferData resizes it - a range may legally
+                // name bytes the buffer does not have yet. Clamping it here answered 0 for the
+                // common conformance shape of binding a range on a buffer that has no storage
+                // yet (KHR-GL43.shader_storage_buffer_object.basic-binding).
                 const Range1D range = bindingPoint.GetRange();
-                const auto start = std::min(range.start, bufferObject->GetSize());
-                const auto end = std::min(range.end, bufferObject->GetSize());
-                *data = static_cast<GLint>(end - start);
+                *data = static_cast<GLint>(range.end - range.start);
                 return;
             }
             default:
@@ -951,9 +955,8 @@ namespace MobileGL::MG_Impl::GLImpl {
                     *data = 0;
                     return;
                 }
-                const auto start = std::min(range.start, bufferObject->GetSize());
-                const auto end = std::min(range.end, bufferObject->GetSize());
-                *data = static_cast<GLint64>(end - start);
+                // Verbatim, unclamped - see the GetIntegeri_v arm.
+                *data = static_cast<GLint64>(range.end - range.start);
                 return;
             }
             default:
@@ -961,15 +964,30 @@ namespace MobileGL::MG_Impl::GLImpl {
             }
         }
 
-        auto getInteger64i = MG_Backend::gBackendFunctionsTable.GL.GetInteger64i_v;
-        if (!getInteger64i) {
-            *data = 0;
-            MG_State::pGLContext->RecordError(
-                ErrorCode::InvalidOperation,
-                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", __func__, "Backend does not support indexed integer queries."));
+        // The one indexed pname whose value genuinely needs 64 bits: a vertex buffer binding
+        // offset is an intptr, so taking the 32-bit route below would truncate it.
+        if (target == GL_VERTEX_BINDING_OFFSET) {
+            if (index >= VertexArrayImpl::GetMaxVertexAttribBindings()) {
+                MG_State::pGLContext->RecordError(
+                    ErrorCode::InvalidValue,
+                    MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", __func__,
+                                                 "Vertex buffer binding index is out of range."));
+                return;
+            }
+            const auto& vao = MG_State::pGLContext->GetBoundVertexArray();
+            *data = vao ? static_cast<GLint64>(vao->GetBindingPoint(index).Offset) : 0;
             return;
         }
-        getInteger64i(target, index, data);
+
+        // Everything else is 32-bit indexed state that the glGetIntegeri_v pname table already
+        // owns, and GL 4.6 core 22.1 says every indexed query answers every indexed pname.
+        // Handing the leftovers straight to the backend instead made glGetInteger64i_v disagree
+        // with glGetIntegeri_v on the very same pname - GL_MAX_COMPUTE_WORK_GROUP_COUNT read
+        // back 0 while the 32-bit view said 65535 (KHR-GL43.compute_shader.max), because a
+        // frontend-only value simply is not in the driver's table.
+        GLint values[4] = {};
+        GetIntegeri_v(target, index, values);
+        *data = static_cast<GLint64>(values[0]);
     }
 
     void GetInteger64v(GLenum pname, GLint64* params) {

MobileGL/MG_Impl/GLImpl/Program/GL_Program.cpp

@@ -744,6 +744,21 @@ namespace MobileGL::MG_Impl::GLImpl {
         CopyStr(bufSize, length, infoLog, log.c_str(), (GLsizei)log.length());
     }
 
+    // MOBILEGL_ASYNC_OPTIMISTIC_SHADER_STATUS: while the compile job is still in flight -
+    // and, via the latch below, for the rest of that node's life once any query was
+    // answered this way - GL_COMPILE_STATUS reads GL_TRUE and the info log reads empty,
+    // WITHOUT joining. The latch (TakeOptimisticCompileAnswer) is what makes the three
+    // sites tell ONE story: without it, a job settling between an application's info-log
+    // read and its status read would produce the torn pair "GL_FALSE with an empty log",
+    // and an application that aborts on that never reaches the link join that carries the
+    // real diagnostic. A failure hidden here still fails the program link, with the
+    // compile log quoted in the program info log (ProgramLinkTask::ConsumeShaders), which
+    // is where the serial compile-then-check applications this exists for do their error
+    // handling.
+    static Bool AnswerCompileOptimistically(const SharedPtr<MG_State::GLState::ShaderObject>& shaderObject) {
+        return MG_Util::Async::OptimisticShaderStatusActive() && shaderObject->TakeOptimisticCompileAnswer();
+    }
+
     void GetShaderiv_State(GLuint shader, GLenum pname, GLint* params) {
         auto& shaderObject = TryToGetShaderObject(shader);
         if (!shaderObject) return;
@@ -756,9 +771,20 @@ namespace MobileGL::MG_Impl::GLImpl {
             *params = shaderObject->GetDeleteStatus();
             break;
         case GL_COMPILE_STATUS:
+            if (AnswerCompileOptimistically(shaderObject)) {
+                *params = GL_TRUE;
+                break;
+            }
             *params = shaderObject->GetCompileStatus();
             break;
         case GL_INFO_LOG_LENGTH:
+            // Not cosmetic: LWJGL's one-argument glGetShaderInfoLog convenience overload
+            // sizes its buffer from this query, so a joining answer here would defeat the
+            // non-joining GetShaderInfoLog below.
+            if (AnswerCompileOptimistically(shaderObject)) {
+                *params = 0;
+                break;
+            }
             *params = shaderObject->GetInfoLog().empty() ? 0 : (GLint)shaderObject->GetInfoLog().length() + 1;
             break;
         case GL_SHADER_SOURCE_LENGTH:
@@ -784,6 +810,15 @@ namespace MobileGL::MG_Impl::GLImpl {
         auto& shaderObject = TryToGetShaderObject(shader);
         if (!shaderObject) return;
 
+        // See AnswerCompileOptimistically: an in-flight compile reads as an empty log. The
+        // cost is a lost compile WARNING (a successful compile whose log the application
+        // reads exactly once, now, and never after the join) - accepted as part of the
+        // opt-in.
+        if (AnswerCompileOptimistically(shaderObject)) {
+            CopyStr(bufSize, length, infoLog, "", 0);
+            return;
+        }
+
         const auto& log = shaderObject->GetInfoLog();
         CopyStr(bufSize, length, infoLog, log.c_str(), (GLsizei)log.length());
     }
@@ -828,12 +863,11 @@ namespace MobileGL::MG_Impl::GLImpl {
     }
 
     // Bytes a uniform actually occupies in the global UBO. It is the tight GL type size for
-    // everything except a float matrix, whose padded columns make it wider.
+    // everything except a float matrix, whose padded columns make it wider. The rule itself
+    // lives on ProgramObject, because the pipeline composite's uniform refresh needs the same
+    // one and two copies of a layout rule is one too many.
     SizeT UniformStorageSpanInBytes(const glslang::TType* ttype, SizeT tightSize) {
-        if (ttype != nullptr && ttype->isMatrix() && ttype->getBasicType() != glslang::EbtDouble) {
-            return static_cast<SizeT>(ttype->getMatrixCols()) * 4 * sizeof(GLfloat);
-        }
-        return tightSize;
+        return MG_State::GLState::ProgramObject::UniformStorageSpanInBytes(ttype, tightSize);
     }
 
     void GetUniform_State(GLuint program, GLint location, void* params) {
@@ -1085,10 +1119,20 @@ namespace MobileGL::MG_Impl::GLImpl {
         if (!programObject.IsUniformOpaqueAtLocation(location)) {
             MGLOG_D("%s: program = %d, location = %d, maxLocation = %d", __func__, programObject.GetExternalIndex(),
                     location, programObject.GetMaxUniformLocation());
+            // Record the write for the pipeline composite's uniform mirror, which copies only
+            // the locations a stage program has actually been written to (see
+            // ProgramObject::MarkUniformWrittenAtLocation). Here rather than further down
+            // because every exit below is still a write as far as GL is concerned: the
+            // buffered-write detour returns early, the bytes-equal dedupe returns early, and
+            // even the no-backing-storage bail is a uniform the application addressed. This is
+            // the funnel EVERY glUniform* and glProgramUniform* entry point reaches, once per
+            // LOCATION - so an array element write marks that element and nothing else. On a
+            // program that can never be a pipeline stage - the monolithic glUseProgram path,
+            // which is where the thousands of calls per frame are - this is one bool branch.
+            programObject.MarkUniformWrittenAtLocation(location);
+            // Everything up to and including the clamp is phase-A data (the uniform's GL type
+            // decides its size), so it is answered without joining anything.
             const SizeT size = programObject.GetUniformSizesInBytes(location);
-            const Uint offset = programObject.GetUniformOffset(location);
-            char* pUBO = static_cast<char*>(programObject.MapUBO());
-            const SizeT uboSize = programObject.GetUBOSize();
             SizeT writeSize = ItemCount * sizeof(T);
             if (size < writeSize) {
                 // Metadata bug: degrade to a clamped copy instead of killing the process.
@@ -1097,6 +1141,18 @@ namespace MobileGL::MG_Impl::GLImpl {
                         __func__, programObject.GetExternalIndex(), location, ItemCount * sizeof(T), size);
                 writeSize = size;
             }
+            // The uniform shadow's LAYOUT is phase-B data, so a write that lands while the
+            // SPIR-V job is still running is recorded and replayed at its publish instead of
+            // joining it. This is the hot path for a shaderpack that sets its uniforms
+            // immediately after glLinkProgram. BufferUniformWrite declines (and we fall
+            // through, joining) only past its size budget.
+            if (programObject.IsSpirvPending() &&
+                programObject.BufferUniformWrite(location, byteOffsetInsideUniform, value, writeSize)) {
+                return;
+            }
+            const Uint offset = programObject.GetUniformOffset(location);
+            char* pUBO = static_cast<char*>(programObject.MapUBO());
+            const SizeT uboSize = programObject.GetUBOSize();
             if (pUBO == nullptr || offset == MG_State::GLState::ProgramObject::kInvalidUniformOffset ||
                 offset + byteOffsetInsideUniform + writeSize > uboSize) {
                 // Should not happen: linking gives every settable uniform backing
@@ -2658,6 +2714,15 @@ namespace MobileGL::MG_Impl::GLImpl {
 
     void GetProgramResourceiv(GLuint program, GLenum programInterface, GLuint index, GLsizei propCount,
                               const GLenum* props, GLsizei bufSize, GLsizei* length, GLint* params) {
+        // Every early-out below reports "nothing was written", and it has to say so before it can
+        // take one: callers legitimately leave *length uninitialised and then loop to it. The CTS
+        // does exactly that (gl4cProgramInterfaceQueryTests.cpp:2172 declares `GLsizei length;` and
+        // walks `for (i = 0; i < length; ++i)` over a 1000-entry stack array), so an untouched
+        // *length turned every error path here into a stack overrun inside the caller -
+        // KHR-GL43.program_interface_query.subroutines-vertex read 0x20202020 entries and died on
+        // both backends. The success path overwrites this with the real count.
+        if (length) *length = 0;
+
         auto& programObject = TryToGetProgramForInterfaceQuery(program, __func__);
         if (!programObject) return;
         if (!ProgramInterface::IsInterfaceEnum(programInterface)) {

MobileGL/MG_Impl/GLImpl/Program/GL_ProgramPipeline.cpp

@@ -19,16 +19,26 @@ namespace MobileGL::MG_Impl::GLImpl {
                 code, MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", function, Move(message)));
         }
 
-        // A pipeline name only names an object once it has been bound or created; querying a
-        // reserved-but-unmaterialised name is INVALID_OPERATION (GL 4.6 core 7.4).
+        // GL 4.6 core 7.4 asks only that the name came from GenProgramPipelines and has not been
+        // deleted - so a name that was reserved and never bound is legal here, and the command
+        // MATERIALIZES it rather than rejecting it.
+        //
+        // Requiring a bound object instead is what broke every separable-program conformance case
+        // across three families: the CTS reserves a name, calls glUseProgramStages three times and
+        // only then binds, which is the order the spec's own example uses. Each of those calls
+        // failed with INVALID_OPERATION, so the stage programs were never recorded - the pipeline
+        // stayed empty, GetProgramForDraw flattened nothing and the draw painted nothing, and the
+        // rejected calls' error was left in the queue for the harness to find. One cause, both
+        // symptoms.
         const SharedPtr<MG_State::GLState::ProgramPipelineObject>* TryGetPipeline(GLuint pipeline,
                                                                                  const char* function) {
-            if (!MG_State::pGLContext->IsProgramPipelineObject(pipeline)) {
+            const auto& object = MG_State::pGLContext->MaterializeProgramPipelineObject(pipeline);
+            if (!object) {
                 RecordPipelineError(ErrorCode::InvalidOperation, function,
                                     std::format("Program pipeline {} does not exist.", pipeline));
                 return nullptr;
             }
-            return &MG_State::pGLContext->GetProgramPipelineObject(pipeline);
+            return &object;
         }
 
         Bool ValidatePipelineCount(GLsizei n, const char* function) {

MobileGL/MG_Impl/GLImpl/RenderState/GL_RenderState.cpp

@@ -8,6 +8,7 @@
 
 #include "GL_RenderState.h"
 #include <cmath>
+#include <MG_Impl/GLImpl/Getter/GL_Getter.h>
 #include <MG_State/GLState/Core.h>
 #include <MG_Util/Converters/GLToStr/GLEnumConverter.h>
 #include <MG_Util/Converters/GLToMG/RenderStateEnumConverter.h>
@@ -380,7 +381,18 @@ namespace MobileGL::MG_Impl::GLImpl {
             return;
         }
 
-        *data = IsEnabledi_State(target, index);
+        // GL 4.6 core 22.1: glGetBooleani_v answers EVERY indexed state, not just the indexed
+        // capabilities - a non-boolean value simply reads back as "is it non-zero". Routing the
+        // non-capability enums to the pname table glGetIntegeri_v already owns is what makes
+        // that true; without it a query like glGetBooleani_v(GL_MAX_COMPUTE_WORK_GROUP_COUNT, 0)
+        // came back GL_INVALID_ENUM (KHR-GL43.compute_shader.max).
+        if (MG_Util::ConvertGLEnumToCapabilityInput(target) != CapabilityInput::Unknown) {
+            *data = IsEnabledi_State(target, index);
+            return;
+        }
+        GLint values[4] = {};
+        GetIntegeri_v(target, index, values);
+        *data = values[0] != 0 ? GL_TRUE : GL_FALSE;
     }
 
     GLboolean IsEnabled_State(GLenum cap) {

MobileGL/MG_Impl/GLImpl/Sampler/GL_Sampler.cpp

@@ -336,8 +336,22 @@ namespace MobileGL::MG_Impl::GLImpl {
             return;
         }
 
+        // ARB_multi_bind adds one rule the single-bind path does not have: "samplers will not be
+        // created if they do not exist", so a name that is not an existing sampler OBJECT is
+        // INVALID_OPERATION here (KHR-GL44.multi_bind.errors_bind_samplers). Per element, not
+        // all-or-nothing - the extension defines glBindSamplers as a loop, so a bad entry costs
+        // its own texture unit and leaves the rest of the range bound.
         for (GLsizei i = 0; i < count; ++i) {
-            BindSampler_State(first + i, samplers ? samplers[i] : 0);
+            const GLuint sampler = samplers ? samplers[i] : 0;
+            if (sampler != 0 && !MG_State::pGLContext->ValidateSamplerObject(sampler)) {
+                MG_State::pGLContext->RecordError(
+                    ErrorCode::InvalidOperation,
+                    MakeUnique<GenericErrorInfo>(
+                        "MG_Impl/GLImpl", "BindSamplers",
+                        std::format("samplers[{}] ({}) is not the name of an existing sampler object.", i, sampler)));
+                continue;
+            }
+            BindSampler_State(first + i, sampler);
         }
     }
 

MobileGL/MG_Impl/GLImpl/Texture/GL_Texture.cpp

@@ -613,6 +613,23 @@ namespace MobileGL::MG_Impl::GLImpl {
                 MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", caller,
                                              "Compressed texture formats are not supported."));
         }
+
+        // glGetTexLevelParameter{i,f}v answers WIDTH/HEIGHT/DEPTH out of the mipmap chain. The only
+        // other storage type the state layer knows is GL_TEXTURE_BUFFER (TextureStorageType is
+        // {Mipmap, Buffer}), whose level geometry this stack does not track yet. Report that instead
+        // of throwing: THROW_UNIMPL_EXCEPTION unwinds a C++ exception through the C GL ABI and takes
+        // the process down, which is never an acceptable answer to a query - see the same reasoning
+        // above for the compressed-format path.
+        void RecordUnsupportedLevelQueryStorage(const char* caller, GLenum pname) {
+            MGLOG_I("%s: glGetTexLevelParameter(pname=%s) is not implemented for texture-buffer "
+                    "storage; recording GL_INVALID_OPERATION instead of terminating",
+                    caller, MG_Util::ConvertGLEnumToString(pname).c_str());
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidOperation,
+                MakeUnique<GenericErrorInfo>(
+                    "MG_Impl/GLImpl", caller,
+                    "Level queries are not supported for texture-buffer storage."));
+        }
     } // namespace
 
     const SharedPtr<MG_State::GLState::ITextureObject>& GetTextureObjectByName(GLuint texture, const char* caller) {
@@ -2910,7 +2927,8 @@ namespace MobileGL::MG_Impl::GLImpl {
                     break;
                 }
                 default:
-                    THROW_UNIMPL_EXCEPTION;
+                    RecordUnsupportedLevelQueryStorage("GetTexLevelParameteriv_State", pname);
+                    break;
                 }
             }
             break;
@@ -2924,7 +2942,8 @@ namespace MobileGL::MG_Impl::GLImpl {
                     break;
                 }
                 default:
-                    THROW_UNIMPL_EXCEPTION;
+                    RecordUnsupportedLevelQueryStorage("GetTexLevelParameteriv_State", pname);
+                    break;
                 }
             }
             break;
@@ -2938,7 +2957,8 @@ namespace MobileGL::MG_Impl::GLImpl {
                     break;
                 }
                 default:
-                    THROW_UNIMPL_EXCEPTION;
+                    RecordUnsupportedLevelQueryStorage("GetTexLevelParameteriv_State", pname);
+                    break;
                 }
             }
             break;
@@ -3045,7 +3065,8 @@ namespace MobileGL::MG_Impl::GLImpl {
                     break;
                 }
                 default:
-                    THROW_UNIMPL_EXCEPTION;
+                    RecordUnsupportedLevelQueryStorage("GetTexLevelParameterfv_State", pname);
+                    break;
                 }
             }
             break;
@@ -3059,7 +3080,8 @@ namespace MobileGL::MG_Impl::GLImpl {
                     break;
                 }
                 default:
-                    THROW_UNIMPL_EXCEPTION;
+                    RecordUnsupportedLevelQueryStorage("GetTexLevelParameterfv_State", pname);
+                    break;
                 }
             }
             break;
@@ -3073,7 +3095,8 @@ namespace MobileGL::MG_Impl::GLImpl {
                     break;
                 }
                 default:
-                    THROW_UNIMPL_EXCEPTION;
+                    RecordUnsupportedLevelQueryStorage("GetTexLevelParameterfv_State", pname);
+                    break;
                 }
             }
             break;
@@ -3403,7 +3426,10 @@ namespace MobileGL::MG_Impl::GLImpl {
             GET_SRC_INTERNAL_FORMAT(readBufferType);
         }
 
-        if (!TextureImpl::ValidateBaseInternalFormatMatch(internalFormat, srcInternalFormat)) THROW_UNIMPL_EXCEPTION;
+        // The validator has already recorded GL_INVALID_OPERATION; just decline. Throwing
+        // here unwound a C++ exception through the C GL ABI and killed the process (see the
+        // same reasoning at :604-609).
+        if (!TextureImpl::ValidateCopyTexImageBaseFormatSubset(internalFormat, srcInternalFormat)) return false;
 
         GLenum outInternalFormat = MG_Util::ConvertTextureInternalFormatToGLEnum(srcInternalFormat);
         GLenum realInternalFormat = GL_RGBA8;
@@ -3426,8 +3452,13 @@ namespace MobileGL::MG_Impl::GLImpl {
 
     void CopyTexImage1D_State(GLenum target, GLint level, GLenum internalformat, GLint x, GLint y, GLsizei width,
                               GLint border) {
-        // TODO: implement
-        THROW_UNIMPL_EXCEPTION;
+        // 1D textures are not implemented by this backend set. Record the error the way every
+        // other unsupported entry point does - throwing unwinds through the C GL ABI and kills
+        // the process, which is never an acceptable answer to an unsupported call.
+        MG_State::pGLContext->RecordError(
+            ErrorCode::InvalidOperation,
+            MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", "CopyTexImage1D",
+                                         "1D textures are not supported by this implementation"));
     }
 
     void CompressedTexSubImage3D_State(GLenum target, GLint level, GLint xoffset, GLint yoffset, GLint zoffset,
@@ -4079,10 +4110,46 @@ namespace MobileGL::MG_Impl::GLImpl {
         textureObject->SetImmutableLevels(static_cast<Uint>(levels));
     }
 
+    // No block-compressed format is defined for a three-dimensional image, so glTexStorage3D on
+    // TEXTURE_3D must reject one - and with INVALID_OPERATION, not the INVALID_ENUM an unknown
+    // sized format gets (GL 4.6 core 8.19 / Khronos bug 11239, KHR-GLxx.texture_storage
+    // .compressed_data). Written against the enum ranges rather than a name list because the
+    // families are contiguous and MobileGL's own internal-format enum drops the ones it cannot
+    // carry, which would make this check silently narrower than the API surface.
+    static Bool IsCompressedGLInternalFormat(GLenum internalformat) {
+        switch (internalformat) {
+        case 0x8225: // GL_COMPRESSED_RED
+        case 0x8226: // GL_COMPRESSED_RG
+        case 0x84ED: // GL_COMPRESSED_RGB
+        case 0x84EE: // GL_COMPRESSED_RGBA
+        case 0x8C48: // GL_COMPRESSED_SRGB
+        case 0x8C49: // GL_COMPRESSED_SRGB_ALPHA
+            return true;
+        default:
+            break;
+        }
+        return (internalformat >= 0x83F0 && internalformat <= 0x83F3) || // S3TC / DXT
+               (internalformat >= 0x8DBB && internalformat <= 0x8DBE) || // RGTC
+               (internalformat >= 0x8E8C && internalformat <= 0x8E8F) || // BPTC
+               (internalformat >= 0x9270 && internalformat <= 0x9279) || // ETC2 / EAC
+               (internalformat >= 0x93B0 && internalformat <= 0x93BD) || // ASTC LDR
+               (internalformat >= 0x93D0 && internalformat <= 0x93DD);   // ASTC sRGB
+    }
+
     void TextureStorage3D(GLuint texture, GLsizei levels, GLenum internalformat, GLsizei width, GLsizei height,
                           GLsizei depth) {
         auto textureObject = GetTextureObjectByName(texture, __func__);
         if (!textureObject) return;
+        if (textureObject->GetTarget() == TextureTarget::Texture3D &&
+            IsCompressedGLInternalFormat(internalformat)) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidOperation,
+                MakeUnique<GenericErrorInfo>(
+                    "MG_Impl/GLImpl", __func__,
+                    std::format("{} is a compressed internal format and cannot back GL_TEXTURE_3D storage.",
+                                MG_Util::ConvertGLEnumToString(internalformat))));
+            return;
+        }
         TextureInternalFormat textureInternalFormat = MG_Util::ConvertGLEnumToTextureInternalFormat(internalformat);
         if (!ValidateTextureStorageInternalFormat(textureInternalFormat, __func__)) return;
         if (!ValidateTextureStorageShape(textureObject, 3, levels, width, height, depth, __func__)) return;

MobileGL/MG_Impl/GLImpl/Texture/Validators.cpp

@@ -424,19 +424,81 @@ namespace MobileGL::MG_Impl::GLImpl::TextureImpl {
         return true;
     }
 
+    namespace {
+        // Component set of an UNSIZED base internal format, as the bitmask GL 4.6 SS 8.6
+        // reasons about. Colour components are independent bits so "subset" is a plain
+        // mask test; depth and stencil are their own components and never satisfy a
+        // colour request (or each other).
+        enum : Uint32 {
+            kComponentR = 1u << 0,
+            kComponentG = 1u << 1,
+            kComponentB = 1u << 2,
+            kComponentA = 1u << 3,
+            kComponentDepth = 1u << 4,
+            kComponentStencil = 1u << 5,
+        };
+
+        Uint32 BaseFormatComponents(TextureInternalFormat unsizedFormat) {
+            switch (unsizedFormat) {
+            case TextureInternalFormat::Red:
+                return kComponentR;
+            case TextureInternalFormat::RG:
+                return kComponentR | kComponentG;
+            case TextureInternalFormat::RGB:
+                return kComponentR | kComponentG | kComponentB;
+            case TextureInternalFormat::RGBA:
+                return kComponentR | kComponentG | kComponentB | kComponentA;
+            case TextureInternalFormat::DepthComponent:
+                return kComponentDepth;
+            case TextureInternalFormat::DepthStencil:
+                return kComponentDepth | kComponentStencil;
+            default:
+                return 0;
+            }
+        }
+    } // namespace
+
     Bool ValidateBaseInternalFormatMatch(TextureInternalFormat format1, TextureInternalFormat format2) {
-        auto unsizedFormat1 = MG_Util::ConvertInternalFormatToUnsized(format1);
-        auto unsizedFormat2 = MG_Util::ConvertInternalFormatToUnsized(format2);
+        const auto unsizedFormat1 = MG_Util::ConvertInternalFormatToUnsized(format1);
+        const auto unsizedFormat2 = MG_Util::ConvertInternalFormatToUnsized(format2);
         if (unsizedFormat1 != unsizedFormat2) {
+            // The 3-argument GenericErrorInfo constructor used to be spelled as a single
+            // std::format() call whose format string was the component name, so every
+            // diagnostic collapsed to the literal "MG_Impl/GLImpl". Format the message, then
+            // hand over component/function/message separately.
             MG_State::pGLContext->RecordError(
                 ErrorCode::InvalidOperation,
                 MakeUnique<GenericErrorInfo>(
-                    std::format("MG_Impl/GLImpl", "ValidateBaseInternalFormatMatch",
-                                "The base internal format of the two formats do not match ({} vs. {})",
-                                MG_Util::ConvertTextureInternalFormatToString(unsizedFormat1).c_str(),
-                                MG_Util::ConvertTextureInternalFormatToString(unsizedFormat2).c_str())));
+                    "MG_Impl/GLImpl", "ValidateBaseInternalFormatMatch",
+                    std::format("The base internal format of the two formats do not match ({} vs. {})",
+                                MG_Util::ConvertTextureInternalFormatToString(unsizedFormat1),
+                                MG_Util::ConvertTextureInternalFormatToString(unsizedFormat2))));
             return false;
         }
         return true;
-    } // namespace TextureImpl
+    }
+
+    Bool ValidateCopyTexImageBaseFormatSubset(TextureInternalFormat destFormat, TextureInternalFormat srcFormat) {
+        const auto unsizedDest = MG_Util::ConvertInternalFormatToUnsized(destFormat);
+        const auto unsizedSrc = MG_Util::ConvertInternalFormatToUnsized(srcFormat);
+        // GL 4.6 SS 8.6: glCopyTexImage* may request a SUBSET of the read buffer's components,
+        // not an exact match - GL_RGB from an RGBA8 framebuffer is textbook legal and is what
+        // Minecraft and its mods do. glCopyTexImage2D used to run the exact-match predicate
+        // above and turn its rejection into an uncaught exception through the C GL ABI, so the
+        // app died rather than seeing a GL error.
+        const Uint32 destComponents = BaseFormatComponents(unsizedDest);
+        const Uint32 srcComponents = BaseFormatComponents(unsizedSrc);
+        if (destComponents == 0 || srcComponents == 0 || (destComponents & ~srcComponents) != 0) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidOperation,
+                MakeUnique<GenericErrorInfo>(
+                    "MG_Impl/GLImpl", "ValidateCopyTexImageBaseFormatSubset",
+                    std::format("the read buffer's base internal format {} does not provide every component of "
+                                "the requested internal format {}",
+                                MG_Util::ConvertTextureInternalFormatToString(unsizedSrc),
+                                MG_Util::ConvertTextureInternalFormatToString(unsizedDest))));
+            return false;
+        }
+        return true;
+    }
 } // namespace MobileGL::MG_Impl::GLImpl::TextureImpl

MobileGL/MG_Impl/GLImpl/Texture/Validators.h

@@ -40,5 +40,9 @@ namespace MobileGL::MG_Impl::GLImpl::TextureImpl {
                                          TextureTarget target);
     Bool ValidateTextureSubImageOffsets(const SharedPtr<MG_State::GLState::ITextureObject>& textureObject, Int xoffset,
                                         Int width, Int yoffset = 0, Int height = 0, Int zoffset = 0, Int depth = 0);
+    // Exact base-format equality - what glCopyImageSubData's format compatibility needs.
     Bool ValidateBaseInternalFormatMatch(TextureInternalFormat format1, TextureInternalFormat format2);
+    // GL 4.6 SS 8.6 subset rule for glCopyTexImage*: the read buffer must supply every component
+    // the requested internalformat asks for, but may supply more.
+    Bool ValidateCopyTexImageBaseFormatSubset(TextureInternalFormat destFormat, TextureInternalFormat srcFormat);
 } // namespace MobileGL::MG_Impl::GLImpl::TextureImpl

MobileGL/MG_Impl/GLImpl/VertexArray/GL_VertexArray.cpp

@@ -179,6 +179,28 @@ namespace MobileGL::MG_Impl::GLImpl {
             return vao;
         }
 
+        // The ARB_vertex_attrib_binding entry points that take no vertex array name modify the
+        // *bound* vertex array, and in a core profile the default vertex array (name 0) is not
+        // one: every one of them is INVALID_OPERATION there (GL 4.6 core 10.3.1, and the tail of
+        // each KHR-GL4x.vertex_attrib_binding.negative-* case checks exactly this). MobileGL
+        // keeps a real object at name 0 for the compatibility paths, so GetBoundVertexArray
+        // never returns null and the rule has to be spelled out - behind the same gate the VAO-0
+        // draw rule already uses (MOBILEGL_RELAXED_SEMANTICS, plus "the context never asked for
+        // a core profile"), so applications that legitimately run relaxed keep working.
+        static SharedPtr<MG_State::GLState::VertexArrayObject> GetBoundVertexArrayForBindingApi(const char* funcName) {
+            auto vao = GetBoundVertexArrayOrError(funcName);
+            if (!vao) return nullptr;
+            if (vao->GetExternalIndex() == 0 && !MG_State::IsRelaxedSemanticsActive()) {
+                MG_State::pGLContext->RecordError(
+                    ErrorCode::InvalidOperation,
+                    MakeUnique<GenericErrorInfo>(
+                        "MG_Impl/GLImpl", funcName,
+                        "The default vertex array object cannot be modified in a core profile."));
+                return nullptr;
+            }
+            return vao;
+        }
+
         static bool ValidateVertexAttribPname(GLenum pname) {
             switch (pname) {
             case GL_VERTEX_ATTRIB_ARRAY_ENABLED:
@@ -944,7 +966,7 @@ namespace MobileGL::MG_Impl::GLImpl {
             params[0] = static_cast<GLfloat>(attr->Size);
             return;
         case GL_VERTEX_ATTRIB_ARRAY_STRIDE:
-            params[0] = static_cast<GLfloat>(attr->Stride);
+            params[0] = static_cast<GLfloat>(attr->LegacyStride);
             return;
         case GL_VERTEX_ATTRIB_ARRAY_TYPE:
             params[0] = static_cast<GLfloat>(MG_Util::ConvertDataTypeToGLEnum(attr->Type));
@@ -1014,7 +1036,7 @@ namespace MobileGL::MG_Impl::GLImpl {
             params[0] = static_cast<GLdouble>(attr->Size);
             return;
         case GL_VERTEX_ATTRIB_ARRAY_STRIDE:
-            params[0] = static_cast<GLdouble>(attr->Stride);
+            params[0] = static_cast<GLdouble>(attr->LegacyStride);
             return;
         case GL_VERTEX_ATTRIB_ARRAY_TYPE:
             params[0] = static_cast<GLdouble>(MG_Util::ConvertDataTypeToGLEnum(attr->Type));
@@ -1079,8 +1101,11 @@ namespace MobileGL::MG_Impl::GLImpl {
         case GL_VERTEX_ATTRIB_ARRAY_SIZE:
             params[0] = attr->Size;
             return;
+        // The legacy shadow, not the resolved draw stride: GL 4.6 core table 23.3 defines this
+        // as the last glVertexAttrib*Pointer argument, which glBindVertexBuffer must not
+        // overwrite even though it does overwrite what the backend actually reads.
         case GL_VERTEX_ATTRIB_ARRAY_STRIDE:
-            params[0] = attr->Stride;
+            params[0] = attr->LegacyStride;
             return;
         case GL_VERTEX_ATTRIB_ARRAY_TYPE:
             params[0] = static_cast<GLint>(MG_Util::ConvertDataTypeToGLEnum(attr->Type));
@@ -1138,7 +1163,7 @@ namespace MobileGL::MG_Impl::GLImpl {
         }
 
         const auto& attr = vao->GetAttribute(index);
-        *pointer = reinterpret_cast<void*>(attr.Offset);
+        *pointer = reinterpret_cast<void*>(attr.LegacyPointer);
     }
 
     void GetVertexAttribIiv(GLuint index, GLenum pname, GLint* params) {
@@ -1222,7 +1247,7 @@ namespace MobileGL::MG_Impl::GLImpl {
             *param = static_cast<GLint>(attr.Size);
             return;
         case GL_VERTEX_ATTRIB_ARRAY_STRIDE:
-            *param = static_cast<GLint>(attr.Stride);
+            *param = static_cast<GLint>(attr.LegacyStride);
             return;
         case GL_VERTEX_ATTRIB_ARRAY_TYPE:
             *param = static_cast<GLint>(MG_Util::ConvertDataTypeToGLEnum(attr.Type));
@@ -1294,14 +1319,14 @@ namespace MobileGL::MG_Impl::GLImpl {
     }
 
     void BindVertexBuffer(GLuint bindingindex, GLuint buffer, GLintptr offset, GLsizei stride) {
-        auto vao = GetBoundVertexArrayOrError("BindVertexBuffer");
+        auto vao = GetBoundVertexArrayForBindingApi("BindVertexBuffer");
         if (!vao) return;
         VertexBufferBinding_State(vao, bindingindex, buffer, offset, stride, "BindVertexBuffer");
     }
 
     void BindVertexBuffers(GLuint first, GLsizei count, const GLuint* buffers, const GLintptr* offsets,
                            const GLsizei* strides) {
-        auto vao = GetBoundVertexArrayOrError("BindVertexBuffers");
+        auto vao = GetBoundVertexArrayForBindingApi("BindVertexBuffers");
         if (!vao) return;
         if (!ValidateVertexBindingRange(first, count, "BindVertexBuffers")) return;
         for (GLsizei i = 0; i < count; ++i) {
@@ -1315,21 +1340,21 @@ namespace MobileGL::MG_Impl::GLImpl {
     }
 
     void VertexAttribFormat(GLuint attribindex, GLint size, GLenum type, GLboolean normalized, GLuint relativeoffset) {
-        auto vao = GetBoundVertexArrayOrError("VertexAttribFormat");
+        auto vao = GetBoundVertexArrayForBindingApi("VertexAttribFormat");
         if (!vao) return;
         VertexAttribFormatSeparate_State(vao, attribindex, size, type, normalized, relativeoffset, false,
                                          "VertexAttribFormat");
     }
 
     void VertexAttribIFormat(GLuint attribindex, GLint size, GLenum type, GLuint relativeoffset) {
-        auto vao = GetBoundVertexArrayOrError("VertexAttribIFormat");
+        auto vao = GetBoundVertexArrayForBindingApi("VertexAttribIFormat");
         if (!vao) return;
         VertexAttribFormatSeparate_State(vao, attribindex, size, type, GL_FALSE, relativeoffset, true,
                                          "VertexAttribIFormat");
     }
 
     void VertexAttribLFormat(GLuint attribindex, GLint size, GLenum type, GLuint relativeoffset) {
-        auto vao = GetBoundVertexArrayOrError("VertexAttribLFormat");
+        auto vao = GetBoundVertexArrayForBindingApi("VertexAttribLFormat");
         if (!vao) return;
         VertexAttribLFormatSeparate_State(vao, attribindex, size, type, relativeoffset);
     }
@@ -1341,7 +1366,7 @@ namespace MobileGL::MG_Impl::GLImpl {
     }
 
     void VertexAttribBinding(GLuint attribindex, GLuint bindingindex) {
-        auto vao = GetBoundVertexArrayOrError("VertexAttribBinding");
+        auto vao = GetBoundVertexArrayForBindingApi("VertexAttribBinding");
         if (!vao) return;
         if (!VertexArrayImpl::ValidateVertexAttributeIndex(attribindex)) return;
         if (!ValidateVertexBindingIndex(bindingindex, "VertexAttribBinding")) return;
@@ -1349,7 +1374,7 @@ namespace MobileGL::MG_Impl::GLImpl {
     }
 
     void VertexBindingDivisor(GLuint bindingindex, GLuint divisor) {
-        auto vao = GetBoundVertexArrayOrError("VertexBindingDivisor");
+        auto vao = GetBoundVertexArrayForBindingApi("VertexBindingDivisor");
         if (!vao) return;
         if (!ValidateVertexBindingIndex(bindingindex, "VertexBindingDivisor")) return;
         vao->SetBindingDivisor(bindingindex, divisor);

MobileGL/MG_IntegrationTest/CMakeLists.txt

@@ -53,6 +53,21 @@ add_executable(MobileGLIntegrationTest
     Scenarios/AsyncCompileScenario.cpp
     Scenarios/XfbAfterClipDistanceScenario.cpp
     Scenarios/ThreeChannelAttachmentScenario.cpp
+    Scenarios/PipelineFailureScenario.cpp
+    Scenarios/AdvertisedLimitsScenario.cpp
+    Scenarios/PixelStoreSweepScenario.cpp
+    Scenarios/FragCoordOriginScenario.cpp
+    Scenarios/ClearThenReadPixelsScenario.cpp
+    Scenarios/DepthStencilReadbackScenario.cpp
+    Scenarios/SsboArrayLengthScenario.cpp
+    Scenarios/UniformInitializerScenario.cpp
+    Scenarios/SwizzleAccessRoutineScenario.cpp
+    Scenarios/ProgramPipelineScenario.cpp
+    Scenarios/ImageLoadStoreSsoScenario.cpp
+    Scenarios/SsboDeclarationFormScenario.cpp
+    Scenarios/Glsl420DeclarationScenario.cpp
+    Scenarios/FragmentOutputArrayIndexScenario.cpp
+    Scenarios/BufferTextureScenario.cpp
 )
 
 target_include_directories(MobileGLIntegrationTest PRIVATE
@@ -169,25 +184,24 @@ endif()
 option(MOBILEGL_ITEST_REQUIRE_GPU
     "Fail (rather than skip) the integration scenarios when the headless harness is unusable" OFF)
 
-# DirectGLES asks the system EGL for a pbuffer config, and on Mesa the default
-# platform is not X11 unless it is said out loud (run_driver_bench.sh sets the
-# same variable). Wrong platform here is not a soft failure: eglCreatePbuffer
-# fails and every scenario skips.
-if (UNIX AND NOT APPLE AND NOT ANDROID)
-    set(MOBILEGL_ITEST_EGL_PLATFORM "x11" CACHE STRING
-        "EGL_PLATFORM for the integration tests (empty: leave the loader alone)")
-else()
-    set(MOBILEGL_ITEST_EGL_PLATFORM "" CACHE STRING
-        "EGL_PLATFORM for the integration tests (empty: leave the loader alone)")
-endif()
+# No EGL_PLATFORM knob here on purpose. The harness pins EGL_PLATFORM=surfaceless
+# itself before its first EGL call (HeadlessGL.cpp, EnsureHeadlessPlatform) so a
+# developer's machine and a CI runner take the SAME path whether or not a window
+# system happens to be running. This used to inject "x11", which is how the lane
+# came up green on a workstation with WSLg and died on a runner with no X server.
+#
+# A build-system knob would not just be redundant, it would be a trap: `set(...
+# CACHE ...)` does not rewrite an existing cache, so every build directory
+# configured before this change would keep injecting EGL_PLATFORM=x11 and go on
+# binding to a window system - silently, and only on the machines that have one.
+# Someone reproducing a platform-specific bug sets EGL_PLATFORM in their own
+# environment, which the harness still honours.
 
 set(MGL_ITEST_COMMON_ENV "")
 if (MOBILEGL_ITEST_EGL_VENDOR)
     list(APPEND MGL_ITEST_COMMON_ENV "__EGL_VENDOR_LIBRARY_FILENAMES=${MOBILEGL_ITEST_EGL_VENDOR}")
 endif()
-if (MOBILEGL_ITEST_EGL_PLATFORM)
-    list(APPEND MGL_ITEST_COMMON_ENV "EGL_PLATFORM=${MOBILEGL_ITEST_EGL_PLATFORM}")
-endif()
+unset(MOBILEGL_ITEST_EGL_PLATFORM CACHE) # see above: an old cache must not resurrect x11
 if (MOBILEGL_ITEST_REQUIRE_GPU)
     list(APPEND MGL_ITEST_COMMON_ENV "MOBILEGL_ITEST_REQUIRE_GPU=1")
 endif()

MobileGL/MG_IntegrationTest/Harness/HeadlessGL.cpp

@@ -74,6 +74,40 @@ namespace MGITest {
             std::string renderer;
         };
 
+        // The harness is headless BY CONSTRUCTION, on every machine: it must never
+        // reach a window system, not even where one happens to be running. This is
+        // not a CI accommodation - it is what keeps a developer's run and a CI run
+        // the same run. The lane was wired up green on a workstation and immediately
+        // died on the runner precisely because the workstation had a DISPLAY (WSLg)
+        // and took Mesa's x11 platform, while the runner has none; that divergence
+        // is the bug, and pinning the platform here is the fix for it.
+        //
+        // Mesa selects its EGL platform from EGL_PLATFORM at loader time, so this
+        // has to run before the first EGL call in the process (see EnsureHeadless
+        // callers). surfaceless is the platform with no window-system dependency at
+        // all; the surface this file then creates is still a pbuffer, which every
+        // platform supports and which the amendment to this rule requires as the
+        // fallback shape. DISPLAY/WAYLAND_DISPLAY are cleared as well so that a
+        // driver that consults them directly cannot reintroduce the dependency
+        // behind EGL's back. Desktop-only file: MG_IntegrationTest never builds
+        // for Android, so no device path is affected.
+        void EnsureHeadlessPlatform() {
+#if defined(__linux__) && !defined(__ANDROID__)
+            static bool done = false;
+            if (done) {
+                return;
+            }
+            done = true;
+            // An explicit EGL_PLATFORM from the operator still wins: pinning a
+            // platform is exactly how someone reproduces a platform-specific bug.
+            if (std::getenv("EGL_PLATFORM") == nullptr) {
+                setenv("EGL_PLATFORM", "surfaceless", 1);
+            }
+            unsetenv("DISPLAY");
+            unsetenv("WAYLAND_DISPLAY");
+#endif
+        }
+
         // THE bring-up, in one function so the pre-flight child and the parent run
         // literally the same sequence - a pre-flight that tests something narrower
         // than what the parent will do is exactly the kind of "predictive" check
@@ -82,6 +116,9 @@ namespace MGITest {
         // Returns 0 on success, or the 1-based index of the step that failed, and
         // fills outReason either way.
         int RunEglBringUp(EglBringUp& out, std::string& outReason) {
+            // Belt and braces: the pre-flight child and the parent both enter here,
+            // and neither may be the first to touch EGL without this having run.
+            EnsureHeadlessPlatform();
             EGLDisplay display = eglGetDisplay(EGL_DEFAULT_DISPLAY);
             if (display == EGL_NO_DISPLAY) {
                 outReason = WithEglError("eglGetDisplay(EGL_DEFAULT_DISPLAY) returned EGL_NO_DISPLAY");
@@ -199,11 +236,10 @@ namespace MGITest {
             }
             if (child == 0) {
                 close(channel[0]);
-                // The child is EXPECTED to die on a signal on an unusable
-                // platform; that is the measurement. Do not let each such
-                // measurement drop a core file next to the test binary.
-                const rlimit noCore{0, 0};
-                setrlimit(RLIMIT_CORE, &noCore);
+                // No core suppression here, deliberately: when the child dies on a
+                // signal, the core IS the diagnosis (an rlimit that used to sit here
+                // made a CI-only crash undebuggable). Machines that do not want
+                // cores control that with the usual ulimit/core_pattern knobs.
                 std::fprintf(stderr, "[itest] pre-flight child: attempting a full EGL bring-up\n");
                 EglBringUp local;
                 std::string reason;
@@ -284,9 +320,19 @@ namespace MGITest {
         }
     } // namespace
 
+    namespace {
+        bool EnvFlag(const char* name) {
+            const char* value = std::getenv(name);
+            return value != nullptr && value[0] != '\0' && std::strcmp(value, "0") != 0;
+        }
+    } // namespace
+
     bool RequireGpu() {
-        const char* value = std::getenv("MOBILEGL_ITEST_REQUIRE_GPU");
-        return value != nullptr && value[0] != '\0' && std::strcmp(value, "0") != 0;
+        return EnvFlag("MOBILEGL_ITEST_REQUIRE_GPU");
+    }
+
+    bool RequireHardwareGpu() {
+        return EnvFlag("MOBILEGL_ITEST_REQUIRE_HARDWARE_GPU");
     }
 
     std::ostream& operator<<(std::ostream& os, const Rgba8& c) {
@@ -390,6 +436,10 @@ namespace MGITest {
     }
 
     HeadlessGL::HeadlessGL() {
+        // Before anything else in this process can reach EGL, and in particular
+        // before the pre-flight forks - the child must measure the same platform
+        // the parent will use.
+        EnsureHeadlessPlatform();
         m_backendName = EnvOr("MOBILEGL_BACKEND_TYPE", "<unset>");
         m_usable = BringUp();
     }
@@ -551,9 +601,13 @@ namespace MGITest {
     }
 
     Image ReadPixels(int width, int height) {
+        return ReadPixelsRect(0, 0, width, height);
+    }
+
+    Image ReadPixelsRect(int x, int y, int width, int height) {
         Image image(width, height);
         glPixelStorei(GL_PACK_ALIGNMENT, 1);
-        glReadPixels(0, 0, width, height, GL_RGBA, GL_UNSIGNED_BYTE, image.Data());
+        glReadPixels(x, y, width, height, GL_RGBA, GL_UNSIGNED_BYTE, image.Data());
         return image;
     }
 

MobileGL/MG_IntegrationTest/Harness/HeadlessGL.h

@@ -41,6 +41,15 @@ namespace MGITest {
     // a job that ran everything.
     bool RequireGpu();
 
+    // True when MOBILEGL_ITEST_REQUIRE_HARDWARE_GPU is set: additionally asserts
+    // that the context did NOT land on a software rasterizer. Deliberately a
+    // SEPARATE switch from RequireGpu - a GPU-less CI runner is a supported and
+    // intended configuration for these scenarios (they pin backend draw logic,
+    // which llvmpipe/lavapipe execute faithfully), so CI wants the falsifiability
+    // of REQUIRE_GPU without the hardware demand. Use this one only where a vendor
+    // pin silently degrading to software would invalidate the measurement.
+    bool RequireHardwareGpu();
+
     struct Rgba8 {
         std::uint8_t r = 0, g = 0, b = 0, a = 0;
 
@@ -175,11 +184,18 @@ namespace MGITest {
 
     void ClearTo(float r, float g, float b, float a);
 
-    // Reads back the whole currently bound READ framebuffer. width/height must
-    // be the target's full size - DirectVulkan's default-framebuffer readback
-    // only re-orients a full-extent read.
+    // Reads back the whole currently bound READ framebuffer.
     Image ReadPixels(int width, int height);
 
+    // A PARTIAL glReadPixels. Row 0 of the returned image is GL row `y` of the
+    // framebuffer, i.e. the bottom row of the requested rect - the same
+    // convention ReadPixels uses, just with an origin. This is the shape the
+    // conformance suite reads in (a random sub-rect of the default
+    // framebuffer), and the shape DirectVulkan's default-FBO readback used to
+    // hand back in Vulkan row order because its re-orientation only ran on an
+    // exact full-extent read.
+    Image ReadPixelsRect(int x, int y, int width, int height);
+
     // Drains any GL error queue and returns the first error, or 0.
     unsigned int FirstGLError();
     const char* GLErrorName(unsigned int error);

MobileGL/MG_IntegrationTest/Harness/ScenarioFixture.h

@@ -45,12 +45,18 @@ namespace MGITest {
                 }
                 GTEST_SKIP() << "no usable GPU/display/ICD for backend " << gl.BackendName() << ": " << gl.SkipReason();
             }
-            if (RequireGpu() && LooksLikeSoftwareRasterizer(gl.RendererString())) {
-                // "Ran on llvmpipe" must not be able to pass as "ran on the GPU":
-                // a misconfigured vendor pin silently lands on the software
-                // rasterizer, and REQUIRE_GPU exists precisely to make that loud.
-                FAIL() << "MOBILEGL_ITEST_REQUIRE_GPU is set but the context landed on a software rasterizer: "
-                       << gl.RendererString();
+            if (RequireHardwareGpu() && LooksLikeSoftwareRasterizer(gl.RendererString())) {
+                // Only when hardware was asked for BY NAME. REQUIRE_GPU means "an
+                // unusable harness is a failure, not a silent skip" - it is the
+                // falsifiability switch, and CI is exactly where it belongs. But CI
+                // runners have no GPU, so folding "must not be llvmpipe" into the
+                // same switch made the CI lane unpassable by construction: the
+                // scenarios pin backend draw logic, which a software rasterizer
+                // executes just as faithfully. Landing on llvmpipe/lavapipe there is
+                // the intended configuration, not a misconfiguration. A vendor pin
+                // that must not silently degrade sets REQUIRE_HARDWARE_GPU.
+                FAIL() << "MOBILEGL_ITEST_REQUIRE_HARDWARE_GPU is set but the context landed on a software "
+                       << "rasterizer: " << gl.RendererString();
             }
             // A scenario starts from a clean slate but shares the context (and so
             // the renderer's memos) with every other scenario in this process -

MobileGL/MG_IntegrationTest/Main.cpp

@@ -26,8 +26,14 @@ namespace {
             const MGITest::HeadlessGL& gl = MGITest::HeadlessGL::Get();
             std::fprintf(stderr, "MobileGL integration scenarios: backend=%s\n", gl.BackendName().c_str());
             if (gl.Usable()) {
-                std::fprintf(stderr, "  renderer: %s\n  surface:  %dx%d pbuffer (headless)\n",
-                             gl.RendererString().c_str(), gl.Width(), gl.Height());
+                // EGL_PLATFORM is echoed because it is the invariant this harness
+                // rests on: the run is headless on every machine, so a run that
+                // silently bound to a workstation's window system is a different
+                // run from CI's and must be visible as one in the log.
+                const char* eglPlatform = std::getenv("EGL_PLATFORM");
+                std::fprintf(stderr, "  renderer: %s\n  surface:  %dx%d pbuffer (headless, EGL_PLATFORM=%s)\n",
+                             gl.RendererString().c_str(), gl.Width(), gl.Height(),
+                             eglPlatform != nullptr ? eglPlatform : "<unset>");
             } else if (MGITest::RequireGpu()) {
                 std::fprintf(stderr,
                              "  FAILING every scenario (MOBILEGL_ITEST_REQUIRE_GPU is set): %s\n",

MobileGL/MG_IntegrationTest/Scenarios/AdvertisedLimitsScenario.cpp

@@ -0,0 +1,120 @@
+// MobileGL - MobileGL/MG_IntegrationTest/Scenarios/AdvertisedLimitsScenario.cpp
+// Copyright (c) 2025-2026 MobileGL-Dev
+// Licensed under the GNU Lesser General Public License v3.0:
+//   https://www.gnu.org/licenses/gpl-3.0.txt
+//   https://www.gnu.org/licenses/lgpl-3.0.txt
+// SPDX-License-Identifier: LGPL-3.0-only
+// End of Source File Header
+//
+// "The limit we advertise is a promise, and an application will hold us to it."
+//
+// DirectVulkan copied Vulkan descriptor limits straight into the GL limit table. Those are not
+// the same quantity: Adreno answers maxPerStageDescriptorUniformBuffers at descriptor-indexing
+// scale, and GL_MAX_COMPUTE_UNIFORM_BLOCKS is a count an app will allocate. KHR-GL44.multi_bind
+// .dispatch_bind_buffers_base does exactly that - createsO(limit) buffers and splices O(limit)
+// UBO declarations into one compute shader - and spent ~14 s allocating before dying on
+// std::bad_alloc. Its sibling dispatch_bind_buffers_range hard-codes 4 buffers and passes.
+//
+// Two failure modes, one table:
+//   - too LARGE: an unusable promise (the OOM above).
+//   - too SMALL or negative: a uint32 limit that lost its top bit on the way to a signed Int -
+//     UINT32_MAX arrived as -1, which every downstream std::min then accepted as "small enough".
+//     A conformant GL 4.x implementation may never advertise below the spec minimum either.
+//
+// Every bound below is checked on BOTH backends, because the loader casts are shared and the
+// DirectGLES lane is the control: it takes its limits from a driver that already reports GL
+// quantities, so an entry that only fails on DirectVulkan is a translation bug and one that
+// fails on both is a table bug.
+
+#include <string>
+#include <vector>
+
+#include "../Harness/HeadlessGL.h"
+#include "../Harness/ScenarioFixture.h"
+
+#ifdef GLAPI
+#undef GLAPI
+#endif
+#define GL_GLEXT_PROTOTYPES
+#include <GL/gl.h>
+#include <GL/glcorearb.h>
+#undef GL_GLEXT_PROTOTYPES
+
+namespace MGITest {
+    namespace {
+
+        struct LimitBound {
+            GLenum pname;
+            const char* name;
+            // The GL 4.x required minimum. A value below this is a conformance failure in its own
+            // right, and is what a sign-flipped uint32 looks like.
+            int minimum;
+            // The largest value this implementation is willing to promise. Chosen well above every
+            // desktop driver's answer, so it can only catch a descriptor-scale number.
+            int ceiling;
+        };
+
+        const std::vector<LimitBound>& BufferLimitTable() {
+            static const std::vector<LimitBound> table = {
+                {GL_MAX_UNIFORM_BUFFER_BINDINGS, "GL_MAX_UNIFORM_BUFFER_BINDINGS", 36, 256},
+                {GL_MAX_COMPUTE_UNIFORM_BLOCKS, "GL_MAX_COMPUTE_UNIFORM_BLOCKS", 12, 256},
+                {GL_MAX_COMPUTE_SHADER_STORAGE_BLOCKS, "GL_MAX_COMPUTE_SHADER_STORAGE_BLOCKS", 8, 256},
+                {GL_MAX_COMBINED_SHADER_STORAGE_BLOCKS, "GL_MAX_COMBINED_SHADER_STORAGE_BLOCKS", 8, 256},
+                {GL_MAX_SHADER_STORAGE_BUFFER_BINDINGS, "GL_MAX_SHADER_STORAGE_BUFFER_BINDINGS", 8, 256},
+                {GL_MAX_TEXTURE_BUFFER_SIZE, "GL_MAX_TEXTURE_BUFFER_SIZE", 65536, 1 << 27},
+                {GL_MAX_UNIFORM_BLOCK_SIZE, "GL_MAX_UNIFORM_BLOCK_SIZE", 16384, 1 << 30},
+                // Already clamped before this campaign; in the table so a regression there is
+                // caught by the same case.
+                {GL_MAX_SHADER_STORAGE_BLOCK_SIZE, "GL_MAX_SHADER_STORAGE_BLOCK_SIZE", 1 << 24, 512 * 1024 * 1024},
+                {GL_MAX_TEXTURE_IMAGE_UNITS, "GL_MAX_TEXTURE_IMAGE_UNITS", 16, 32},
+                {GL_MAX_COMBINED_TEXTURE_IMAGE_UNITS, "GL_MAX_COMBINED_TEXTURE_IMAGE_UNITS", 48, 192},
+            };
+            return table;
+        }
+
+        class AdvertisedLimitsScenario : public ScenarioTest {};
+
+        TEST_F(AdvertisedLimitsScenario, EveryBufferLimitIsWithinItsAdvertisedRange) {
+            for (const LimitBound& bound : BufferLimitTable()) {
+                GLint value = -424242;
+                glGetIntegerv(bound.pname, &value);
+                const unsigned int error = FirstGLError();
+                EXPECT_EQ(error, GLenum(GL_NO_ERROR))
+                    << bound.name << " is not answerable: " << GLErrorName(error);
+                if (error != GL_NO_ERROR) continue;
+
+                EXPECT_GE(value, bound.minimum)
+                    << bound.name << " = " << value << " is below the GL required minimum "
+                    << bound.minimum << " (a negative or tiny value here is a uint32 limit that lost "
+                                        "its top bit on the way to a signed Int)";
+                EXPECT_LE(value, bound.ceiling)
+                    << bound.name << " = " << value << " exceeds the ceiling " << bound.ceiling
+                    << " this implementation is willing to promise - an application that allocates "
+                       "what we advertise will run out of memory";
+            }
+        }
+
+        // The OOM case in isolation, because it is the one with a known CTS victim and the one a
+        // future refactor is most likely to reintroduce by copying the Vulkan limit back.
+        TEST_F(AdvertisedLimitsScenario, ComputeUniformBlocksIsAnAmountAnApplicationCouldActuallyAllocate) {
+            GLint blocks = -1;
+            glGetIntegerv(GL_MAX_COMPUTE_UNIFORM_BLOCKS, &blocks);
+            ASSERT_EQ(FirstGLError(), GLenum(GL_NO_ERROR));
+            EXPECT_GE(blocks, 12);
+            EXPECT_LE(blocks, 256) << "KHR-GL44.multi_bind.dispatch_bind_buffers_base creates one GL buffer "
+                                      "and one UBO declaration per advertised block";
+
+            GLint blockSize = -1;
+            glGetIntegerv(GL_MAX_UNIFORM_BLOCK_SIZE, &blockSize);
+            ASSERT_EQ(FirstGLError(), GLenum(GL_NO_ERROR));
+            EXPECT_GT(blockSize, 0);
+            // GL_MAX_COMBINED_COMPUTE_UNIFORM_COMPONENTS is derived from the product of these two,
+            // so their product has to stay representable.
+            EXPECT_LE(static_cast<long long>(blocks) * blockSize,
+                      static_cast<long long>(2147483647))
+                << "blocks(" << blocks << ") * blockSize(" << blockSize << ") overflows the GLint the "
+                   "derived component limits are computed in";
+        }
+
+    } // namespace
+} // namespace MGITest

MobileGL/MG_IntegrationTest/Scenarios/AsyncCompileScenario.cpp

@@ -157,6 +157,22 @@ void main() {
             const QuirkOverride m_saved;
         };
 
+        // MOBILEGL_ASYNC_OPTIMISTIC_SHADER_STATUS, forced in-process for the same reason
+        // as AsyncModeScope: one ctest run asserts the quirk against the ambient default.
+        class OptimisticStatusScope {
+        public:
+            explicit OptimisticStatusScope(const QuirkOverride mode)
+                : m_saved(MobileGL::MG_Config::Features.AsyncOptimisticShaderStatus) {
+                MobileGL::MG_Config::Features.AsyncOptimisticShaderStatus = mode;
+            }
+            ~OptimisticStatusScope() { MobileGL::MG_Config::Features.AsyncOptimisticShaderStatus = m_saved; }
+            OptimisticStatusScope(const OptimisticStatusScope&) = delete;
+            OptimisticStatusScope& operator=(const OptimisticStatusScope&) = delete;
+
+        private:
+            const QuirkOverride m_saved;
+        };
+
         // glMaxShaderCompilerThreadsKHR writes process-wide state; a scenario that calls
         // it has to put the pool back or it changes how every scenario after it compiles.
         class CompilerThreadScope {
@@ -463,5 +479,81 @@ void main() {
             EXPECT_EQ(FirstGLError(), GLenum(GL_NO_ERROR));
         }
 
+        // The Iris two-phase shape end to end on a real driver, with the optimistic-status
+        // quirk on: phase 1 compiles each stage and reads its log then its status (both
+        // answered optimistically), links, detaches and deletes the shaders for every
+        // program with no program-level read anywhere; phase 2 then checks every link and
+        // draws every program. Deliberately NOT built on the harness CompileProgram(),
+        // whose status read would join and collapse the phase-1 overlap this exists to
+        // exercise. What the unit suite cannot see - worker-produced artifacts the backend
+        // then mis-renders - shows up here as a wrong quadrant signature.
+        TEST_F(AsyncCompileScenario, IrisShapedTwoPhaseBatchRendersCorrectly) {
+            if (!Ready()) return;
+            constexpr int kPrograms = 12;
+
+            // Distinct per program (so neither the source memo nor the adoption map turns
+            // a compile into a no-op) but a pure pass-through at runtime: the bulk sits in
+            // a branch a zero-initialised uniform never takes.
+            const auto fragmentSource = [](const int index) {
+                std::string source = "#version 330 core\nin vec3 vColor;\nout vec4 oColor;\n";
+                source += "uniform float uGate" + std::to_string(index) + ";\n";
+                source += "void main() {\n    oColor = vec4(vColor, 1.0);\n";
+                source += "    if (uGate" + std::to_string(index) + " > 1e30) {\n        float acc = 1.0;\n";
+                for (int i = 0; i < 60; ++i) {
+                    source += "        acc = acc * 1.0001 + sin(acc + " + std::to_string(i) + ".0);\n";
+                }
+                source += "        oColor = vec4(acc);\n    }\n}\n";
+                return source;
+            };
+
+            std::vector<GLuint> programs;
+            {
+                const AsyncModeScope async(true);
+                const OptimisticStatusScope quirk(QuirkOverride::ForceOn);
+                const CompilerThreadScope threads;
+                glMaxShaderCompilerThreadsKHR(1);
+
+                for (int i = 0; i < kPrograms; ++i) {
+                    m_sources.push_back(fragmentSource(i));
+                    const char* fsText = m_sources.back().c_str();
+
+                    const GLuint vs = glCreateShader(GL_VERTEX_SHADER);
+                    glShaderSource(vs, 1, &kVertexSource, nullptr);
+                    glCompileShader(vs);
+                    (void)ShaderInfoLog(vs);       // Iris's exact order: the log first...
+                    (void)ShaderCompileStatus(vs); // ...then the status; both optimistic.
+
+                    const GLuint fs = glCreateShader(GL_FRAGMENT_SHADER);
+                    glShaderSource(fs, 1, &fsText, nullptr);
+                    glCompileShader(fs);
+                    (void)ShaderInfoLog(fs);
+                    (void)ShaderCompileStatus(fs);
+
+                    const GLuint program = glCreateProgram();
+                    glAttachShader(program, vs);
+                    glAttachShader(program, fs);
+                    glBindAttribLocation(program, 0, "aPos");
+                    glBindAttribLocation(program, 1, "aColor");
+                    glLinkProgram(program);
+                    glDetachShader(program, vs);
+                    glDetachShader(program, fs);
+                    glDeleteShader(vs);
+                    glDeleteShader(fs);
+                    programs.push_back(program);
+                }
+            }
+
+            for (int i = 0; i < kPrograms; ++i) {
+                const GLuint program = programs[static_cast<std::size_t>(i)];
+                GLint linked = GL_FALSE;
+                glGetProgramiv(program, GL_LINK_STATUS, &linked);
+                ASSERT_EQ(linked, GL_TRUE) << "program " << i;
+                const Image image = DrawFrameWith(program);
+                EXPECT_EQ(image.QuadrantSignature(), "blue,green,red,white") << "program " << i;
+            }
+            for (const GLuint program : programs) glDeleteProgram(program);
+            EXPECT_EQ(FirstGLError(), GLenum(GL_NO_ERROR));
+        }
+
     } // namespace
 } // namespace MGITest

MobileGL/MG_IntegrationTest/Scenarios/BufferTextureScenario.cpp

@@ -0,0 +1,174 @@
+// MobileGL - MobileGL/MG_IntegrationTest/Scenarios/BufferTextureScenario.cpp
+// Copyright (c) 2025-2026 MobileGL-Dev
+// Licensed under the GNU Lesser General Public License v3.0:
+//   https://www.gnu.org/licenses/gpl-3.0.txt
+//   https://www.gnu.org/licenses/lgpl-3.0.txt
+// SPDX-License-Identifier: LGPL-3.0-only
+// End of Source File Header
+//
+// Scenario - A BUFFER TEXTURE IS SAMPLED FROM THE VERTEX STAGE, AND TRACKS ITS BUFFER.
+//
+// Buffer textures are core in OpenGL 3.1 and MobileGL advertises a 4.x context, so an
+// application may build geometry out of one without asking whether the host can. Minecraft
+// 26.3 does exactly that: its cloud layer has no vertex attributes at all, only gl_VertexID
+// and texelFetch on a GL_R8I buffer texture. Nothing covered that path end to end on either
+// backend - the frontend unit tests stop at glTexBuffer's state, and no scenario ever drew
+// with the result - which is how DirectGLES came to emit `#extension GL_EXT_texture_buffer :
+// require` unconditionally, compile nothing on a host without the extension, and lose the
+// whole cloud layer with no diagnostic anywhere.
+//
+// Two claims, in the order they can break:
+//   1. a vertex-stage texelFetch on an R8I buffer texture reads the byte the application put
+//      in the buffer (the shape of the real workload: no attributes, index from gl_VertexID);
+//   2. a later glBufferSubData is visible to the next draw WITHOUT re-specifying the texture.
+//      glTexBuffer attaches storage, it does not copy: the texture is a live view of the
+//      buffer, so a backend that only refreshes the view when the texture's own state changes
+//      must still show the new bytes. DirectGLES' respecify gate is keyed on the texture info
+//      and deliberately does not include the buffer's contents, so this is the assertion that
+//      says that is safe rather than merely untested.
+//
+// NOTE ON A HOST WITHOUT BUFFER TEXTURES: this scenario is expected to FAIL there, and that is
+// the honest outcome - MobileGL keeps advertising GL_MAX_TEXTURE_BUFFER_SIZE (an OpenGL 4.x
+// context may not answer 0), so there is no capability an application, or this test, could
+// branch on. The driver POST's "Buffer textures" row is where that verdict is stated.
+
+#include <cstdint>
+#include <string>
+#include <vector>
+
+#include "../Harness/HeadlessGL.h"
+#include "../Harness/ScenarioFixture.h"
+
+#ifdef GLAPI
+#undef GLAPI
+#endif
+#define GL_GLEXT_PROTOTYPES
+#include <GL/gl.h>
+#include <GL/glcorearb.h>
+#undef GL_GLEXT_PROTOTYPES
+
+namespace MGITest {
+    namespace {
+
+        // No vertex attributes: the quad's corners come from gl_VertexID, exactly like the
+        // workload this exists for. The texel is fetched in the VERTEX stage - the stage where
+        // buffer-texture support is scarcest across ES drivers - and carried flat so every
+        // fragment of the quad reports the same byte and the readback is exact.
+        constexpr const char* kVS = R"(#version 330 core
+uniform isamplerBuffer uFaces;
+flat out int vFace;
+void main() {
+    vec2 corner = vec2((gl_VertexID & 1) == 0 ? -1.0 : 1.0,
+                       (gl_VertexID & 2) == 0 ? -1.0 : 1.0);
+    vFace = texelFetch(uFaces, 0).r;
+    gl_Position = vec4(corner, 0.0, 1.0);
+}
+)";
+
+        // 1/255 steps survive an RGBA8 round trip exactly, so the readback byte IS the value
+        // the vertex shader fetched.
+        constexpr const char* kFS = R"(#version 330 core
+flat in int vFace;
+out vec4 o_color;
+void main() { o_color = vec4(float(vFace) / 255.0, 0.0, 0.0, 1.0); }
+)";
+
+        class BufferTextureScenario : public ScenarioTest {};
+
+        // Draws the full-viewport quad and returns the red byte every fragment was painted with,
+        // or -1 if the quad did not come out uniform (which would mean the flat varying, not the
+        // fetch, is what this test is measuring).
+        int PaintedValue(unsigned int program, int width, int height) {
+            ClearTo(0.0f, 0.0f, 0.0f, 1.0f);
+            GLuint vao = 0;
+            glGenVertexArrays(1, &vao);
+            glBindVertexArray(vao);
+            glUseProgram(program);
+            glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
+            glBindVertexArray(0);
+            glDeleteVertexArrays(1, &vao);
+
+            const Image image = ReadPixels(width, height);
+            if (image.Empty()) {
+                return -1;
+            }
+            const int first = image.At(0, 0).r;
+            for (int y = 0; y < image.Height(); ++y) {
+                for (int x = 0; x < image.Width(); ++x) {
+                    if (image.At(x, y).r != first) {
+                        return -1;
+                    }
+                }
+            }
+            return first;
+        }
+
+    } // namespace
+
+    TEST_F(BufferTextureScenario, VertexStageTexelFetchReadsTheBufferAndTracksItsUpdates) {
+        if (!Ready()) return;
+        HeadlessGL& gl = Gl();
+
+        std::string error;
+        const unsigned int program = CompileProgram(kVS, kFS, &error);
+        ASSERT_NE(program, 0u) << error;
+
+        // GL_R8I is the format the real workload uses. Signed, so the values stay well inside
+        // [0, 127] to keep the readback arithmetic honest.
+        constexpr signed char kInitial = 37;
+        constexpr signed char kUpdated = 91;
+        std::vector<signed char> texels(64, 0);
+        texels[0] = kInitial;
+
+        // The harness shares one context across every scenario in the process, so an error left
+        // by an earlier one would surface below as "glTexBuffer was refused".
+        FirstGLError();
+
+        GLuint buffer = 0;
+        glGenBuffers(1, &buffer);
+        glBindBuffer(GL_TEXTURE_BUFFER, buffer);
+        glBufferData(GL_TEXTURE_BUFFER, static_cast<GLsizeiptr>(texels.size()), texels.data(),
+                     GL_DYNAMIC_DRAW);
+
+        GLuint texture = 0;
+        glGenTextures(1, &texture);
+        glBindTexture(GL_TEXTURE_BUFFER, texture);
+        glTexBuffer(GL_TEXTURE_BUFFER, GL_R8I, buffer);
+        ASSERT_EQ(FirstGLError(), 0u) << "glTexBuffer(GL_R8I) was refused";
+
+        ColorFbo target = MakeColorFbo(64, 64);
+        ASSERT_NE(target.fbo, 0u) << "could not create the render target";
+        BindFbo(target);
+
+        glActiveTexture(GL_TEXTURE0);
+        glBindTexture(GL_TEXTURE_BUFFER, texture);
+        glUseProgram(program);
+        const GLint location = glGetUniformLocation(program, "uFaces");
+        ASSERT_NE(location, -1) << "the buffer sampler was optimized away or never reflected";
+        glUniform1i(location, 0);
+
+        EXPECT_EQ(PaintedValue(program, target.width, target.height), static_cast<int>(kInitial))
+            << "a vertex-stage texelFetch on an R8I buffer texture did not read the byte the "
+               "application stored (a uniform -1 here means the quad was not uniform at all)";
+
+        // The texture is a VIEW of the buffer: no glTexBuffer call follows, and none should be
+        // needed for the new bytes to be visible.
+        glBindBuffer(GL_TEXTURE_BUFFER, buffer);
+        glBufferSubData(GL_TEXTURE_BUFFER, 0, 1, &kUpdated);
+        ASSERT_EQ(FirstGLError(), 0u) << "glBufferSubData on the texture's buffer was refused";
+
+        EXPECT_EQ(PaintedValue(program, target.width, target.height), static_cast<int>(kUpdated))
+            << "the buffer texture kept showing the old contents after glBufferSubData; the "
+               "texture must track its buffer without being re-specified";
+
+        BindDefaultFramebuffer();
+        DestroyColorFbo(target);
+        glUseProgram(0);
+        glDeleteProgram(program);
+        glDeleteTextures(1, &texture);
+        glDeleteBuffers(1, &buffer);
+        glViewport(0, 0, gl.Width(), gl.Height());
+        EXPECT_EQ(FirstGLError(), 0u);
+    }
+
+} // namespace MGITest

MobileGL/MG_IntegrationTest/Scenarios/ClearThenReadPixelsScenario.cpp

@@ -0,0 +1,335 @@
+// MobileGL - MobileGL/MG_IntegrationTest/Scenarios/ClearThenReadPixelsScenario.cpp
+// Copyright (c) 2025-2026 MobileGL-Dev
+// Licensed under the GNU Lesser General Public License v3.0:
+//   https://www.gnu.org/licenses/gpl-3.0.txt
+//   https://www.gnu.org/licenses/lgpl-3.0.txt
+// SPDX-License-Identifier: LGPL-3.0-only
+// End of Source File Header
+//
+// Scenario - A CLEAR OF THE DEFAULT FRAMEBUFFER IS VISIBLE TO glReadPixels WITH NO DRAW BETWEEN.
+//
+// DirectVulkan parks a glClear as a pending clear and folds it into the next render pass's
+// loadOp. When nothing is drawn after the clear there is no render pass, and the readback path
+// used to materialize pending clears only for USER framebuffers - so a readback right after a
+// clear of the DEFAULT framebuffer blitted the untouched swapchain image and handed back the
+// previous frame's colour.
+//
+// That is the whole of KHR-GL40.draw_indirect.negative-* (12 Magma failures): each case clears,
+// issues a draw that correctly raises INVALID_OPERATION and therefore never executes, then reads
+// the frame back expecting (0,0,0,0) and gets the previous case's (0.1,0.2,0.3,1). The staleness
+// cannot appear in one frame, so the scenario paints a frame first and clears in the next.
+//
+// The alpha assertion is the second half of the same census finding: a cleared default
+// framebuffer read back (0,0,0,1) where (0,0,0,0) was written, because the clear was routed
+// through the default FBO's placeholder attachment, whose format can lack alpha, rather than
+// through the swapchain image that actually has one.
+//
+// DirectGLES is the built-in control: a native GL driver has no deferred-clear model at all, so
+// a failure there would mean the scenario, not the backend.
+
+#include <string>
+#include <vector>
+
+#include "../Harness/HeadlessGL.h"
+#include "../Harness/ScenarioFixture.h"
+
+#ifdef GLAPI
+#undef GLAPI
+#endif
+#define GL_GLEXT_PROTOTYPES
+#include <GL/gl.h>
+#include <GL/glcorearb.h>
+#undef GL_GLEXT_PROTOTYPES
+
+namespace MGITest {
+    namespace {
+
+        constexpr const char* kVS = R"(#version 330 core
+in vec2 aPos;
+void main() { gl_Position = vec4(aPos, 0.0, 1.0); }
+)";
+
+        // The colour KHR-GL40.draw_indirect's fshSimple paints, so a stale readback shows up as
+        // the same value the conformance log reports.
+        constexpr const char* kFS = R"(#version 330 core
+out vec4 o_color;
+void main() { o_color = vec4(0.1, 0.2, 0.3, 1.0); }
+)";
+
+        class ClearThenReadPixelsScenario : public ScenarioTest {};
+
+        void DrawFullViewportQuad(unsigned int program) {
+            static const float kQuad[] = {-1.0f, -1.0f, 1.0f, -1.0f, -1.0f, 1.0f, 1.0f, 1.0f};
+            GLuint vao = 0, vbo = 0;
+            glGenVertexArrays(1, &vao);
+            glBindVertexArray(vao);
+            glGenBuffers(1, &vbo);
+            glBindBuffer(GL_ARRAY_BUFFER, vbo);
+            glBufferData(GL_ARRAY_BUFFER, sizeof(kQuad), kQuad, GL_STATIC_DRAW);
+            glEnableVertexAttribArray(0);
+            glVertexAttribPointer(0, 2, GL_FLOAT, GL_FALSE, 2 * sizeof(float), nullptr);
+            glUseProgram(program);
+            glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
+            glBindVertexArray(0);
+            glDeleteBuffers(1, &vbo);
+            glDeleteVertexArrays(1, &vao);
+        }
+
+    } // namespace
+
+    TEST_F(ClearThenReadPixelsScenario, ClearWithNoDrawIsVisibleToDefaultFramebufferReadPixels) {
+        if (!Ready()) return;
+        HeadlessGL& gl = Gl();
+        const int width = gl.Width();
+        const int height = gl.Height();
+        ASSERT_GE(width, 8);
+        ASSERT_GE(height, 8);
+
+        std::string error;
+        const unsigned int program = CompileProgram(kVS, kFS, &error);
+        ASSERT_NE(program, 0u) << error;
+
+        // Frame 1: paint the whole default framebuffer, so there IS something stale to return.
+        BindDefaultFramebuffer();
+        glViewport(0, 0, width, height);
+        glDisable(GL_SCISSOR_TEST);
+        glDisable(GL_DEPTH_TEST);
+        ClearTo(1.0f, 1.0f, 1.0f, 1.0f);
+        DrawFullViewportQuad(program);
+        {
+            const Image painted = ReadPixels(width, height);
+            const Rgba8 centre = painted.At(width / 2, height / 2);
+            ASSERT_NEAR(centre.r, 26, 2) << "the setup frame did not paint; the staleness test would be vacuous";
+            ASSERT_NEAR(centre.g, 51, 2);
+            ASSERT_NEAR(centre.b, 77, 2);
+        }
+        gl.EndFrame();
+
+        // Frame 2: clear to transparent black and read back with NO draw at all.
+        BindDefaultFramebuffer();
+        glViewport(0, 0, width, height);
+        ClearTo(0.0f, 0.0f, 0.0f, 0.0f);
+        const Image cleared = ReadPixels(width, height);
+        EXPECT_EQ(FirstGLError(), 0u);
+
+        int nonZero = 0;
+        int firstX = -1;
+        int firstY = -1;
+        Rgba8 firstOffender{};
+        for (int y = 0; y < height; ++y) {
+            for (int x = 0; x < width; ++x) {
+                const Rgba8 pixel = cleared.At(x, y);
+                if (pixel.r == 0 && pixel.g == 0 && pixel.b == 0 && pixel.a == 0) continue;
+                if (nonZero == 0) {
+                    firstX = x;
+                    firstY = y;
+                    firstOffender = pixel;
+                }
+                ++nonZero;
+            }
+        }
+        EXPECT_EQ(nonZero, 0) << "glClear(0,0,0,0) followed by glReadPixels with no draw returned " << nonZero
+                              << " of " << (width * height) << " non-zero pixels; first at (" << firstX << ", "
+                              << firstY << ") = (" << static_cast<int>(firstOffender.r) << ", "
+                              << static_cast<int>(firstOffender.g) << ", " << static_cast<int>(firstOffender.b)
+                              << ", " << static_cast<int>(firstOffender.a) << ")";
+
+        gl.EndFrame();
+        glDeleteProgram(program);
+    }
+
+    // The same claim for a sub-rect read, which is the shape the conformance suite uses most and
+    // the one whose orientation handling is separate (see OrientationScenario).
+    TEST_F(ClearThenReadPixelsScenario, ClearWithNoDrawIsVisibleToASubRectReadback) {
+        if (!Ready()) return;
+        HeadlessGL& gl = Gl();
+        const int width = gl.Width();
+        const int height = gl.Height();
+        ASSERT_GE(width, 8);
+        ASSERT_GE(height, 8);
+
+        std::string error;
+        const unsigned int program = CompileProgram(kVS, kFS, &error);
+        ASSERT_NE(program, 0u) << error;
+
+        BindDefaultFramebuffer();
+        glViewport(0, 0, width, height);
+        glDisable(GL_SCISSOR_TEST);
+        glDisable(GL_DEPTH_TEST);
+        DrawFullViewportQuad(program);
+        gl.EndFrame();
+
+        BindDefaultFramebuffer();
+        glViewport(0, 0, width, height);
+        ClearTo(0.0f, 0.0f, 0.0f, 0.0f);
+        const int rectWidth = width / 2;
+        const int rectHeight = height / 2;
+        const Image cleared = ReadPixelsRect(width / 4, height / 4, rectWidth, rectHeight);
+        EXPECT_EQ(FirstGLError(), 0u);
+
+        int nonZero = 0;
+        for (int y = 0; y < rectHeight; ++y) {
+            for (int x = 0; x < rectWidth; ++x) {
+                const Rgba8 pixel = cleared.At(x, y);
+                if (pixel.r != 0 || pixel.g != 0 || pixel.b != 0 || pixel.a != 0) ++nonZero;
+            }
+        }
+        EXPECT_EQ(nonZero, 0) << nonZero << " of " << (rectWidth * rectHeight)
+                              << " pixels in a sub-rect read after a draw-free clear were not zero";
+
+        gl.EndFrame();
+        glDeleteProgram(program);
+    }
+
+    // The other half of the same rule, and the one the first version of this fix got wrong: a
+    // parked clear must be executed BEFORE whatever writes the framebuffer next, not whenever the
+    // readback happens to notice it. Minecraft clears the default framebuffer, renders the world
+    // into its own framebuffer and blits the result out; nothing in between opens a render pass on
+    // the default framebuffer, so the clear stays parked across the whole frame. Materializing it
+    // at readback time therefore ran it AFTER the blit and returned a blank frame - which is what
+    // took every DirectVulkan retrace to ssim 0.000005.
+    TEST_F(ClearThenReadPixelsScenario, ABlitIntoTheDefaultFramebufferSurvivesAnEarlierClear) {
+        if (!Ready()) return;
+        HeadlessGL& gl = Gl();
+        const int width = gl.Width();
+        const int height = gl.Height();
+
+        std::string error;
+        const unsigned int program = CompileProgram(kVS, kFS, &error);
+        ASSERT_NE(program, 0u) << error;
+
+        // Paint a source framebuffer, exactly as a game renders its world off-screen.
+        ColorFbo source = MakeColorFbo(width, height);
+        ASSERT_NE(source.fbo, 0u);
+        BindFbo(source);
+        glDisable(GL_SCISSOR_TEST);
+        glDisable(GL_DEPTH_TEST);
+        ClearTo(0.0f, 0.0f, 0.0f, 1.0f);
+        DrawFullViewportQuad(program);
+
+        // Clear the DEFAULT framebuffer, then blit the source over it. The clear is white so a
+        // frame that lost the blit is unmistakable, and the blit's colour is fshSimple's.
+        BindDefaultFramebuffer();
+        glViewport(0, 0, width, height);
+        ClearTo(1.0f, 1.0f, 1.0f, 1.0f);
+        glBindFramebuffer(GL_READ_FRAMEBUFFER, source.fbo);
+        glBindFramebuffer(GL_DRAW_FRAMEBUFFER, 0);
+        glBlitFramebuffer(0, 0, width, height, 0, 0, width, height, GL_COLOR_BUFFER_BIT, GL_NEAREST);
+        glBindFramebuffer(GL_FRAMEBUFFER, 0);
+        EXPECT_EQ(FirstGLError(), 0u);
+
+        const Image blitted = ReadPixels(width, height);
+        EXPECT_EQ(FirstGLError(), 0u);
+        const Rgba8 centre = blitted.At(width / 2, height / 2);
+        EXPECT_NEAR(centre.r, 26, 2) << "the blit into the default framebuffer did not survive the clear that "
+                                        "preceded it; read back rgba(" << static_cast<int>(centre.r) << ", "
+                                     << static_cast<int>(centre.g) << ", " << static_cast<int>(centre.b) << ", "
+                                     << static_cast<int>(centre.a) << ")";
+        EXPECT_NEAR(centre.g, 51, 2);
+        EXPECT_NEAR(centre.b, 77, 2);
+
+        DestroyColorFbo(source);
+        gl.EndFrame();
+        glDeleteProgram(program);
+    }
+
+    // A MULTISAMPLE-RESOLVE blit into the default framebuffer has to change orientation like any
+    // other, but vkCmdResolveImage takes one offset per side and cannot invert an axis, so it used
+    // to land the mirrored band. The renderer now resolves into a single-sample scratch image and
+    // blits from there. The source is painted in two horizontal bands so the mirror is visible;
+    // a full-extent uniform blit is a fixed point of the flip and would prove nothing.
+    TEST_F(ClearThenReadPixelsScenario, AMultisampleResolveBlitIntoTheDefaultFramebufferKeepsItsOrientation) {
+        if (!Ready()) return;
+        HeadlessGL& gl = Gl();
+        const int width = gl.Width();
+        const int height = gl.Height();
+        ASSERT_GE(height, 8);
+
+        GLint maxSamples = 0;
+        glGetIntegerv(GL_MAX_SAMPLES, &maxSamples);
+        if (maxSamples < 2) {
+            GTEST_SKIP() << "GL_MAX_SAMPLES is " << maxSamples << "; this needs a multisample renderbuffer";
+        }
+
+        GLuint fbo = 0, rbo = 0;
+        glGenFramebuffers(1, &fbo);
+        glBindFramebuffer(GL_FRAMEBUFFER, fbo);
+        glGenRenderbuffers(1, &rbo);
+        glBindRenderbuffer(GL_RENDERBUFFER, rbo);
+        glRenderbufferStorageMultisample(GL_RENDERBUFFER, 2, GL_RGBA8, width, height);
+        glFramebufferRenderbuffer(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_RENDERBUFFER, rbo);
+        if (glCheckFramebufferStatus(GL_FRAMEBUFFER) != GL_FRAMEBUFFER_COMPLETE) {
+            glDeleteRenderbuffers(1, &rbo);
+            glDeleteFramebuffers(1, &fbo);
+            glBindFramebuffer(GL_FRAMEBUFFER, 0);
+            GTEST_SKIP() << "no complete 2x multisample RGBA8 renderbuffer on this driver";
+        }
+        glViewport(0, 0, width, height);
+
+        // Bottom half red, top half blue - via scissored clears, so no shader is involved.
+        glEnable(GL_SCISSOR_TEST);
+        glScissor(0, 0, width, height / 2);
+        ClearTo(1.0f, 0.0f, 0.0f, 1.0f);
+        glScissor(0, height / 2, width, height - height / 2);
+        ClearTo(0.0f, 0.0f, 1.0f, 1.0f);
+        glDisable(GL_SCISSOR_TEST);
+
+        BindDefaultFramebuffer();
+        glViewport(0, 0, width, height);
+        ClearTo(0.0f, 0.0f, 0.0f, 1.0f);
+        glBindFramebuffer(GL_READ_FRAMEBUFFER, fbo);
+        glBindFramebuffer(GL_DRAW_FRAMEBUFFER, 0);
+        glBlitFramebuffer(0, 0, width, height, 0, 0, width, height, GL_COLOR_BUFFER_BIT, GL_NEAREST);
+        glBindFramebuffer(GL_FRAMEBUFFER, 0);
+        EXPECT_EQ(FirstGLError(), 0u);
+
+        const Image resolved = ReadPixels(width, height);
+        EXPECT_EQ(FirstGLError(), 0u);
+        const Rgba8 bottom = resolved.At(width / 2, height / 4);
+        const Rgba8 top = resolved.At(width / 2, height - 1 - height / 4);
+        EXPECT_GT(bottom.r, 200) << "the bottom band should be red after the resolve, got rgba("
+                                 << static_cast<int>(bottom.r) << ", " << static_cast<int>(bottom.g) << ", "
+                                 << static_cast<int>(bottom.b) << ") - blue there means the resolve landed "
+                                 << "in the mirrored band";
+        EXPECT_LT(bottom.b, 60);
+        EXPECT_GT(top.b, 200) << "the top band should be blue after the resolve, got rgba("
+                              << static_cast<int>(top.r) << ", " << static_cast<int>(top.g) << ", "
+                              << static_cast<int>(top.b) << ")";
+        EXPECT_LT(top.r, 60);
+
+        glDeleteRenderbuffers(1, &rbo);
+        glDeleteFramebuffers(1, &fbo);
+        gl.EndFrame();
+    }
+
+    // The same ordering claim for the path that DOES open a render pass. It passes today (the
+    // render pass folds the clear into its loadOp and pops it), and it is here so a future change
+    // to the pending-clear lifecycle cannot quietly reverse clear and draw.
+    TEST_F(ClearThenReadPixelsScenario, ADrawIntoTheDefaultFramebufferSurvivesAnEarlierClear) {
+        if (!Ready()) return;
+        HeadlessGL& gl = Gl();
+        const int width = gl.Width();
+        const int height = gl.Height();
+
+        std::string error;
+        const unsigned int program = CompileProgram(kVS, kFS, &error);
+        ASSERT_NE(program, 0u) << error;
+
+        BindDefaultFramebuffer();
+        glViewport(0, 0, width, height);
+        glDisable(GL_SCISSOR_TEST);
+        glDisable(GL_DEPTH_TEST);
+        ClearTo(1.0f, 1.0f, 1.0f, 1.0f);
+        DrawFullViewportQuad(program);
+        EXPECT_EQ(FirstGLError(), 0u);
+
+        const Image painted = ReadPixels(width, height);
+        const Rgba8 centre = painted.At(width / 2, height / 2);
+        EXPECT_NEAR(centre.r, 26, 2) << "the draw did not survive the clear that preceded it";
+        EXPECT_NEAR(centre.g, 51, 2);
+        EXPECT_NEAR(centre.b, 77, 2);
+
+        gl.EndFrame();
+        glDeleteProgram(program);
+    }
+} // namespace MGITest

MobileGL/MG_IntegrationTest/Scenarios/DepthStencilReadbackScenario.cpp

@@ -0,0 +1,297 @@
+// MobileGL - MobileGL/MG_IntegrationTest/Scenarios/DepthStencilReadbackScenario.cpp
+// Copyright (c) 2025-2026 MobileGL-Dev
+// Licensed under the GNU Lesser General Public License v3.0:
+//   https://www.gnu.org/licenses/gpl-3.0.txt
+//   https://www.gnu.org/licenses/lgpl-3.0.txt
+// SPDX-License-Identifier: LGPL-3.0-only
+// End of Source File Header
+//
+// Scenario - glReadPixels OF DEPTH AND STENCIL FROM THE DEFAULT FRAMEBUFFER.
+//
+// DirectVulkan's depth/stencil readback used to decline the default framebuffer outright
+// (`ReadDepthStencilPixels` returned at its first line) because that framebuffer's depth and
+// stencil "attachments" are placeholder texture objects backing no image - the real one is the
+// swapchain's depth/stencil twin. Declining meant the call raised no GL error and wrote NOTHING,
+// so the caller kept whatever its buffer already held.
+//
+// That silence is what the framebuffer_blit family trips over. Every one of its cases begins by
+// clearing the default framebuffer's depth and stencil and reading them straight back as a
+// sanity check, into a local pre-initialised to 0.2 (depth) and 50 (stencil); an untouched
+// buffer therefore reports "expected DEPTH[0.25] but got DEPTH[0.2]" and "expected STENCIL[1] but
+// got STENCIL[50]" - the exact strings in the 15 Magma failures - long before any blit happens.
+// A test that only checked "no GL error" would pass against the broken path, so every case here
+// poisons its destination with a value the correct answer cannot be.
+//
+// The orientation case is the second half. This renderer stores the default framebuffer
+// display-side-up and converts GL rects on their way in, so the depth copy needs the same rect
+// mapping and row re-ordering the colour readback got in the M-1 fix; without them a
+// vertically-varying depth buffer reads back mirrored, which no full-extent uniform-value test
+// can see.
+//
+// Depth/stencil readback through a USER framebuffer already worked and is asserted here too, as
+// the built-in control: it shares ReadDepthStencilImageToClient with the default-framebuffer
+// path, so it is what says a failure is about the default framebuffer specifically.
+
+#include <cmath>
+#include <string>
+#include <vector>
+
+#include "../Harness/HeadlessGL.h"
+#include "../Harness/ScenarioFixture.h"
+
+#ifdef GLAPI
+#undef GLAPI
+#endif
+#define GL_GLEXT_PROTOTYPES
+#include <GL/gl.h>
+#include <GL/glcorearb.h>
+#undef GL_GLEXT_PROTOTYPES
+
+namespace MGITest {
+    namespace {
+
+        // Values no correct read can produce, so "the backend wrote nothing" fails loudly instead
+        // of passing on whatever happened to be in the variable. These are the CTS's own poison
+        // values, which is why its logs report exactly them.
+        constexpr float kDepthPoison = 0.2f;
+        constexpr int kStencilPoison = 50;
+
+        class DepthStencilReadbackScenario : public ScenarioTest {
+        protected:
+            // DirectGLES reads depth and stencil back through the ES driver, which has no
+            // guaranteed path for either (GL_NV_read_depth / GL_NV_read_stencil are optional and
+            // absent on both the Adreno device and Mesa's ES). That gap is tracked separately as
+            // the packed_depth_stencil cluster and needs a shader-sampling emulation, not this
+            // change; asserting it here would only pin a known-missing feature.
+            bool BackendReadsDepthStencil() const { return Gl().BackendName() == "DirectVulkan"; }
+
+            float ReadDepthAt(int x, int y) const {
+                float depth = kDepthPoison;
+                glReadPixels(x, y, 1, 1, GL_DEPTH_COMPONENT, GL_FLOAT, &depth);
+                return depth;
+            }
+
+            int ReadStencilAt(int x, int y) const {
+                int stencil = kStencilPoison;
+                glReadPixels(x, y, 1, 1, GL_STENCIL_INDEX, GL_INT, &stencil);
+                return stencil;
+            }
+        };
+
+        // A depth buffer whose value depends on the row: bottom half `bottom`, top half `top`.
+        // Built with a scissored clear rather than a draw so the test stays independent of
+        // depth-test and shader behaviour.
+        void ClearDepthInBands(int width, int height, float bottom, float top) {
+            glEnable(GL_SCISSOR_TEST);
+            glScissor(0, 0, width, height / 2);
+            glClearDepth(bottom);
+            glClear(GL_DEPTH_BUFFER_BIT);
+            glScissor(0, height / 2, width, height - height / 2);
+            glClearDepth(top);
+            glClear(GL_DEPTH_BUFFER_BIT);
+            glDisable(GL_SCISSOR_TEST);
+        }
+
+    } // namespace
+
+    TEST_F(DepthStencilReadbackScenario, DefaultFramebufferDepthClearIsVisibleToReadPixels) {
+        if (!Ready()) return;
+        if (!BackendReadsDepthStencil()) {
+            GTEST_SKIP() << "backend " << Gl().BackendName()
+                         << " has no depth readback path (ES lacks GL_NV_read_depth); see the packed_depth_stencil "
+                            "cluster";
+        }
+        HeadlessGL& gl = Gl();
+        const int width = gl.Width();
+        const int height = gl.Height();
+
+        BindDefaultFramebuffer();
+        glViewport(0, 0, width, height);
+        glDisable(GL_SCISSOR_TEST);
+        glDepthMask(GL_TRUE);
+        glClearDepth(0.25);
+        glClear(GL_DEPTH_BUFFER_BIT);
+
+        const float centre = ReadDepthAt(width / 2, height / 2);
+        EXPECT_EQ(FirstGLError(), 0u);
+        EXPECT_NEAR(centre, 0.25f, 1.0f / 4096.0f)
+            << "glReadPixels(GL_DEPTH_COMPONENT) of the default framebuffer returned " << centre
+            << (std::fabs(centre - kDepthPoison) < 1e-6f ? " - the destination was never written at all" : "");
+
+        gl.EndFrame();
+    }
+
+    TEST_F(DepthStencilReadbackScenario, DefaultFramebufferStencilClearIsVisibleToReadPixels) {
+        if (!Ready()) return;
+        if (!BackendReadsDepthStencil()) {
+            GTEST_SKIP() << "backend " << Gl().BackendName()
+                         << " has no stencil readback path (ES lacks GL_NV_read_stencil); see the "
+                            "packed_depth_stencil cluster";
+        }
+        HeadlessGL& gl = Gl();
+        const int width = gl.Width();
+        const int height = gl.Height();
+
+        BindDefaultFramebuffer();
+        glViewport(0, 0, width, height);
+        glDisable(GL_SCISSOR_TEST);
+        glStencilMask(0xFFu);
+        glClearStencil(3);
+        glClear(GL_STENCIL_BUFFER_BIT);
+
+        const int centre = ReadStencilAt(width / 2, height / 2);
+        EXPECT_EQ(FirstGLError(), 0u);
+        EXPECT_EQ(centre, 3) << "glReadPixels(GL_STENCIL_INDEX) of the default framebuffer returned " << centre
+                             << (centre == kStencilPoison ? " - the destination was never written at all" : "");
+
+        gl.EndFrame();
+    }
+
+    // The orientation half: a depth buffer that varies with the row must read back in GL's
+    // bottom-up order. A full-extent uniform clear is a fixed point of the flip, so only a banded
+    // buffer can tell the two apart.
+    TEST_F(DepthStencilReadbackScenario, DefaultFramebufferDepthReadbackKeepsTheGLRowOrder) {
+        if (!Ready()) return;
+        if (!BackendReadsDepthStencil()) {
+            GTEST_SKIP() << "backend " << Gl().BackendName() << " has no depth readback path";
+        }
+        HeadlessGL& gl = Gl();
+        const int width = gl.Width();
+        const int height = gl.Height();
+        ASSERT_GE(height, 8);
+
+        BindDefaultFramebuffer();
+        glViewport(0, 0, width, height);
+        glDepthMask(GL_TRUE);
+        ClearDepthInBands(width, height, /*bottom=*/0.25f, /*top=*/0.75f);
+        EXPECT_EQ(FirstGLError(), 0u);
+
+        const float bottom = ReadDepthAt(width / 2, height / 4);
+        const float top = ReadDepthAt(width / 2, height - 1 - height / 4);
+        EXPECT_EQ(FirstGLError(), 0u);
+        EXPECT_NEAR(bottom, 0.25f, 1.0f / 4096.0f)
+            << "GL row " << (height / 4) << " is in the bottom band and was cleared to 0.25, but read back " << bottom
+            << " (0.75 there means the readback is upside down)";
+        EXPECT_NEAR(top, 0.75f, 1.0f / 4096.0f)
+            << "GL row " << (height - 1 - height / 4) << " is in the top band and was cleared to 0.75, but read back "
+            << top << " (0.25 there means the readback is upside down)";
+
+        gl.EndFrame();
+    }
+
+    // A depth blit INTO the default framebuffer has to convert its rect out of GL's bottom-origin
+    // space, exactly as the colour blit does. The colour path had that conversion and the
+    // depth path did not, so a scissored depth blit landed in the mirrored band - which is the
+    // whole of KHR-GL*.framebuffer_blit.scissor_blit once the readback above works well enough to
+    // see it (before that the test died on the poison values and never reached the blit).
+    TEST_F(DepthStencilReadbackScenario, AScissoredDepthBlitIntoTheDefaultFramebufferLandsInTheScissorBox) {
+        if (!Ready()) return;
+        if (!BackendReadsDepthStencil()) {
+            GTEST_SKIP() << "backend " << Gl().BackendName() << " has no depth readback path";
+        }
+        HeadlessGL& gl = Gl();
+        const int width = gl.Width();
+        const int height = gl.Height();
+        ASSERT_GE(width, 8);
+        ASSERT_GE(height, 8);
+
+        // Source: a user framebuffer whose depth is uniformly 0.75.
+        GLuint fbo = 0, colorTex = 0, depthTex = 0;
+        glGenFramebuffers(1, &fbo);
+        glBindFramebuffer(GL_FRAMEBUFFER, fbo);
+        glGenTextures(1, &colorTex);
+        glBindTexture(GL_TEXTURE_2D, colorTex);
+        glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA8, width, height, 0, GL_RGBA, GL_UNSIGNED_BYTE, nullptr);
+        glFramebufferTexture2D(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_TEXTURE_2D, colorTex, 0);
+        glGenTextures(1, &depthTex);
+        glBindTexture(GL_TEXTURE_2D, depthTex);
+        glTexImage2D(GL_TEXTURE_2D, 0, GL_DEPTH24_STENCIL8, width, height, 0, GL_DEPTH_STENCIL,
+                     GL_UNSIGNED_INT_24_8, nullptr);
+        glFramebufferTexture2D(GL_FRAMEBUFFER, GL_DEPTH_STENCIL_ATTACHMENT, GL_TEXTURE_2D, depthTex, 0);
+        ASSERT_EQ(glCheckFramebufferStatus(GL_FRAMEBUFFER), GLenum(GL_FRAMEBUFFER_COMPLETE));
+        glViewport(0, 0, width, height);
+        glDisable(GL_SCISSOR_TEST);
+        glDepthMask(GL_TRUE);
+        glClearDepth(0.75);
+        glClear(GL_DEPTH_BUFFER_BIT);
+
+        // Destination: the default framebuffer, depth 0 everywhere.
+        BindDefaultFramebuffer();
+        glViewport(0, 0, width, height);
+        glClearDepth(0.0);
+        glClear(GL_DEPTH_BUFFER_BIT);
+
+        // Blit the whole rect, but scissored to the BOTTOM-LEFT quadrant in GL coordinates.
+        glEnable(GL_SCISSOR_TEST);
+        glScissor(0, 0, width / 2, height / 2);
+        glBindFramebuffer(GL_READ_FRAMEBUFFER, fbo);
+        glBindFramebuffer(GL_DRAW_FRAMEBUFFER, 0);
+        glBlitFramebuffer(0, 0, width, height, 0, 0, width, height, GL_DEPTH_BUFFER_BIT, GL_NEAREST);
+        glDisable(GL_SCISSOR_TEST);
+        glBindFramebuffer(GL_FRAMEBUFFER, 0);
+        EXPECT_EQ(FirstGLError(), 0u);
+
+        const float inside = ReadDepthAt(width / 4, height / 4);
+        const float above = ReadDepthAt(width / 4, height - 1 - height / 4);
+        EXPECT_EQ(FirstGLError(), 0u);
+        EXPECT_NEAR(inside, 0.75f, 1.0f / 4096.0f)
+            << "GL (" << (width / 4) << ", " << (height / 4) << ") is inside the scissor box and should hold the "
+            << "blitted 0.75, but read back " << inside;
+        EXPECT_NEAR(above, 0.0f, 1.0f / 4096.0f)
+            << "GL (" << (width / 4) << ", " << (height - 1 - height / 4)
+            << ") is ABOVE the scissor box and must still hold the cleared 0.0, but read back " << above
+            << " (0.75 there means the depth blit landed in the mirrored band)";
+
+        glDeleteTextures(1, &depthTex);
+        glDeleteTextures(1, &colorTex);
+        glDeleteFramebuffers(1, &fbo);
+        gl.EndFrame();
+    }
+
+    // The control: the same read against a user framebuffer, which never went through the
+    // declined path. It is what makes a failure above specific to the default framebuffer.
+    TEST_F(DepthStencilReadbackScenario, UserFramebufferDepthClearIsVisibleToReadPixels) {
+        if (!Ready()) return;
+        if (!BackendReadsDepthStencil()) {
+            GTEST_SKIP() << "backend " << Gl().BackendName() << " has no depth readback path";
+        }
+        HeadlessGL& gl = Gl();
+        const int width = 64;
+        const int height = 48;
+
+        GLuint fbo = 0, colorTex = 0, depthTex = 0;
+        glGenFramebuffers(1, &fbo);
+        glBindFramebuffer(GL_FRAMEBUFFER, fbo);
+        glGenTextures(1, &colorTex);
+        glBindTexture(GL_TEXTURE_2D, colorTex);
+        glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA8, width, height, 0, GL_RGBA, GL_UNSIGNED_BYTE, nullptr);
+        glFramebufferTexture2D(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_TEXTURE_2D, colorTex, 0);
+        glGenTextures(1, &depthTex);
+        glBindTexture(GL_TEXTURE_2D, depthTex);
+        glTexImage2D(GL_TEXTURE_2D, 0, GL_DEPTH24_STENCIL8, width, height, 0, GL_DEPTH_STENCIL,
+                     GL_UNSIGNED_INT_24_8, nullptr);
+        glFramebufferTexture2D(GL_FRAMEBUFFER, GL_DEPTH_STENCIL_ATTACHMENT, GL_TEXTURE_2D, depthTex, 0);
+        ASSERT_EQ(glCheckFramebufferStatus(GL_FRAMEBUFFER), GLenum(GL_FRAMEBUFFER_COMPLETE));
+        ASSERT_EQ(FirstGLError(), 0u);
+
+        glViewport(0, 0, width, height);
+        glDisable(GL_SCISSOR_TEST);
+        glDepthMask(GL_TRUE);
+        glStencilMask(0xFFu);
+        glClearDepth(0.5);
+        glClearStencil(7);
+        glClear(GL_DEPTH_BUFFER_BIT | GL_STENCIL_BUFFER_BIT);
+
+        const float depth = ReadDepthAt(width / 2, height / 2);
+        const int stencil = ReadStencilAt(width / 2, height / 2);
+        EXPECT_EQ(FirstGLError(), 0u);
+        EXPECT_NEAR(depth, 0.5f, 1.0f / 4096.0f) << "user-framebuffer depth readback returned " << depth;
+        EXPECT_EQ(stencil, 7) << "user-framebuffer stencil readback returned " << stencil;
+
+        glBindFramebuffer(GL_FRAMEBUFFER, 0);
+        glDeleteTextures(1, &depthTex);
+        glDeleteTextures(1, &colorTex);
+        glDeleteFramebuffers(1, &fbo);
+        gl.EndFrame();
+    }
+} // namespace MGITest

MobileGL/MG_IntegrationTest/Scenarios/FragCoordOriginScenario.cpp

@@ -0,0 +1,139 @@
+// MobileGL - MobileGL/MG_IntegrationTest/Scenarios/FragCoordOriginScenario.cpp
+// Copyright (c) 2025-2026 MobileGL-Dev
+// Licensed under the GNU Lesser General Public License v3.0:
+//   https://www.gnu.org/licenses/gpl-3.0.txt
+//   https://www.gnu.org/licenses/lgpl-3.0.txt
+// SPDX-License-Identifier: LGPL-3.0-only
+// End of Source File Header
+//
+// Scenario - gl_FragCoord ON THE DEFAULT FRAMEBUFFER CARRIES GL'S WINDOW ORIGIN.
+//
+// GL measures gl_FragCoord.y from the BOTTOM of the window. Vulkan's gl_FragCoord.y is the
+// framebuffer ROW being written, and DirectVulkan stores the default framebuffer display-side-up
+// (compensating for vertices by negating gl_Position.y), so a fragment's reported Y there was
+// `height - y_GL` - flipped, and for a viewport that does not span the full height, outside the
+// range GL promises entirely. GL CTS
+// `KHR-GL42.shader_image_load_store.basic-{allTargets-atomic,glsl-earlyFragTests,glsl-misc}`
+// caught it: each sets a small viewport at GL y=0 and does
+// `imageStore(image, ivec2(gl_FragCoord.xy), ...)` into an image exactly that size, so on a
+// 256-tall surface every store addressed rows 224..255 of a 32-row image and was dropped.
+//
+// The shader here paints each row with its own GL window Y, which is the whole claim in one
+// value: row j of the readback must be j, for a full-height viewport and for a half-height one
+// (the case where a flip and an offset can no longer hide each other). DirectGLES is the
+// built-in control - a native GL driver gets this right by construction, so a failure there
+// would mean the test, not the backend.
+
+#include <cstdint>
+#include <string>
+#include <vector>
+
+#include "../Harness/HeadlessGL.h"
+#include "../Harness/ScenarioFixture.h"
+
+#ifdef GLAPI
+#undef GLAPI
+#endif
+#define GL_GLEXT_PROTOTYPES
+#include <GL/gl.h>
+#include <GL/glcorearb.h>
+#undef GL_GLEXT_PROTOTYPES
+
+namespace MGITest {
+    namespace {
+
+        constexpr const char* kVS = R"(#version 330 core
+in vec2 aPos;
+void main() { gl_Position = vec4(aPos, 0.0, 1.0); }
+)";
+
+        // floor(gl_FragCoord.y) is the fragment's window row; 1/255 steps survive an RGBA8
+        // round trip exactly, so the readback byte IS the row the shader believes it is on.
+        constexpr const char* kFS = R"(#version 330 core
+out vec4 o_color;
+void main() { o_color = vec4(floor(gl_FragCoord.y) / 255.0, 0.0, 0.0, 1.0); }
+)";
+
+        class FragCoordOriginScenario : public ScenarioTest {};
+
+        // A quad covering the whole viewport, drawn with attribute 0 = aPos.
+        void DrawFullViewportQuad(unsigned int program) {
+            static const float kQuad[] = {-1.0f, -1.0f, 1.0f, -1.0f, -1.0f, 1.0f, 1.0f, 1.0f};
+            GLuint vao = 0, vbo = 0;
+            glGenVertexArrays(1, &vao);
+            glBindVertexArray(vao);
+            glGenBuffers(1, &vbo);
+            glBindBuffer(GL_ARRAY_BUFFER, vbo);
+            glBufferData(GL_ARRAY_BUFFER, sizeof(kQuad), kQuad, GL_STATIC_DRAW);
+            glEnableVertexAttribArray(0);
+            glVertexAttribPointer(0, 2, GL_FLOAT, GL_FALSE, 2 * sizeof(float), nullptr);
+            glUseProgram(program);
+            glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
+            glBindVertexArray(0);
+            glDeleteBuffers(1, &vbo);
+            glDeleteVertexArrays(1, &vao);
+        }
+
+        // Paints `viewportHeight` rows starting at GL y=0 and returns the red byte of each row.
+        std::vector<int> RowsPaintedWithTheirOwnWindowY(unsigned int program, int width, int viewportHeight) {
+            BindDefaultFramebuffer();
+            glViewport(0, 0, width, viewportHeight);
+            glDisable(GL_SCISSOR_TEST);
+            glDisable(GL_DEPTH_TEST);
+            ClearTo(0.0f, 0.0f, 1.0f, 1.0f);
+            DrawFullViewportQuad(program);
+
+            const Image image = ReadPixelsRect(0, 0, width, viewportHeight);
+            std::vector<int> rows;
+            rows.reserve(static_cast<std::size_t>(viewportHeight));
+            for (int y = 0; y < viewportHeight; ++y) {
+                rows.push_back(image.At(width / 2, y).r);
+            }
+            return rows;
+        }
+
+        ::testing::AssertionResult RowsAreTheirOwnIndex(const std::vector<int>& rows, const char* when) {
+            for (std::size_t y = 0; y < rows.size(); ++y) {
+                if (rows[y] != static_cast<int>(y)) {
+                    return ::testing::AssertionFailure()
+                           << when << ": GL window row " << y << " reported gl_FragCoord.y = " << rows[y]
+                           << " (expected " << y << "). Rows 0.." << (rows.size() - 1) << " read back as ["
+                           << rows.front() << " .. " << rows.back() << "].";
+                }
+            }
+            return ::testing::AssertionSuccess();
+        }
+
+    } // namespace
+
+    TEST_F(FragCoordOriginScenario, DefaultFramebufferFragCoordCountsFromTheBottom) {
+        if (!Ready()) return;
+        HeadlessGL& gl = Gl();
+        // 1/255 steps only stay distinguishable while the row index fits in a byte.
+        const int width = gl.Width();
+        const int fullHeight = std::min(gl.Height(), 256);
+        ASSERT_GE(fullHeight, 8) << "the harness surface is too small to tell rows apart";
+
+        std::string error;
+        const unsigned int program = CompileProgram(kVS, kFS, &error);
+        ASSERT_NE(program, 0u) << error;
+
+        // Full height first: this one passed even before the fix (a flip alone maps the row set
+        // onto itself), so it is the control that the shader and the readback agree at all.
+        EXPECT_TRUE(RowsAreTheirOwnIndex(RowsPaintedWithTheirOwnWindowY(program, width, fullHeight),
+                                         "full-height viewport"));
+
+        // Half height at GL y=0: the case the CTS failures were made of. A backend that reports
+        // the stored row here answers `height - y` for every row - off the bottom of the range,
+        // not merely reversed within it.
+        const int halfHeight = fullHeight / 2;
+        EXPECT_TRUE(RowsAreTheirOwnIndex(RowsPaintedWithTheirOwnWindowY(program, width, halfHeight),
+                                         "half-height viewport at GL y=0"));
+
+        glUseProgram(0);
+        glDeleteProgram(program);
+        glViewport(0, 0, gl.Width(), gl.Height());
+        EXPECT_EQ(FirstGLError(), 0u);
+    }
+
+} // namespace MGITest

MobileGL/MG_IntegrationTest/Scenarios/FragmentOutputArrayIndexScenario.cpp

@@ -0,0 +1,227 @@
+// MobileGL - MobileGL/MG_IntegrationTest/Scenarios/FragmentOutputArrayIndexScenario.cpp
+// Copyright (c) 2025-2026 MobileGL-Dev
+// Licensed under the GNU Lesser General Public License v3.0:
+//   https://www.gnu.org/licenses/gpl-3.0.txt
+//   https://www.gnu.org/licenses/lgpl-3.0.txt
+// SPDX-License-Identifier: LGPL-3.0-only
+// End of Source File Header
+//
+// Scenario - DYNAMICALLY INDEXED FRAGMENT OUTPUT ARRAYS, on a live driver.
+//
+// The bug: GLSL ES requires a *constant integral expression* to index a fragment output array
+// (GLSL ES 3.00 4.3.6); SPIR-V has no such rule. A shader that writes `coeff[i]` from a loop
+// therefore travels through glslang and SPIRV-Cross intact and lands on the ES driver as ESSL it
+// refuses outright - "array indexes for fragment outputs must be constant integral expressions".
+// The program links nothing and every draw that uses it becomes a silent no-op. That is the whole
+// of improved-transparency-minecraft-26.3 on the Android DirectGLES lane: Minecraft 26.3's OIT
+// coefficient shader has exactly this shape, and losing it empties the entire translucent layer
+// (clouds and water) while the opaque geometry stays pixel-exact.
+//
+// WHY THIS SCENARIO EXISTS RATHER THAN A UNIT TEST. The unit tests in MG_Test/Program (see
+// ProgramUtilTest, LoopDerivedFragmentOutputIndexFoldsToConstantIndices and its
+// genuinely-dynamic sibling) prove the SPIR-V comes out with constant indices, validates, and
+// decompiles to ESSL with only literal indices. What they cannot prove is that a real driver
+// then ACCEPTS and RUNS it - and acceptance is the whole failure mode, because Mesa accepts the
+// illegal form too. Only a live glCompileShader/glLinkProgram followed by a draw can tell the two
+// apart, and only reading the pixels back can tell "linked" from "wrote the right attachment".
+//
+// Both backends run this: on DirectVulkan the original module is already legal (the legalization
+// is DirectGLES-only, deliberately), so this doubles as the check that the two backends agree
+// about what such a shader means.
+
+#include <cmath>
+#include <string>
+#include <vector>
+
+#include "../Harness/HeadlessGL.h"
+#include "../Harness/ScenarioFixture.h"
+
+#ifdef GLAPI
+#undef GLAPI
+#endif
+#define GL_GLEXT_PROTOTYPES
+#include <GL/gl.h>
+#include <GL/glcorearb.h>
+#undef GL_GLEXT_PROTOTYPES
+
+namespace MGITest {
+    namespace {
+
+        constexpr const char* kVS = R"(#version 330 core
+in vec2 aPos;
+void main() {
+    gl_Position = vec4(aPos, 0.0, 1.0);
+}
+)";
+
+        // The Minecraft 26.3 OIT coefficient shape: both the attachment index and the component
+        // index come from loop counters, so nothing but the loop bounds decides where each value
+        // lands. Attachment 0 gets (0.0, 0.1, 0.2, 0.3) and attachment 1 gets (0.5, 0.6, 0.7, 0.8) -
+        // values that are only correct if the two indices were folded to the RIGHT constants, not
+        // merely to some constant.
+        constexpr const char* kLoopIndexedFS = R"(#version 330 core
+out vec4 coeff[2];
+void main() {
+    for (int attachmentIndex = 0; attachmentIndex < 2; ++attachmentIndex) {
+        for (int i = 0; i < 4; ++i) {
+            coeff[attachmentIndex][i] = float(attachmentIndex) * 0.5 + float(i) * 0.1;
+        }
+    }
+}
+)";
+
+        // No loop can fold this one: the index arrives in a uniform. It exercises the fallback
+        // lowering (a switch over the array range for the write, constant-indexed loads and a
+        // select for the read) and it checks the untargeted attachment is left ALONE, which a
+        // lowering that wrote every element unconditionally would break.
+        constexpr const char* kUniformIndexedFS = R"(#version 330 core
+uniform int uTarget;
+out vec4 coeff[2];
+void main() {
+    coeff[0] = vec4(0.25, 0.25, 0.25, 1.0);
+    coeff[1] = vec4(0.75, 0.75, 0.75, 1.0);
+    coeff[uTarget] = coeff[uTarget] + vec4(0.25, 0.0, 0.0, 0.0);
+}
+)";
+
+        constexpr int kSize = 8;
+
+        class FragmentOutputArrayIndexScenario : public ScenarioTest {
+        protected:
+            void SetUp() override {
+                ScenarioTest::SetUp();
+                if (!Ready()) return;
+
+                glGenFramebuffers(1, &m_fbo);
+                glBindFramebuffer(GL_FRAMEBUFFER, m_fbo);
+                for (int i = 0; i < 2; ++i) {
+                    glGenTextures(1, &m_color[i]);
+                    glBindTexture(GL_TEXTURE_2D, m_color[i]);
+                    glTexStorage2D(GL_TEXTURE_2D, 1, GL_RGBA8, kSize, kSize);
+                    glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
+                    glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
+                    glFramebufferTexture2D(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0 + i, GL_TEXTURE_2D,
+                                           m_color[i], 0);
+                }
+                const GLenum drawBuffers[2] = {GL_COLOR_ATTACHMENT0, GL_COLOR_ATTACHMENT1};
+                glDrawBuffers(2, drawBuffers);
+                ASSERT_EQ(glCheckFramebufferStatus(GL_FRAMEBUFFER),
+                          static_cast<GLenum>(GL_FRAMEBUFFER_COMPLETE));
+
+                const float quad[] = {-1.0f, -1.0f, 1.0f, -1.0f, -1.0f, 1.0f, 1.0f, 1.0f};
+                glGenVertexArrays(1, &m_vao);
+                glBindVertexArray(m_vao);
+                glGenBuffers(1, &m_vbo);
+                glBindBuffer(GL_ARRAY_BUFFER, m_vbo);
+                glBufferData(GL_ARRAY_BUFFER, sizeof(quad), quad, GL_STATIC_DRAW);
+                glEnableVertexAttribArray(0);
+                glVertexAttribPointer(0, 2, GL_FLOAT, GL_FALSE, 2 * sizeof(float), nullptr);
+                glViewport(0, 0, kSize, kSize);
+            }
+
+            void TearDown() override {
+                if (Ready()) {
+                    glBindFramebuffer(GL_FRAMEBUFFER, 0);
+                    glDeleteFramebuffers(1, &m_fbo);
+                    glDeleteTextures(2, m_color);
+                    glDeleteBuffers(1, &m_vbo);
+                    glDeleteVertexArrays(1, &m_vao);
+                }
+                ScenarioTest::TearDown();
+            }
+
+            // Clears both attachments to a colour no shader below writes, so an attachment that
+            // was never written reads back as the sentinel rather than as a plausible value.
+            void ClearToSentinel() {
+                glClearColor(0.0f, 0.0f, 1.0f, 1.0f);
+                glClear(GL_COLOR_BUFFER_BIT);
+            }
+
+            std::vector<float> ReadAttachment(int index) {
+                std::vector<unsigned char> bytes(static_cast<std::size_t>(kSize) * kSize * 4, 0);
+                glReadBuffer(GL_COLOR_ATTACHMENT0 + index);
+                glReadPixels(0, 0, kSize, kSize, GL_RGBA, GL_UNSIGNED_BYTE, bytes.data());
+                std::vector<float> centre(4, -1.0f);
+                // The middle pixel: the quad covers the whole target, so every pixel is the same,
+                // and the middle one cannot be a rasterization edge case.
+                const std::size_t offset = (static_cast<std::size_t>(kSize / 2) * kSize + kSize / 2) * 4;
+                for (int i = 0; i < 4; ++i) {
+                    centre[static_cast<std::size_t>(i)] = static_cast<float>(bytes[offset + i]) / 255.0f;
+                }
+                return centre;
+            }
+
+            GLuint m_fbo = 0;
+            GLuint m_color[2] = {0, 0};
+            GLuint m_vao = 0;
+            GLuint m_vbo = 0;
+        };
+
+        // The gate for the whole defect: before the legalization this program did not link on a
+        // strict ES driver (ANGLE), so the draw wrote nothing and BOTH attachments kept the
+        // sentinel. Now each attachment must carry the value its loop iteration produced.
+        TEST_F(FragmentOutputArrayIndexScenario, LoopIndexedOutputArrayWritesEveryAttachment) {
+            if (!Ready() || IsSkipped()) return;
+
+            std::string error;
+            const GLuint program = CompileProgram(kVS, kLoopIndexedFS, &error);
+            ASSERT_NE(program, 0u) << "a loop-indexed fragment output array must compile and link: "
+                                   << error;
+
+            ClearToSentinel();
+            glUseProgram(program);
+            glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
+
+            const std::vector<float> first = ReadAttachment(0);
+            EXPECT_NEAR(first[0], 0.0f, 0.02f) << "attachment 0 red";
+            EXPECT_NEAR(first[1], 0.1f, 0.02f) << "attachment 0 green";
+            EXPECT_NEAR(first[2], 0.2f, 0.02f)
+                << "attachment 0 blue - a sentinel 1.0 here means the draw never ran";
+            EXPECT_NEAR(first[3], 0.3f, 0.02f) << "attachment 0 alpha";
+
+            const std::vector<float> second = ReadAttachment(1);
+            EXPECT_NEAR(second[0], 0.5f, 0.02f)
+                << "attachment 1 red - the second loop iteration must reach the second draw buffer";
+            EXPECT_NEAR(second[1], 0.6f, 0.02f) << "attachment 1 green";
+            EXPECT_NEAR(second[2], 0.7f, 0.02f) << "attachment 1 blue";
+            EXPECT_NEAR(second[3], 0.8f, 0.02f) << "attachment 1 alpha";
+
+            glDeleteProgram(program);
+            EXPECT_EQ(FirstGLError(), 0u) << GLErrorName(FirstGLError());
+        }
+
+        // The fallback half, on a live driver, for both values of the uniform: the targeted
+        // attachment is read, incremented and written back; the other one keeps exactly what the
+        // constant-indexed store put there.
+        TEST_F(FragmentOutputArrayIndexScenario, UniformIndexedOutputArrayWritesOnlyTheSelectedAttachment) {
+            if (!Ready() || IsSkipped()) return;
+
+            std::string error;
+            const GLuint program = CompileProgram(kVS, kUniformIndexedFS, &error);
+            ASSERT_NE(program, 0u) << "a uniform-indexed fragment output array must compile and link: "
+                                   << error;
+            const GLint targetLocation = glGetUniformLocation(program, "uTarget");
+            ASSERT_GE(targetLocation, 0);
+            glUseProgram(program);
+
+            for (int target = 0; target < 2; ++target) {
+                ClearToSentinel();
+                glUniform1i(targetLocation, target);
+                glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
+
+                const std::vector<float> first = ReadAttachment(0);
+                const std::vector<float> second = ReadAttachment(1);
+                EXPECT_NEAR(first[0], target == 0 ? 0.5f : 0.25f, 0.02f)
+                    << "attachment 0 red with uTarget=" << target;
+                EXPECT_NEAR(first[1], 0.25f, 0.02f) << "attachment 0 green with uTarget=" << target;
+                EXPECT_NEAR(second[0], target == 1 ? 1.0f : 0.75f, 0.02f)
+                    << "attachment 1 red with uTarget=" << target;
+                EXPECT_NEAR(second[1], 0.75f, 0.02f) << "attachment 1 green with uTarget=" << target;
+            }
+
+            glDeleteProgram(program);
+            EXPECT_EQ(FirstGLError(), 0u) << GLErrorName(FirstGLError());
+        }
+
+    } // namespace
+} // namespace MGITest

MobileGL/MG_IntegrationTest/Scenarios/Glsl420DeclarationScenario.cpp

@@ -0,0 +1,476 @@
+// MobileGL - MobileGL/MG_IntegrationTest/Scenarios/Glsl420DeclarationScenario.cpp
+// Copyright (c) 2025-2026 MobileGL-Dev
+// Licensed under the GNU Lesser General Public License v3.0:
+//   https://www.gnu.org/licenses/gpl-3.0.txt
+//   https://www.gnu.org/licenses/lgpl-3.0.txt
+// SPDX-License-Identifier: LGPL-3.0-only
+// End of Source File Header
+//
+// Scenario - GLSL 4.20 DECLARATIONS THE FRONTEND USED TO REJECT OR COLLAPSE.
+//
+// GLSL 4.20 gives an array of opaque uniforms or of block instances CONSECUTIVE binding
+// points: "layout(binding = 1) uniform sampler2D goku[7]" puts goku[0] on texture unit 1
+// and goku[6] on unit 7, and the same rule holds for "layout(binding = 2) uniform GOKU
+// {...} goku[14]" over uniform buffer binding points 2..15 (GLSL 4.20 4.4.5, GL 4.6 7.6.2).
+// One qualifier, N bindings - which is exactly the part that is easy to get wrong, because
+// every element shares one declaration and one reflection record.
+//
+// Three separate mechanisms all collapsed that array down to its first element, and the
+// three cases below pin one each:
+//
+//   * the SAMPLER array (Espryt): reflection names an array after its first element at
+//     every location it spans, so the backend resolved "goku[0]" once per element, got one
+//     backend location N times, and the per-draw pass's last glUniform1i was the only one
+//     that survived. goku[0] ended up holding the LAST element's unit and goku[1..N-1] kept
+//     unit 0 - so every element sampled whatever was bound to unit 0.
+//   * the uniform BLOCK array (both backends): glslang reports the declared binding for
+//     every expanded instance, so nothing added the element offset. glGetActiveUniformBlockiv
+//     answered the base binding for all of them, and since both backends feed a block from
+//     that same number at draw time, all instances also read one buffer.
+//   * 'invariant' on a non-vertex stage's INPUT: legal desktop GLSL at every version, and
+//     ignored where it is written, but glslang rejected it from 4.20 up - so a shader that
+//     compiled as "#version 400" stopped compiling as "#version 420".
+//
+// The fourth case is the same species as the third - a legal 4.20 shader the frontend
+// refused - and lives here for that reason: atomicCounterIncrement() was rejected because
+// glslang applied its atomicAdd() extension gate to the atomicAdd() its own Vulkan-relaxed
+// lowering had just synthesized.
+//
+// Conformance cases behind these: KHR-GL42.shading_language_420pack.binding_sampler_array,
+// .binding_uniform_block_array, .qualifier_order[_block]_test_id_*, and
+// KHR-GL42.shader_image_load_store.advanced-sso-atomicCounters.
+
+#include <cstdint>
+#include <string>
+#include <vector>
+
+#include "../Harness/HeadlessGL.h"
+#include "../Harness/ScenarioFixture.h"
+
+#ifdef GLAPI
+#undef GLAPI
+#endif
+#define GL_GLEXT_PROTOTYPES
+#include <GL/gl.h>
+#include <GL/glcorearb.h>
+#undef GL_GLEXT_PROTOTYPES
+
+namespace MGITest {
+    namespace {
+
+        constexpr int kElements = 4;
+
+        // No vertex attributes: the quad comes from gl_VertexID, so nothing here depends on
+        // the harness's attribute pinning and the fragment stage is the only thing under test.
+        constexpr const char* kQuadVS = R"(#version 420 core
+void main()
+{
+    switch (gl_VertexID)
+    {
+      case 0: gl_Position = vec4(-1.0, -1.0, 0.0, 1.0); break;
+      case 1: gl_Position = vec4( 1.0, -1.0, 0.0, 1.0); break;
+      case 2: gl_Position = vec4(-1.0,  1.0, 0.0, 1.0); break;
+      default: gl_Position = vec4( 1.0,  1.0, 0.0, 1.0); break;
+    }
+}
+)";
+
+        // The red channel comes back as a BITMASK of which elements read the wrong thing, so
+        // a failure names the element instead of just saying "not green". float(bad)/255.0
+        // round-trips exactly through an RGBA8 target for every mask this can produce.
+        constexpr const char* kSamplerArrayFS = R"(#version 420 core
+layout(binding = 1) uniform sampler2D goku[4];
+out vec4 o_color;
+void main()
+{
+    const vec2 uv = vec2(0.5, 0.5);
+    int bad = 0;
+    if (texture(goku[0], uv) != vec4(1.0, 0.0, 0.0, 1.0)) bad |= 1;
+    if (texture(goku[1], uv) != vec4(0.0, 0.0, 1.0, 1.0)) bad |= 2;
+    if (texture(goku[2], uv) != vec4(1.0, 1.0, 0.0, 1.0)) bad |= 4;
+    if (texture(goku[3], uv) != vec4(0.0, 1.0, 1.0, 1.0)) bad |= 8;
+    o_color = vec4(float(bad) / 255.0, bad == 0 ? 1.0 : 0.0, 0.0, 1.0);
+}
+)";
+
+        // Same declaration one dimension deeper. GLSL 4.30 arrays of arrays are legal here, and
+        // the elements still take consecutive units (1..4) in declaration order - but the two
+        // reflections disagree about how to count them, which is the whole point of this case.
+        constexpr const char* kSamplerArrayOfArraysFS = R"(#version 430 core
+layout(binding = 1) uniform sampler2D goku[2][2];
+out vec4 o_color;
+void main()
+{
+    const vec2 uv = vec2(0.5, 0.5);
+    int bad = 0;
+    if (texture(goku[0][0], uv) != vec4(1.0, 0.0, 0.0, 1.0)) bad |= 1;
+    if (texture(goku[0][1], uv) != vec4(0.0, 0.0, 1.0, 1.0)) bad |= 2;
+    if (texture(goku[1][0], uv) != vec4(1.0, 1.0, 0.0, 1.0)) bad |= 4;
+    if (texture(goku[1][1], uv) != vec4(0.0, 1.0, 1.0, 1.0)) bad |= 8;
+    o_color = vec4(float(bad) / 255.0, bad == 0 ? 1.0 : 0.0, 0.0, 1.0);
+}
+)";
+
+        constexpr const char* kBlockArrayFS = R"(#version 420 core
+layout(std140, binding = 2) uniform GOKU
+{
+    vec4 gohan;
+} goku[4];
+out vec4 o_color;
+void main()
+{
+    int bad = 0;
+    if (goku[0].gohan != vec4(1.0, 0.0, 0.0, 1.0)) bad |= 1;
+    if (goku[1].gohan != vec4(0.0, 0.0, 1.0, 1.0)) bad |= 2;
+    if (goku[2].gohan != vec4(1.0, 1.0, 0.0, 1.0)) bad |= 4;
+    if (goku[3].gohan != vec4(0.0, 1.0, 1.0, 1.0)) bad |= 8;
+    o_color = vec4(float(bad) / 255.0, bad == 0 ? 1.0 : 0.0, 0.0, 1.0);
+}
+)";
+
+        // The producing stage declares the varying invariant (always legal) and the consuming
+        // stage redeclares it (the part that regressed at 4.20). The qualifier ORDER is the
+        // shuffled one 420pack exists to allow, so this also covers the parse path the
+        // qualifier_order cases exercise.
+        constexpr const char* kInvariantInVS = R"(#version 420 core
+smooth invariant out highp vec4 v_data;
+void main()
+{
+    v_data = vec4(0.0, 1.0, 0.0, 1.0);
+    switch (gl_VertexID)
+    {
+      case 0: gl_Position = vec4(-1.0, -1.0, 0.0, 1.0); break;
+      case 1: gl_Position = vec4( 1.0, -1.0, 0.0, 1.0); break;
+      case 2: gl_Position = vec4(-1.0,  1.0, 0.0, 1.0); break;
+      default: gl_Position = vec4( 1.0,  1.0, 0.0, 1.0); break;
+    }
+}
+)";
+
+        constexpr const char* kInvariantInFS = R"(#version 420 core
+highp in smooth invariant vec4 v_data;
+out vec4 o_color;
+void main() { o_color = v_data; }
+)";
+
+        // atomicCounterIncrement() is core GLSL from 4.20 and needs no extension. MobileGL
+        // parses under Vulkan-relaxed rules, which rewrite it into an atomicAdd() on a buffer
+        // block - and glslang then applied to its OWN rewrite the desktop-below-430 gate that
+        // demands GL_ARB_shader_storage_buffer_object for atomicAdd, rejecting a shader it had
+        // just accepted. The shape is lifted from
+        // KHR-GL42.shader_image_load_store.advanced-sso-atomicCounters.
+        constexpr const char* kAtomicCounterVS = R"(#version 420 core
+layout(binding = 0, offset = 0) uniform atomic_uint g_counter;
+out flat uint v_index;
+void main()
+{
+    v_index = atomicCounterIncrement(g_counter);
+    switch (gl_VertexID)
+    {
+      case 0: gl_Position = vec4(-1.0, -1.0, 0.0, 1.0); break;
+      case 1: gl_Position = vec4( 1.0, -1.0, 0.0, 1.0); break;
+      case 2: gl_Position = vec4(-1.0,  1.0, 0.0, 1.0); break;
+      default: gl_Position = vec4( 1.0,  1.0, 0.0, 1.0); break;
+    }
+}
+)";
+
+        constexpr const char* kAtomicCounterFS = R"(#version 420 core
+in flat uint v_index;
+out vec4 o_color;
+void main() { o_color = vec4(0.0, 1.0, 0.0, 1.0); }
+)";
+
+        class Glsl420DeclarationScenario : public ScenarioTest {
+        protected:
+            void TearDown() override {
+                if (!Ready()) return;
+                glUseProgram(0);
+                if (!m_textures.empty()) glDeleteTextures(static_cast<GLsizei>(m_textures.size()), m_textures.data());
+                if (!m_buffers.empty()) glDeleteBuffers(static_cast<GLsizei>(m_buffers.size()), m_buffers.data());
+                for (GLuint p : m_programs) glDeleteProgram(p);
+                if (m_vao != 0) glDeleteVertexArrays(1, &m_vao);
+                m_textures.clear();
+                m_buffers.clear();
+                m_programs.clear();
+                m_vao = 0;
+            }
+
+            GLuint Build(const char* vs, const char* fs) {
+                std::string error;
+                const GLuint program = CompileProgram(vs, fs, &error);
+                if (program == 0) {
+                    ADD_FAILURE() << "program did not build: " << error;
+                    return 0;
+                }
+                m_programs.push_back(program);
+                return program;
+            }
+
+            // One 1x1 RGBA8 texture per element, each a colour whose channels are exactly 0 or
+            // 255 so the shader's == comparisons are exact.
+            void MakeElementTextures(const std::uint8_t colors[kElements][4]) {
+                m_textures.assign(kElements, 0);
+                glGenTextures(kElements, m_textures.data());
+                for (int i = 0; i < kElements; ++i) {
+                    glActiveTexture(GL_TEXTURE0 + 1 + i);
+                    glBindTexture(GL_TEXTURE_2D, m_textures[i]);
+                    glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA8, 1, 1, 0, GL_RGBA, GL_UNSIGNED_BYTE, colors[i]);
+                    glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
+                    glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
+                    glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAX_LEVEL, 0);
+                }
+                glActiveTexture(GL_TEXTURE0);
+            }
+
+            void MakeElementBuffers(const float values[kElements][4], GLuint firstBinding) {
+                m_buffers.assign(kElements, 0);
+                glGenBuffers(kElements, m_buffers.data());
+                for (int i = 0; i < kElements; ++i) {
+                    glBindBuffer(GL_UNIFORM_BUFFER, m_buffers[i]);
+                    glBufferData(GL_UNIFORM_BUFFER, 4 * sizeof(float), values[i], GL_STATIC_DRAW);
+                    glBindBufferBase(GL_UNIFORM_BUFFER, firstBinding + i, m_buffers[i]);
+                }
+                glBindBuffer(GL_UNIFORM_BUFFER, 0);
+            }
+
+            // Draws the full-screen quad and hands back the centre pixel.
+            Rgba8 DrawAndRead(GLuint program) {
+                HeadlessGL& gl = Gl();
+                if (m_vao == 0) glGenVertexArrays(1, &m_vao);
+                glBindVertexArray(m_vao);
+                BindDefaultFramebuffer();
+                glViewport(0, 0, gl.Width(), gl.Height());
+                glDisable(GL_SCISSOR_TEST);
+                glDisable(GL_DEPTH_TEST);
+                ClearTo(0.0f, 0.0f, 0.0f, 1.0f);
+                glUseProgram(program);
+                glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
+                const Image image = ReadPixels(gl.Width(), gl.Height());
+                glUseProgram(0);
+                return image.At(gl.Width() / 2, gl.Height() / 2);
+            }
+
+            // An array of ARRAYS is declined by Magma (ProgramFactory::ReflectLayout logs it and
+            // VkProgramObject::declinedDescriptors then refuses every draw), which is a defined
+            // outcome the case below can assert. Espryt has no such gate: it bakes the units the
+            // frontend reports into its ESSL, and since the binding-qualifier seeding does not
+            // walk the inner dimension every element reports unit 0 - so it samples one texture
+            // four times and paints a mismatch. That gap is in the FRONTEND, one level below
+            // either backend, and fixing it is the feature that would make this shape work
+            // everywhere; it is not part of wiring descriptor arrays through Magma, so the
+            // Espryt arm is SCOPED and the reflection half is asserted on both backends.
+            bool MultiDimensionalSamplerArraysAreDeclined() const { return Gl().BackendName() == "DirectVulkan"; }
+
+            // Same shape, different gap: with the compile fixed, this shader now links on
+            // both backends but paints nothing on Magma - the atomic counter becomes a
+            // buffer descriptor there and that half is not wired up yet (the conformance
+            // case KHR-GL42.shader_image_load_store.advanced-sso-atomicCounters is where it
+            // is measured). The regression this case exists for is the COMPILE, which is
+            // asserted on both backends above; only the paint is scoped.
+            bool AtomicCounterDrawsAreSupported() const { return Gl().BackendName() != "DirectVulkan"; }
+
+            static std::string BadElements(std::uint8_t mask) {
+                if (mask == 0) return "none";
+                std::string out;
+                for (int i = 0; i < kElements; ++i) {
+                    if ((mask & (1u << i)) == 0) continue;
+                    if (!out.empty()) out += ", ";
+                    out += "[" + std::to_string(i) + "]";
+                }
+                return out;
+            }
+
+            std::vector<GLuint> m_textures;
+            std::vector<GLuint> m_buffers;
+            std::vector<GLuint> m_programs;
+            GLuint m_vao = 0;
+        };
+
+    } // namespace
+
+    // Element k of a sampler array samples texture unit N+k - both as the API reports it and,
+    // the part that was actually broken, as the draw behaves.
+    TEST_F(Glsl420DeclarationScenario, SamplerArrayElementsSampleConsecutiveTextureUnits) {
+        if (!Ready()) return;
+
+        static const std::uint8_t colors[kElements][4] = {
+            {255, 0, 0, 255}, {0, 0, 255, 255}, {255, 255, 0, 255}, {0, 255, 255, 255}};
+        MakeElementTextures(colors);
+
+        const GLuint program = Build(kQuadVS, kSamplerArrayFS);
+        if (program == 0) return;
+
+        // The reported unit is the shadow the frontend seeds from the qualifier. It was
+        // already right when the draw was wrong, so checking only this would have passed
+        // straight through the bug - it is here to separate a reflection regression from a
+        // backend one if this case ever fails again.
+        glUseProgram(program);
+        for (int i = 0; i < kElements; ++i) {
+            const std::string name = "goku[" + std::to_string(i) + "]";
+            const GLint location = glGetUniformLocation(program, name.c_str());
+            ASSERT_GE(location, 0) << name << " has no location";
+            GLint unit = -1;
+            glGetUniformiv(program, location, &unit);
+            EXPECT_EQ(unit, 1 + i) << name << " should default to texture unit " << (1 + i);
+        }
+        glUseProgram(0);
+
+        const Rgba8 centre = DrawAndRead(program);
+        EXPECT_EQ(FirstGLError(), 0u);
+        EXPECT_EQ(centre.r, 0) << "sampler array elements that read the wrong texture: " << BadElements(centre.r);
+        EXPECT_EQ(centre.g, 255) << "the draw did not reach the fragment stage at all";
+    }
+
+    // An array of ARRAYS of samplers is the shape the two reflections count differently:
+    // SPIRV-Reflect reports one binding of 4 flattened descriptors, while the frontend hands out
+    // uniform locations along the outer dimension only and keys the uniform by its full
+    // "goku[0][0]" spelling. Magma therefore cannot address elements 1..3 of that binding, and
+    // the contract this case pins is that it says so and DECLINES - the failure it must never
+    // return to is resolving those elements onto whatever uniform got the next locations, which
+    // is a silently wrong texture rather than a missing draw.
+    //
+    // Deliberately weak on the pixels for that reason: what is asserted on every backend is that
+    // the program builds, the draw raises no GL error, and the process survives. Where the
+    // descriptors do resolve, the colours are checked too.
+    TEST_F(Glsl420DeclarationScenario, AnArrayOfSamplerArraysIsHonouredOrDeclinedCleanly) {
+        if (!Ready()) return;
+
+        static const std::uint8_t colors[kElements][4] = {
+            {255, 0, 0, 255}, {0, 0, 255, 255}, {255, 255, 0, 255}, {0, 255, 255, 255}};
+        MakeElementTextures(colors);
+
+        std::string error;
+        const GLuint program = CompileProgram(kQuadVS, kSamplerArrayOfArraysFS, &error);
+        if (program == 0) {
+            GTEST_SKIP() << "the frontend does not build an array of sampler arrays: " << error;
+        }
+        m_programs.push_back(program);
+
+        // The reflection DOES reserve one location per flattened element, in the order
+        // SPIRV-Reflect flattens them - which is the whole reason baseLocation + element is the
+        // right addressing rule for a descriptor array, and would be right for this shape too.
+        // What is missing is one level up: the `layout(binding = 1)` unit seeding walks the outer
+        // dimension only, so all four elements report unit 0 instead of 1..4. That is why this
+        // shape is declined rather than supported, and it is asserted here because the day the
+        // seeding learns arrays of arrays, the decline should be revisited rather than kept.
+        glUseProgram(program);
+        for (int outer = 0; outer < 2; ++outer) {
+            for (int inner = 0; inner < 2; ++inner) {
+                const std::string name = "goku[" + std::to_string(outer) + "][" + std::to_string(inner) + "]";
+                EXPECT_EQ(glGetUniformLocation(program, name.c_str()), outer * 2 + inner)
+                    << name << " should hold the flattened element's own location";
+            }
+        }
+        glUseProgram(0);
+
+        const Rgba8 centre = DrawAndRead(program);
+        EXPECT_EQ(FirstGLError(), 0u) << "declining a descriptor array must not raise a GL error";
+
+        if (!MultiDimensionalSamplerArraysAreDeclined()) {
+            GTEST_SKIP() << "the frontend's binding-qualifier seeding does not walk an array of arrays, so "
+                         << Gl().BackendName() << " samples unit 0 for every element; the locations "
+                         << "asserted above are the half of this case it can answer";
+        }
+
+        // Three outcomes are possible and only two are acceptable. Green means every element
+        // sampled its own unit. Black - the untouched clear - means the program was declined and
+        // painted nothing, which is the documented Magma outcome. A non-zero red channel is the
+        // third: the draw DID reach the fragment stage and elements read the wrong textures,
+        // which is exactly the silent mismatch this decline exists to prevent.
+        if (centre.g == 255) {
+            EXPECT_EQ(centre.r, 0) << "elements of the array of arrays that read the wrong texture: "
+                                   << BadElements(centre.r);
+            return;
+        }
+        EXPECT_EQ(centre.r, 0) << "the array of arrays was not resolved, but the draw still painted "
+                                  "a mismatch instead of being declined: " << BadElements(centre.r);
+    }
+
+    // Instance k of a uniform block array sits on buffer binding point N+k - again both as
+    // reported and as fed to the shader.
+    TEST_F(Glsl420DeclarationScenario, UniformBlockArrayInstancesTakeConsecutiveBindings) {
+        if (!Ready()) return;
+
+        static const float values[kElements][4] = {
+            {1.0f, 0.0f, 0.0f, 1.0f}, {0.0f, 0.0f, 1.0f, 1.0f}, {1.0f, 1.0f, 0.0f, 1.0f}, {0.0f, 1.0f, 1.0f, 1.0f}};
+        constexpr GLuint kFirstBinding = 2;
+        MakeElementBuffers(values, kFirstBinding);
+
+        const GLuint program = Build(kQuadVS, kBlockArrayFS);
+        if (program == 0) return;
+
+        for (int i = 0; i < kElements; ++i) {
+            const std::string name = "GOKU[" + std::to_string(i) + "]";
+            const GLuint index = glGetUniformBlockIndex(program, name.c_str());
+            ASSERT_NE(index, static_cast<GLuint>(GL_INVALID_INDEX)) << name << " is not an active block";
+            GLint binding = -1;
+            glGetActiveUniformBlockiv(program, index, GL_UNIFORM_BLOCK_BINDING, &binding);
+            EXPECT_EQ(binding, static_cast<GLint>(kFirstBinding) + i)
+                << name << " should start on binding point " << (kFirstBinding + i);
+        }
+        EXPECT_EQ(FirstGLError(), 0u) << "the block queries left a GL error behind";
+
+        const Rgba8 centre = DrawAndRead(program);
+        EXPECT_EQ(FirstGLError(), 0u);
+        EXPECT_EQ(centre.r, 0) << "block array instances that read the wrong buffer: " << BadElements(centre.r);
+        EXPECT_EQ(centre.g, 255) << "the draw did not reach the fragment stage at all";
+    }
+
+    // 'invariant' written on a fragment input at #version 420. The same source compiles at
+    // #version 400 on any implementation, so a version-dependent rejection is the defect.
+    TEST_F(Glsl420DeclarationScenario, InvariantIsAcceptedOnANonVertexStageInput) {
+        if (!Ready()) return;
+
+        const GLuint program = Build(kInvariantInVS, kInvariantInFS);
+        if (program == 0) return;
+
+        const Rgba8 centre = DrawAndRead(program);
+        EXPECT_EQ(FirstGLError(), 0u);
+        EXPECT_EQ(centre.g, 255) << "the invariant-qualified varying did not arrive";
+        EXPECT_EQ(centre.r, 0);
+    }
+
+    // A #version 420 shader may call atomicCounterIncrement() with no extension at all. The
+    // assertion is deliberately the COMPILE, because the defect was a compile-time gate on
+    // glslang's own atomic-counter lowering; the draw that follows only checks the shader
+    // survives the rest of the pipeline without leaving an error behind.
+    TEST_F(Glsl420DeclarationScenario, AnAtomicCounterCompilesWithoutTheSsboExtension) {
+        if (!Ready()) return;
+
+        const GLuint shader = glCreateShader(GL_VERTEX_SHADER);
+        glShaderSource(shader, 1, &kAtomicCounterVS, nullptr);
+        glCompileShader(shader);
+        GLint compiled = GL_FALSE;
+        glGetShaderiv(shader, GL_COMPILE_STATUS, &compiled);
+        if (compiled == GL_FALSE) {
+            char log[2048] = {};
+            glGetShaderInfoLog(shader, sizeof(log) - 1, nullptr, log);
+            glDeleteShader(shader);
+            FAIL() << "atomicCounterIncrement() at #version 420 core did not compile: " << log;
+        }
+        glDeleteShader(shader);
+
+        const GLuint program = Build(kAtomicCounterVS, kAtomicCounterFS);
+        if (program == 0) return;
+
+        GLuint counter = 0;
+        glGenBuffers(1, &counter);
+        m_buffers.push_back(counter);
+        const GLuint zero = 0;
+        glBindBuffer(GL_ATOMIC_COUNTER_BUFFER, counter);
+        glBufferData(GL_ATOMIC_COUNTER_BUFFER, sizeof(GLuint), &zero, GL_DYNAMIC_DRAW);
+        glBindBufferBase(GL_ATOMIC_COUNTER_BUFFER, 0, counter);
+        glBindBuffer(GL_ATOMIC_COUNTER_BUFFER, 0);
+
+        if (!AtomicCounterDrawsAreSupported()) {
+            GTEST_SKIP() << "atomic-counter draws do not paint on " << Gl().BackendName()
+                         << " yet; the compile above is what this case pins";
+        }
+
+        const Rgba8 centre = DrawAndRead(program);
+        EXPECT_EQ(FirstGLError(), 0u);
+        EXPECT_EQ(centre.g, 255) << "the atomic-counter shader linked but painted nothing";
+    }
+
+} // namespace MGITest

MobileGL/MG_IntegrationTest/Scenarios/ImageLoadStoreSsoScenario.cpp

@@ -0,0 +1,473 @@
+// MobileGL - MobileGL/MG_IntegrationTest/Scenarios/ImageLoadStoreSsoScenario.cpp
+// Copyright (c) 2025-2026 MobileGL-Dev
+// Licensed under the GNU Lesser General Public License v3.0:
+//   https://www.gnu.org/licenses/gpl-3.0.txt
+//   https://www.gnu.org/licenses/lgpl-3.0.txt
+// SPDX-License-Identifier: LGPL-3.0-only
+// End of Source File Header
+//
+// Scenario - IMAGE UNIFORMS REACHED THROUGH A PROGRAM PIPELINE.
+//
+// KHR-GL42.shader_image_load_store.advanced-sso-simple reduced to its mechanism. An ARRAY of
+// image uniforms lives in a separable FRAGMENT program; the application assigns each element its
+// own image unit with glProgramUniform1i, on a program that is not current and whose pipeline is
+// not even bound yet; the draw then goes through the pipeline, i.e. through the flattened
+// composite program (MG_State/GLState/Core.cpp, GetProgramForDraw) rather than through the stage
+// program the units were written to.
+//
+// Three separate things have to survive that indirection, and each one is a different mechanism:
+//
+//   1. the units themselves, which are per-program state on a DIFFERENT object from the one the
+//      draw reads (the composite mirror carries them);
+//   2. the units as seen by a backend that cannot take them at draw time - Espryt has to BAKE an
+//      image unit into the ESSL it generates, because ES forbids glUniform1i on image uniforms,
+//      so a change has to invalidate the generated program;
+//   3. per-ELEMENT assignment, which is what makes this different from every sampler case: the
+//      four elements of g_image[] are four locations with four different units, and nothing may
+//      collapse them to the array's base.
+//
+// Two pipelines that SHARE their vertex stage program and differ only in the fragment one are
+// used exactly as the conformance case does, because that is what makes the composite cache and
+// the stage programs' separate uniform storage both load-bearing at once.
+
+#include <string>
+#include <vector>
+
+#include "../Harness/HeadlessGL.h"
+#include "../Harness/ScenarioFixture.h"
+
+#ifdef GLAPI
+#undef GLAPI
+#endif
+#define GL_GLEXT_PROTOTYPES
+#include <GL/gl.h>
+#include <GL/glcorearb.h>
+#undef GL_GLEXT_PROTOTYPES
+
+namespace MGITest {
+    namespace {
+
+        constexpr const char* kSsoVS = R"(#version 420 core
+out gl_PerVertex { vec4 gl_Position; };
+void main()
+{
+    switch (gl_VertexID)
+    {
+      case 0: gl_Position = vec4(-1.0, -1.0, 0.0, 1.0); break;
+      case 1: gl_Position = vec4( 1.0, -1.0, 0.0, 1.0); break;
+      case 2: gl_Position = vec4(-1.0,  1.0, 0.0, 1.0); break;
+      case 3: gl_Position = vec4( 1.0,  1.0, 0.0, 1.0); break;
+    }
+}
+)";
+
+        // The conformance case's two fragment programs: one with an explicit format qualifier,
+        // one writeonly with none. Both write every element of a four-image array and discard.
+        constexpr const char* kImageFS0 = R"(#version 420 core
+layout(rgba32f) uniform image2D g_image[4];
+void main()
+{
+    for (int i = 0; i < g_image.length(); ++i) {
+        imageStore(g_image[i], ivec2(gl_FragCoord), vec4(1.0));
+    }
+    discard;
+}
+)";
+
+        constexpr const char* kImageFS1 = R"(#version 420 core
+writeonly uniform image2D g_image[4];
+void main()
+{
+    for (int i = 0; i < g_image.length(); ++i) {
+        imageStore(g_image[i], ivec2(gl_FragCoord), vec4(2.0));
+    }
+    discard;
+}
+)";
+
+        class ImageLoadStoreSsoScenario : public ScenarioTest {
+        protected:
+            void TearDown() override {
+                if (!Ready()) return;
+                glBindProgramPipeline(0);
+                glUseProgram(0);
+                for (GLuint p : m_programs) glDeleteProgram(p);
+                for (GLuint p : m_pipelines) glDeleteProgramPipelines(1, &p);
+                m_programs.clear();
+                m_pipelines.clear();
+            }
+
+            GLuint MakeSeparable(GLenum stage, const char* source) {
+                const GLuint program = glCreateShaderProgramv(stage, 1, &source);
+                if (program != 0) m_programs.push_back(program);
+                EXPECT_EQ(FirstGLError(), 0u)
+                    << "glCreateShaderProgramv(stage 0x" << std::hex << stage << std::dec << ") left a GL error";
+                GLint linked = GL_FALSE;
+                glGetProgramiv(program, GL_LINK_STATUS, &linked);
+                if (linked == GL_FALSE) {
+                    char log[2048] = {};
+                    glGetProgramInfoLog(program, sizeof(log) - 1, nullptr, log);
+                    ADD_FAILURE() << "glCreateShaderProgramv(stage 0x" << std::hex << stage << std::dec
+                                  << ") did not link: " << log;
+                    return 0;
+                }
+                return program;
+            }
+
+            GLuint MakePipeline() {
+                GLuint pipeline = 0;
+                glGenProgramPipelines(1, &pipeline);
+                m_pipelines.push_back(pipeline);
+                return pipeline;
+            }
+
+            // Espryt reaches the GPU through an ES driver, and ES forbids glUniform1i on an
+            // image uniform: the unit has to be BAKED into the generated ESSL as
+            // layout(binding = N) (RebindImageUniformsToFrontendUnits, MG_Backend/DirectGLES).
+            // One qualifier is all an ARRAY declaration can carry, and ESSL then gives the
+            // array's elements the CONSECUTIVE units N, N+1, N+2, ... - so a per-element
+            // assignment that is not consecutive (the conformance case uses 0, 2, 4, 6) has no
+            // spelling in a single declaration and cannot be expressed at all without splitting
+            // the array into one declaration per element and rewriting every use of it.
+            //
+            // Scoped rather than disabled, exactly as ProgramPipelineScenario scopes its
+            // storage-block rebinding cases: the defect is per-backend and the frontend
+            // mechanism these cases exist for - per-element units surviving the trip to the
+            // pipeline composite - is fully exercised on Magma.
+            bool PerElementImageUnitsAreHonoured() const { return Gl().BackendName() == "DirectVulkan"; }
+
+            // The scenarios below need image load/store at all; a driver without it should skip
+            // rather than fail.
+            bool ImagesAreUsable() const {
+                GLint maxImageUnits = 0;
+                glGetIntegerv(GL_MAX_IMAGE_UNITS, &maxImageUnits);
+                while (glGetError() != GL_NO_ERROR) {
+                }
+                return maxImageUnits >= 8;
+            }
+
+            std::vector<GLuint> m_programs;
+            std::vector<GLuint> m_pipelines;
+        };
+
+    } // namespace
+
+    // The whole conformance shape in one case: two pipelines sharing a vertex stage, four image
+    // array elements each pointed at a different unit through glProgramUniform1i, eight layers of
+    // one array texture bound one per unit, and every layer checked.
+    //
+    // Layers alternate 1.0 / 2.0 because the two fragment programs interleave their units
+    // (0,2,4,6 and 1,3,5,7) - so a defect that collapses an image array to its base element, or
+    // that loses the units on the way to the composite, does not merely dim the result: it puts
+    // the wrong VALUE in a layer and names which one.
+    TEST_F(ImageLoadStoreSsoScenario, PerElementImageUnitsReachAPipelineDraw) {
+        if (!Ready()) return;
+        if (!ImagesAreUsable()) GTEST_SKIP() << "fewer than 8 image units";
+        if (!PerElementImageUnitsAreHonoured()) {
+            GTEST_SKIP() << "non-consecutive per-element image units cannot be baked into ESSL";
+        }
+        HeadlessGL& gl = Gl();
+
+        constexpr int kWidth = 8;
+        constexpr int kHeight = 8;
+        constexpr int kLayers = 8;
+
+        const GLuint vs = MakeSeparable(GL_VERTEX_SHADER, kSsoVS);
+        const GLuint fs0 = MakeSeparable(GL_FRAGMENT_SHADER, kImageFS0);
+        const GLuint fs1 = MakeSeparable(GL_FRAGMENT_SHADER, kImageFS1);
+        if (vs == 0 || fs0 == 0 || fs1 == 0) return;
+
+        // Per ELEMENT, by name, on programs that are neither current nor attached to a bound
+        // pipeline yet - exactly the conformance call order.
+        const int units0[4] = {0, 2, 4, 6};
+        const int units1[4] = {1, 3, 5, 7};
+        for (int i = 0; i < 4; ++i) {
+            const std::string name = "g_image[" + std::to_string(i) + "]";
+            const GLint loc0 = glGetUniformLocation(fs0, name.c_str());
+            const GLint loc1 = glGetUniformLocation(fs1, name.c_str());
+            ASSERT_NE(loc0, -1) << "fs0 has no location for " << name;
+            ASSERT_NE(loc1, -1) << "fs1 has no location for " << name;
+            glProgramUniform1i(fs0, loc0, units0[i]);
+            glProgramUniform1i(fs1, loc1, units1[i]);
+        }
+        ASSERT_EQ(FirstGLError(), 0u) << "assigning image units with glProgramUniform1i errored";
+
+        const GLuint pipeline0 = MakePipeline();
+        const GLuint pipeline1 = MakePipeline();
+        glUseProgramStages(pipeline0, GL_VERTEX_SHADER_BIT, vs);
+        glUseProgramStages(pipeline0, GL_FRAGMENT_SHADER_BIT, fs0);
+        glUseProgramStages(pipeline1, GL_VERTEX_SHADER_BIT, vs);
+        glUseProgramStages(pipeline1, GL_FRAGMENT_SHADER_BIT, fs1);
+        ASSERT_EQ(FirstGLError(), 0u) << "pipeline setup errored";
+
+        GLuint texture = 0;
+        glGenTextures(1, &texture);
+        glBindTexture(GL_TEXTURE_2D_ARRAY, texture);
+        glTexParameteri(GL_TEXTURE_2D_ARRAY, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
+        glTexParameteri(GL_TEXTURE_2D_ARRAY, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
+        const std::vector<float> zeros(static_cast<size_t>(kWidth) * kHeight * kLayers * 4, 0.0f);
+        glTexImage3D(GL_TEXTURE_2D_ARRAY, 0, GL_RGBA32F, kWidth, kHeight, kLayers, 0, GL_RGBA, GL_FLOAT, zeros.data());
+        ASSERT_EQ(FirstGLError(), 0u) << "creating the RGBA32F array texture errored";
+
+        // One LAYER of the array texture per unit, which is what makes each element's unit
+        // independently observable in the readback.
+        for (int unit = 0; unit < kLayers; ++unit) {
+            glBindImageTexture(static_cast<GLuint>(unit), texture, 0, GL_FALSE, unit, GL_READ_WRITE, GL_RGBA32F);
+        }
+        ASSERT_EQ(FirstGLError(), 0u) << "glBindImageTexture errored";
+
+        GLuint vao = 0;
+        glGenVertexArrays(1, &vao);
+        glBindVertexArray(vao);
+        BindDefaultFramebuffer();
+        glViewport(0, 0, kWidth, kHeight);
+        glDisable(GL_SCISSOR_TEST);
+        glDisable(GL_DEPTH_TEST);
+        glUseProgram(0);
+
+        glBindProgramPipeline(pipeline0);
+        glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
+        glBindProgramPipeline(pipeline1);
+        glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
+        glMemoryBarrier(GL_TEXTURE_UPDATE_BARRIER_BIT | GL_SHADER_IMAGE_ACCESS_BARRIER_BIT);
+        EXPECT_EQ(FirstGLError(), 0u) << "the two pipeline draws leaked a GL error";
+
+        std::vector<float> readback(static_cast<size_t>(kWidth) * kHeight * kLayers * 4, -1.0f);
+        glBindTexture(GL_TEXTURE_2D_ARRAY, texture);
+        glGetTexImage(GL_TEXTURE_2D_ARRAY, 0, GL_RGBA, GL_FLOAT, readback.data());
+        ASSERT_EQ(FirstGLError(), 0u) << "reading the array texture back errored";
+
+        // Even layers were written through fs0's units, odd layers through fs1's.
+        for (int layer = 0; layer < kLayers; ++layer) {
+            const float expected = (layer % 2) ? 2.0f : 1.0f;
+            int offenders = 0;
+            float firstSeen = 0.0f;
+            for (int y = 0; y < kHeight; ++y) {
+                for (int x = 0; x < kWidth; ++x) {
+                    const size_t base =
+                        (static_cast<size_t>(layer) * kHeight * kWidth + static_cast<size_t>(y) * kWidth + x) * 4;
+                    for (int c = 0; c < 4; ++c) {
+                        if (readback[base + c] != expected) {
+                            if (offenders == 0) firstSeen = readback[base + c];
+                            ++offenders;
+                        }
+                    }
+                }
+            }
+            EXPECT_EQ(offenders, 0) << "layer " << layer << " (image unit " << layer << ") expected " << expected
+                                    << " but " << offenders << " components differ; first was " << firstSeen;
+        }
+
+        glBindVertexArray(0);
+        glDeleteVertexArrays(1, &vao);
+        glDeleteTextures(1, &texture);
+        gl.EndFrame();
+    }
+
+    // An image ARRAY sharing a program with another descriptor, which is the shape that makes
+    // the SPIR-V binding remap load-bearing.
+    //
+    // The remap (ProgramFactory::RemapDescriptorBindingsForVulkan) is what unifies bindings
+    // across stages and normalises every descriptor onto set 0; glslang hands it per-stage
+    // numbering that starts at 0 in EACH stage. It used to refuse any descriptor array that was
+    // not a UBO, and its only complaint was an assert that compiles out above DEBUG - so a
+    // release build carried on with the un-remapped numbering and a program holding an image
+    // array plus a second descriptor could see the two alias onto one binding, while a DEBUG
+    // build trapped on the very same program.
+    //
+    // A case with ONE descriptor cannot see any of that: with a single resource there is nothing
+    // to collide with and skipping the remap is indistinguishable from running it. Hence this
+    // one - an image array AND a uniform block in the same fragment program, with the block
+    // supplying the value that gets stored, so a mis-assigned binding shows up as the wrong
+    // colour rather than as nothing at all.
+    TEST_F(ImageLoadStoreSsoScenario, AnImageArrayAlongsideAnotherDescriptorKeepsBothBindings) {
+        if (!Ready()) return;
+        if (!ImagesAreUsable()) GTEST_SKIP() << "fewer than 8 image units";
+        if (!PerElementImageUnitsAreHonoured()) {
+            GTEST_SKIP() << "non-consecutive per-element image units cannot be baked into ESSL";
+        }
+        HeadlessGL& gl = Gl();
+
+        constexpr int kWidth = 8;
+        constexpr int kHeight = 8;
+        constexpr int kLayers = 2;
+
+        static const char* kMixedFS = R"(#version 420 core
+layout(rgba32f) uniform image2D g_image[2];
+layout(std140) uniform Value { vec4 u_value; };
+void main()
+{
+    for (int i = 0; i < g_image.length(); ++i) {
+        imageStore(g_image[i], ivec2(gl_FragCoord), u_value);
+    }
+    discard;
+}
+)";
+        const GLuint vs = MakeSeparable(GL_VERTEX_SHADER, kSsoVS);
+        const GLuint fs = MakeSeparable(GL_FRAGMENT_SHADER, kMixedFS);
+        if (vs == 0 || fs == 0) return;
+
+        // Consecutive units here on purpose: this case is about the two descriptor KINDS
+        // coexisting, not about non-consecutive assignment, which the case above covers.
+        for (int i = 0; i < 2; ++i) {
+            const std::string name = "g_image[" + std::to_string(i) + "]";
+            const GLint loc = glGetUniformLocation(fs, name.c_str());
+            ASSERT_NE(loc, -1) << "no location for " << name;
+            glProgramUniform1i(fs, loc, i);
+        }
+
+        const GLfloat value[4] = {7.0f, 7.0f, 7.0f, 7.0f};
+        GLuint ubo = 0;
+        glGenBuffers(1, &ubo);
+        glBindBuffer(GL_UNIFORM_BUFFER, ubo);
+        glBufferData(GL_UNIFORM_BUFFER, sizeof(value), value, GL_STATIC_DRAW);
+        const GLuint blockIndex = glGetUniformBlockIndex(fs, "Value");
+        ASSERT_NE(blockIndex, GL_INVALID_INDEX);
+        glUniformBlockBinding(fs, blockIndex, 0);
+        glBindBufferBase(GL_UNIFORM_BUFFER, 0, ubo);
+        glBindBuffer(GL_UNIFORM_BUFFER, 0);
+        ASSERT_EQ(FirstGLError(), 0u) << "uniform block setup errored";
+
+        const GLuint pipeline = MakePipeline();
+        glUseProgramStages(pipeline, GL_VERTEX_SHADER_BIT, vs);
+        glUseProgramStages(pipeline, GL_FRAGMENT_SHADER_BIT, fs);
+
+        GLuint texture = 0;
+        glGenTextures(1, &texture);
+        glBindTexture(GL_TEXTURE_2D_ARRAY, texture);
+        glTexParameteri(GL_TEXTURE_2D_ARRAY, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
+        glTexParameteri(GL_TEXTURE_2D_ARRAY, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
+        const std::vector<float> zeros(static_cast<size_t>(kWidth) * kHeight * kLayers * 4, 0.0f);
+        glTexImage3D(GL_TEXTURE_2D_ARRAY, 0, GL_RGBA32F, kWidth, kHeight, kLayers, 0, GL_RGBA, GL_FLOAT, zeros.data());
+        glBindImageTexture(0, texture, 0, GL_FALSE, 0, GL_READ_WRITE, GL_RGBA32F);
+        glBindImageTexture(1, texture, 0, GL_FALSE, 1, GL_READ_WRITE, GL_RGBA32F);
+        ASSERT_EQ(FirstGLError(), 0u) << "image texture setup errored";
+
+        GLuint vao = 0;
+        glGenVertexArrays(1, &vao);
+        glBindVertexArray(vao);
+        BindDefaultFramebuffer();
+        glViewport(0, 0, kWidth, kHeight);
+        glDisable(GL_SCISSOR_TEST);
+        glDisable(GL_DEPTH_TEST);
+        glUseProgram(0);
+        glBindProgramPipeline(pipeline);
+        glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
+        glMemoryBarrier(GL_TEXTURE_UPDATE_BARRIER_BIT | GL_SHADER_IMAGE_ACCESS_BARRIER_BIT);
+        EXPECT_EQ(FirstGLError(), 0u) << "the mixed-descriptor pipeline draw leaked a GL error";
+
+        std::vector<float> readback(static_cast<size_t>(kWidth) * kHeight * kLayers * 4, -1.0f);
+        glBindTexture(GL_TEXTURE_2D_ARRAY, texture);
+        glGetTexImage(GL_TEXTURE_2D_ARRAY, 0, GL_RGBA, GL_FLOAT, readback.data());
+        ASSERT_EQ(FirstGLError(), 0u) << "reading the array texture back errored";
+
+        for (int layer = 0; layer < kLayers; ++layer) {
+            int offenders = 0;
+            float firstSeen = 0.0f;
+            for (size_t i = 0; i < static_cast<size_t>(kWidth) * kHeight * 4; ++i) {
+                const size_t index = static_cast<size_t>(layer) * kHeight * kWidth * 4 + i;
+                if (readback[index] != 7.0f) {
+                    if (offenders == 0) firstSeen = readback[index];
+                    ++offenders;
+                }
+            }
+            EXPECT_EQ(offenders, 0) << "layer " << layer << ": " << offenders
+                                    << " components are not the uniform block's value; first was " << firstSeen
+                                    << " (an image-array binding and a uniform block did not both survive)";
+        }
+
+        glBindVertexArray(0);
+        glDeleteVertexArrays(1, &vao);
+        glDeleteTextures(1, &texture);
+        glDeleteBuffers(1, &ubo);
+        gl.EndFrame();
+    }
+
+    // The same units, reassigned BETWEEN draws through the same pipeline. This is the half that
+    // the composite cache key change put weight on: the composite object now survives a
+    // glProgramUniform1i, so nothing rebuilds by accident and the new unit has to be carried by
+    // the refresh path (and, on Espryt, by regenerating the program the unit is baked into).
+    TEST_F(ImageLoadStoreSsoScenario, ReassigningAnImageUnitBetweenDrawsReachesTheNextDraw) {
+        if (!Ready()) return;
+        if (!ImagesAreUsable()) GTEST_SKIP() << "fewer than 8 image units";
+        HeadlessGL& gl = Gl();
+
+        constexpr int kWidth = 8;
+        constexpr int kHeight = 8;
+        constexpr int kLayers = 2;
+
+        static const char* kSingleImageFS = R"(#version 420 core
+layout(rgba32f) uniform image2D g_image;
+void main()
+{
+    imageStore(g_image, ivec2(gl_FragCoord), vec4(3.0));
+    discard;
+}
+)";
+        const GLuint vs = MakeSeparable(GL_VERTEX_SHADER, kSsoVS);
+        const GLuint fs = MakeSeparable(GL_FRAGMENT_SHADER, kSingleImageFS);
+        if (vs == 0 || fs == 0) return;
+
+        const GLuint pipeline = MakePipeline();
+        glUseProgramStages(pipeline, GL_VERTEX_SHADER_BIT, vs);
+        glUseProgramStages(pipeline, GL_FRAGMENT_SHADER_BIT, fs);
+
+        GLuint texture = 0;
+        glGenTextures(1, &texture);
+        glBindTexture(GL_TEXTURE_2D_ARRAY, texture);
+        glTexParameteri(GL_TEXTURE_2D_ARRAY, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
+        glTexParameteri(GL_TEXTURE_2D_ARRAY, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
+        const std::vector<float> zeros(static_cast<size_t>(kWidth) * kHeight * kLayers * 4, 0.0f);
+        glTexImage3D(GL_TEXTURE_2D_ARRAY, 0, GL_RGBA32F, kWidth, kHeight, kLayers, 0, GL_RGBA, GL_FLOAT, zeros.data());
+        glBindImageTexture(0, texture, 0, GL_FALSE, 0, GL_READ_WRITE, GL_RGBA32F);
+        glBindImageTexture(1, texture, 0, GL_FALSE, 1, GL_READ_WRITE, GL_RGBA32F);
+        ASSERT_EQ(FirstGLError(), 0u) << "image texture setup errored";
+
+        GLuint vao = 0;
+        glGenVertexArrays(1, &vao);
+        glBindVertexArray(vao);
+        BindDefaultFramebuffer();
+        glViewport(0, 0, kWidth, kHeight);
+        glDisable(GL_SCISSOR_TEST);
+        glDisable(GL_DEPTH_TEST);
+        glUseProgram(0);
+        glBindProgramPipeline(pipeline);
+
+        const GLint location = glGetUniformLocation(fs, "g_image");
+        ASSERT_NE(location, -1);
+
+        // Draw one against unit 0 (layer 0)...
+        glProgramUniform1i(fs, location, 0);
+        glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
+        // ...and draw two against unit 1 (layer 1), with the composite already built and cached.
+        glProgramUniform1i(fs, location, 1);
+        glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
+        glMemoryBarrier(GL_TEXTURE_UPDATE_BARRIER_BIT | GL_SHADER_IMAGE_ACCESS_BARRIER_BIT);
+        EXPECT_EQ(FirstGLError(), 0u) << "the two pipeline draws leaked a GL error";
+
+        std::vector<float> readback(static_cast<size_t>(kWidth) * kHeight * kLayers * 4, -1.0f);
+        glBindTexture(GL_TEXTURE_2D_ARRAY, texture);
+        glGetTexImage(GL_TEXTURE_2D_ARRAY, 0, GL_RGBA, GL_FLOAT, readback.data());
+        ASSERT_EQ(FirstGLError(), 0u) << "reading the array texture back errored";
+
+        for (int layer = 0; layer < kLayers; ++layer) {
+            int offenders = 0;
+            float firstSeen = 0.0f;
+            for (size_t i = 0; i < static_cast<size_t>(kWidth) * kHeight * 4; ++i) {
+                const size_t index = static_cast<size_t>(layer) * kHeight * kWidth * 4 + i;
+                if (readback[index] != 3.0f) {
+                    if (offenders == 0) firstSeen = readback[index];
+                    ++offenders;
+                }
+            }
+            EXPECT_EQ(offenders, 0) << "layer " << layer << " was not written; " << offenders
+                                    << " components differ, first was " << firstSeen
+                                    << " (the image unit reassignment did not reach the draw)";
+        }
+
+        glBindVertexArray(0);
+        glDeleteVertexArrays(1, &vao);
+        glDeleteTextures(1, &texture);
+        gl.EndFrame();
+    }
+} // namespace MGITest

MobileGL/MG_IntegrationTest/Scenarios/OrientationScenario.cpp

@@ -55,6 +55,7 @@
 
 #include <algorithm>
 #include <cstdint>
+#include <cstring>
 #include <string>
 #include <vector>
 
@@ -101,6 +102,39 @@ void main() {
         // "every single pixel" an achievable (and therefore useful) demand.
         constexpr int kQuadrantInset = 2;
 
+        // A deliberately asymmetric sub-rect of the 128x96 surface: neither centred nor
+        // full-extent in either axis, mirroring the conformance suite's randomised
+        // sub-viewport geometry (glcShaderRenderCase.cpp:735-741). Asymmetry is the whole
+        // point - y == H - y - h is exactly the case an unconverted Y origin gets right by
+        // accident, and it is the only case the shipped code ever exercised.
+        //   correct band  = GL rows [13, 55)
+        //   mirrored band = GL rows [41, 83)   (what H-y-h produces)
+        constexpr int kSubX = 17;
+        constexpr int kSubY = 13;
+        constexpr int kSubW = 60;
+        constexpr int kSubH = 42;
+
+        Image CropRect(const Image& source, int x0, int y0, int width, int height) {
+            Image out(width, height);
+            const std::size_t rowBytes = static_cast<std::size_t>(width) * 4;
+            for (int y = 0; y < height; ++y) {
+                const std::uint8_t* sourceRow =
+                    source.Data() + (static_cast<std::size_t>(y0 + y) * source.Width() + x0) * 4;
+                std::memcpy(out.Data() + static_cast<std::size_t>(y) * rowBytes, sourceRow, rowBytes);
+            }
+            return out;
+        }
+
+        Image VFlip(const Image& source) {
+            Image out(source.Width(), source.Height());
+            const std::size_t rowBytes = static_cast<std::size_t>(source.Width()) * 4;
+            for (int y = 0; y < source.Height(); ++y) {
+                std::memcpy(out.Data() + static_cast<std::size_t>(y) * rowBytes,
+                            source.Data() + static_cast<std::size_t>(source.Height() - 1 - y) * rowBytes, rowBytes);
+            }
+            return out;
+        }
+
         struct Vertex {
             float x, y;
             float r, g, b;
@@ -377,5 +411,174 @@ void main() {
             }
         }
 
+        // ------------------------------------------------------------------ sub-rect / M-1 ----
+        //
+        // Everything above reads the FULL extent of its target, which is the one case
+        // DirectVulkan's default-framebuffer readback ever re-oriented: the remap at
+        // VulkanRenderer.cpp:2042 had no rect parameters at all, so :8278 gated it on
+        // `width == swapchainExtent.width && height == swapchainExtent.height` and fell back to a
+        // raw copy otherwise. Meanwhile the viewport (:422), the scissor (:506-546) and the
+        // ReadPixels copy offset (:8238) all used the GL bottom-origin Y verbatim as a Vulkan
+        // top-origin Y.
+        //
+        // In the conformance suite those defects CANCEL in placement - the draw lands in Vulkan
+        // rows [y, y+h) and the readback copies the same rows back - and compose into an exact
+        // vertical flip of a correct image. That is 1,759 of Magma's 1,793 non-pass cases, and
+        // image forensics over all 861 gl33 failures found 861 vertical flips and nothing else.
+        // Taken apart, they are two independent user-visible bugs, so they are tested apart:
+        // SubViewportDraw pins placement with a full-extent read, SubRectReadback pins the
+        // readback rect after a full-viewport draw, and SubViewportSubRectRoundTrip is the CTS
+        // shape where the two cancel.
+
+        // Placement: a sub-viewport draw must land in GL rows [y0, y0+h), not mirrored about the
+        // surface centre. Read back full-extent, which is the path that already worked, so a
+        // failure here can only be the viewport's Y origin.
+        TEST_F(OrientationScenario, SubViewportDrawLandsWhereGLPutsIt) {
+            BindDefaultFramebuffer();
+            ClearTo(0.0f, 0.0f, 0.0f, 1.0f);
+            glViewport(kSubX, kSubY, kSubW, kSubH);
+            DrawQuadrants();
+            glViewport(0, 0, Gl().Width(), Gl().Height());
+
+            const Image whole = ReadPixels(Gl().Width(), Gl().Height());
+            EXPECT_EQ(FirstGLError(), GLenum(GL_NO_ERROR));
+
+            const Image placed = CropRect(whole, kSubX, kSubY, kSubW, kSubH);
+            EXPECT_EQ(placed.QuadrantSignature(), kUprightSignature)
+                << "the sub-viewport draw is not upright inside its own rect";
+            ExpectUprightQuadrants(placed, "sub-viewport draw, cropped out of a full-extent read");
+
+            // Nothing may have been painted outside the viewport. This is what catches the
+            // mirrored placement: the drawn band would sit at GL rows [41, 83) instead.
+            EXPECT_TRUE(RegionIsMostly(whole, 0, Gl().Width() - 1, 0, kSubY - 2, "black", 0.0,
+                                       "below the sub-viewport"));
+            EXPECT_TRUE(RegionIsMostly(whole, 0, Gl().Width() - 1, kSubY + kSubH + 1, Gl().Height() - 1, "black",
+                                       0.0, "above the sub-viewport"));
+        }
+
+        // Readback: a full-viewport draw read back through a sub-rect must return the requested
+        // band, in GL row order. Band and orientation are asserted separately so that fixing only
+        // one of the two cannot pass this case.
+        TEST_F(OrientationScenario, SubRectReadbackReturnsTheRequestedBandUpright) {
+            BindDefaultFramebuffer();
+            ClearTo(0.0f, 0.0f, 0.0f, 1.0f);
+            DrawQuadrants();
+
+            const Image whole = ReadPixels(Gl().Width(), Gl().Height());
+            ASSERT_EQ(whole.QuadrantSignature(), kUprightSignature)
+                << "the full-extent read is already wrong, so nothing below can be trusted";
+
+            const Image sub = ReadPixelsRect(kSubX, kSubY, kSubW, kSubH);
+            EXPECT_EQ(FirstGLError(), GLenum(GL_NO_ERROR));
+            ASSERT_EQ(sub.Width(), kSubW);
+            ASSERT_EQ(sub.Height(), kSubH);
+
+            const Image requestedBand = CropRect(whole, kSubX, kSubY, kSubW, kSubH);
+            const Image mirroredBand = CropRect(whole, kSubX, Gl().Height() - kSubY - kSubH, kSubW, kSubH);
+
+            // The geometry has to be able to see both mistakes; if a future surface size made the
+            // band symmetric these assertions would be vacuous, so say so loudly instead.
+            ASSERT_FALSE(requestedBand == VFlip(requestedBand))
+                << "the chosen sub-rect is vertically symmetric - it cannot detect a row flip";
+            ASSERT_FALSE(requestedBand == mirroredBand)
+                << "the chosen sub-rect equals its mirror band - it cannot detect a wrong band";
+
+            EXPECT_FALSE(sub == VFlip(requestedBand))
+                << "ORIENTATION: the requested band came back with its rows in Vulkan (top-first) order";
+            EXPECT_FALSE(sub == mirroredBand || sub == VFlip(mirroredBand))
+                << "BAND: the read returned GL rows [H-y-h, H-y) instead of [y, y+h)";
+            EXPECT_TRUE(sub == requestedBand)
+                << "the sub-rect readback differs from the same rect of the full-extent read in "
+                << sub.ByteDiffCount(requestedBand) << " bytes";
+        }
+
+        // The exact conformance-suite shape: an asymmetric sub-viewport draw read back through the
+        // very same sub-rect. The placement and readback errors cancel, leaving an image that is
+        // correct in every pixel VALUE and vertically flipped - which is precisely the 861-case
+        // signature. One assertion, and it pins all of them.
+        TEST_F(OrientationScenario, SubViewportSubRectRoundTripIsUpright) {
+            BindDefaultFramebuffer();
+            ClearTo(0.0f, 0.0f, 0.0f, 1.0f);
+            glViewport(kSubX, kSubY, kSubW, kSubH);
+            DrawQuadrants();
+            const Image sub = ReadPixelsRect(kSubX, kSubY, kSubW, kSubH);
+            glViewport(0, 0, Gl().Width(), Gl().Height());
+
+            EXPECT_EQ(FirstGLError(), GLenum(GL_NO_ERROR));
+            EXPECT_EQ(sub.QuadrantSignature(), kUprightSignature)
+                << "sub-viewport draw + same-rect readback came back flipped - this is the shape "
+                   "behind KHR-GL33/GL40.shaders.* (861 cases each)";
+            ExpectUprightQuadrants(sub, "sub-viewport draw read back through the same sub-rect");
+        }
+
+        // The same conversion, on the other rect consumer that reads the default framebuffer.
+        // glBlitFramebuffer already converted its DESTINATION rect when the draw framebuffer was
+        // the default one (ApplyNativeBlitDefaultFramebufferTransform), but never its SOURCE rect,
+        // so a blit OUT of the default framebuffer took the mirrored band and wrote it upside
+        // down. Blitting a sub-rect and comparing against the same sub-rect of a direct read pins
+        // both halves at once.
+        TEST_F(OrientationScenario, BlitOutOfTheDefaultFramebufferKeepsBandAndOrientation) {
+            BindDefaultFramebuffer();
+            ClearTo(0.0f, 0.0f, 0.0f, 1.0f);
+            DrawQuadrants();
+            const Image whole = ReadPixels(Gl().Width(), Gl().Height());
+            ASSERT_EQ(whole.QuadrantSignature(), kUprightSignature)
+                << "the full-extent read is already wrong, so nothing below can be trusted";
+
+            BindFbo(m_offscreen);
+            ClearTo(0.0f, 0.0f, 0.0f, 1.0f);
+            glBindFramebuffer(GL_READ_FRAMEBUFFER, 0);
+            glBindFramebuffer(GL_DRAW_FRAMEBUFFER, m_offscreen.fbo);
+            glBlitFramebuffer(kSubX, kSubY, kSubX + kSubW, kSubY + kSubH, kSubX, kSubY, kSubX + kSubW,
+                              kSubY + kSubH, GL_COLOR_BUFFER_BIT, GL_NEAREST);
+            const unsigned int blitError = FirstGLError();
+            if (blitError != GL_NO_ERROR) {
+                GTEST_SKIP() << "this backend refused the default-framebuffer blit: "
+                             << GLErrorName(blitError);
+            }
+
+            glBindFramebuffer(GL_FRAMEBUFFER, m_offscreen.fbo);
+            const Image blitted = ReadPixels(m_offscreen.width, m_offscreen.height);
+            EXPECT_EQ(FirstGLError(), GLenum(GL_NO_ERROR));
+
+            const Image landed = CropRect(blitted, kSubX, kSubY, kSubW, kSubH);
+            const Image expected = CropRect(whole, kSubX, kSubY, kSubW, kSubH);
+            EXPECT_FALSE(landed == VFlip(expected))
+                << "ORIENTATION: the blitted band arrived upside down";
+            EXPECT_TRUE(landed == expected)
+                << "the blitted sub-rect differs from the same sub-rect of a direct read in "
+                << landed.ByteDiffCount(expected) << " bytes";
+        }
+
+        // Negative control. A non-default framebuffer is already self-consistent - no
+        // gl_Position.y negation, GL row 0 IS Vulkan row 0 - so none of the fixes above may touch
+        // it. If this ever starts failing, the default-FBO remap has leaked into the FBO path.
+        TEST_F(OrientationScenario, FboSubRectReadbackAndSubViewportAreUnaffected) {
+            BindFbo(m_offscreen);
+            ClearTo(0.0f, 0.0f, 0.0f, 1.0f);
+            DrawQuadrants();
+
+            const Image whole = ReadPixels(m_offscreen.width, m_offscreen.height);
+            const Image sub = ReadPixelsRect(kSubX, kSubY, kSubW, kSubH);
+            EXPECT_EQ(FirstGLError(), GLenum(GL_NO_ERROR));
+            EXPECT_TRUE(sub == CropRect(whole, kSubX, kSubY, kSubW, kSubH))
+                << "an FBO sub-rect readback differs from the same rect of its full-extent read in "
+                << sub.ByteDiffCount(CropRect(whole, kSubX, kSubY, kSubW, kSubH)) << " bytes";
+
+            BindFbo(m_offscreen);
+            ClearTo(0.0f, 0.0f, 0.0f, 1.0f);
+            glViewport(kSubX, kSubY, kSubW, kSubH);
+            DrawQuadrants();
+            glViewport(0, 0, m_offscreen.width, m_offscreen.height);
+            const Image placedWhole = ReadPixels(m_offscreen.width, m_offscreen.height);
+            EXPECT_EQ(FirstGLError(), GLenum(GL_NO_ERROR));
+            EXPECT_EQ(CropRect(placedWhole, kSubX, kSubY, kSubW, kSubH).QuadrantSignature(), kUprightSignature)
+                << "an FBO sub-viewport draw must land in GL rows [y0, y0+h) upright";
+            EXPECT_TRUE(RegionIsMostly(placedWhole, 0, m_offscreen.width - 1, 0, kSubY - 2, "black", 0.0,
+                                       "below an FBO sub-viewport"));
+            EXPECT_TRUE(RegionIsMostly(placedWhole, 0, m_offscreen.width - 1, kSubY + kSubH + 1,
+                                       m_offscreen.height - 1, "black", 0.0, "above an FBO sub-viewport"));
+        }
+
     } // namespace
 } // namespace MGITest

MobileGL/MG_IntegrationTest/Scenarios/PipelineFailureScenario.cpp

@@ -0,0 +1,197 @@
+// MobileGL - MobileGL/MG_IntegrationTest/Scenarios/PipelineFailureScenario.cpp
+// Copyright (c) 2025-2026 MobileGL-Dev
+// Licensed under the GNU Lesser General Public License v3.0:
+//   https://www.gnu.org/licenses/gpl-3.0.txt
+//   https://www.gnu.org/licenses/lgpl-3.0.txt
+// SPDX-License-Identifier: LGPL-3.0-only
+// End of Source File Header
+//
+// "The draw had no pipeline, so we bound null."
+//
+// DirectVulkan's SetupDraw called GetOrCreatePipeline - a function that DOCUMENTS a
+// VK_NULL_HANDLE return - and passed the result straight to vkCmdBindPipeline. When the
+// Adreno driver answered vkCreateGraphicsPipelines with VK_ERROR_UNKNOWN, the next
+// instruction dereferenced null inside the driver: SIGSEGV at fault addr 0x8, and that one
+// shape accounted for 9 of the 15 process deaths in the 2026-08-10 GL-CTS run
+// (KHR-GL33/GL40.shaders.struct.uniform.sampler_array_vertex, six
+// KHR-GL42.shader_image_load_store cases, one shader_storage_buffer_object case).
+//
+// It was made permanent by a second defect: PipelineFactory memoized the failure, so the
+// null was served for the rest of the process. Every later draw with the same state died
+// too, which is why a single bad program took whole CTS groups down with it.
+//
+// What this scenario pins, on both backends:
+//   1. The GL program shape the CTS crashed on (an array of structs each containing a
+//      sampler, sampled from the VERTEX stage) draws without killing the process.
+//   2. It draws AGAIN and produces the identical image. A second draw is the only thing
+//      that can tell a working pipeline apart from a poisoned cache entry: if the first
+//      creation had failed and been memoized, the second draw is where the null would be
+//      served back.
+//
+// A deterministic driver-side pipeline-creation FAILURE is not reachable from the GL API on
+// the llvmpipe/lavapipe lanes - both accept every pipeline these scenarios can describe - so
+// the guard itself is proven structurally (PipelineFactory returns before it can emplace a
+// VK_NULL_HANDLE, SetupDraw returns false before it can bind one) and this scenario holds
+// the surrounding path honest.
+
+#include <string>
+#include <vector>
+
+#include "../Harness/HeadlessGL.h"
+#include "../Harness/ScenarioFixture.h"
+
+#ifdef GLAPI
+#undef GLAPI
+#endif
+#define GL_GLEXT_PROTOTYPES
+#include <GL/gl.h>
+#include <GL/glcorearb.h>
+#undef GL_GLEXT_PROTOTYPES
+
+namespace MGITest {
+    namespace {
+
+        // Lifted from KHR-GL33.shaders.struct.uniform.sampler_array_vertex (the QPA records the
+        // source verbatim): an array of structs, each carrying an opaque sampler, sampled in the
+        // vertex stage. The fragment sibling of this case only FAILS on Magma; only the vertex one
+        // takes the process down, so the stage matters and is kept.
+        constexpr const char* kSamplerArrayVertexSource = R"(#version 330 core
+struct S {
+    float a;
+    vec3 b;
+    sampler2D c;
+};
+uniform S s[2];
+in vec2 aPos;
+out vec4 vColor;
+void main() {
+    vec2 coords = aPos * 0.5 + 0.5;
+    vColor = vec4(texture(s[1].c, coords * s[0].b.xy + s[1].b.z).rgb, s[0].a);
+    gl_Position = vec4(aPos, 0.0, 1.0);
+}
+)";
+
+        constexpr const char* kPassthroughFragmentSource = R"(#version 330 core
+in vec4 vColor;
+out vec4 oColor;
+void main() {
+    oColor = vColor;
+}
+)";
+
+        struct Vertex {
+            float x, y;
+        };
+
+        std::vector<Vertex> FullscreenTriangleStrip() {
+            return {{-1.0f, -1.0f}, {1.0f, -1.0f}, {-1.0f, 1.0f}, {1.0f, 1.0f}};
+        }
+
+        class PipelineFailureScenario : public ScenarioTest {
+        protected:
+            void SetUp() override {
+                ScenarioTest::SetUp();
+                if (!Ready()) return;
+
+                std::string error;
+                m_program = CompileProgram(kSamplerArrayVertexSource, kPassthroughFragmentSource, &error);
+                ASSERT_NE(m_program, 0u) << error;
+
+                const std::vector<Vertex> vertices = FullscreenTriangleStrip();
+                m_vertexCount = static_cast<int>(vertices.size());
+                glGenVertexArrays(1, &m_vao);
+                glBindVertexArray(m_vao);
+                glGenBuffers(1, &m_vbo);
+                glBindBuffer(GL_ARRAY_BUFFER, m_vbo);
+                glBufferData(GL_ARRAY_BUFFER, GLsizeiptr(vertices.size() * sizeof(Vertex)), vertices.data(),
+                             GL_STATIC_DRAW);
+                glEnableVertexAttribArray(0);
+                glVertexAttribPointer(0, 2, GL_FLOAT, GL_FALSE, sizeof(Vertex), reinterpret_cast<void*>(0));
+                glBindVertexArray(0);
+
+                // A solid red 2x2 texture, so the sampled colour is the same wherever the
+                // (deliberately degenerate) coordinates land.
+                const unsigned char red[] = {255, 0, 0, 255, 255, 0, 0, 255,
+                                             255, 0, 0, 255, 255, 0, 0, 255};
+                glGenTextures(1, &m_texture);
+                glActiveTexture(GL_TEXTURE0);
+                glBindTexture(GL_TEXTURE_2D, m_texture);
+                glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA8, 2, 2, 0, GL_RGBA, GL_UNSIGNED_BYTE, red);
+                glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
+                glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
+                glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
+                glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
+
+                glUseProgram(m_program);
+                const int samplerLocation = glGetUniformLocation(m_program, "s[1].c");
+                if (samplerLocation >= 0) glUniform1i(samplerLocation, 0);
+                const int alphaLocation = glGetUniformLocation(m_program, "s[0].a");
+                if (alphaLocation >= 0) glUniform1f(alphaLocation, 1.0f);
+                glUseProgram(0);
+
+                m_target = MakeColorFbo(Gl().Width(), Gl().Height());
+                ASSERT_NE(m_target.fbo, 0u) << "offscreen FBO is not framebuffer-complete";
+
+                ASSERT_EQ(FirstGLError(), GLenum(GL_NO_ERROR)) << "setup left a GL error behind";
+            }
+
+            void TearDown() override {
+                if (!Ready()) return;
+                DestroyColorFbo(m_target);
+                if (m_texture != 0) glDeleteTextures(1, &m_texture);
+                if (m_vbo != 0) glDeleteBuffers(1, &m_vbo);
+                if (m_vao != 0) glDeleteVertexArrays(1, &m_vao);
+                if (m_program != 0) glDeleteProgram(m_program);
+            }
+
+            Image DrawOnce() {
+                BindFbo(m_target);
+                ClearTo(0.0f, 0.0f, 0.0f, 1.0f);
+                glDisable(GL_DEPTH_TEST);
+                glDisable(GL_BLEND);
+                glUseProgram(m_program);
+                glActiveTexture(GL_TEXTURE0);
+                glBindTexture(GL_TEXTURE_2D, m_texture);
+                glBindVertexArray(m_vao);
+                glDrawArrays(GL_TRIANGLE_STRIP, 0, m_vertexCount);
+                glBindVertexArray(0);
+                return ReadPixels(m_target.width, m_target.height);
+            }
+
+            unsigned int m_program = 0;
+            unsigned int m_vao = 0;
+            unsigned int m_vbo = 0;
+            unsigned int m_texture = 0;
+            int m_vertexCount = 0;
+            ColorFbo m_target;
+        };
+
+        // Reaching the assertion at all is most of the point: the shipped code SIGSEGV'd inside
+        // the driver on this draw.
+        TEST_F(PipelineFailureScenario, SamplerArrayInAStructDrawsWithoutKillingTheProcess) {
+            const Image drawn = DrawOnce();
+            EXPECT_EQ(FirstGLError(), GLenum(GL_NO_ERROR));
+            EXPECT_TRUE(RegionIsMostly(drawn, 2, drawn.Width() - 3, 2, drawn.Height() - 3, "red", 0.0,
+                                       "sampler-array-in-struct draw"));
+        }
+
+        // The second draw is what a poisoned cache entry cannot survive: a memoized
+        // VK_NULL_HANDLE is served on every subsequent lookup, so a run that dies (or silently
+        // stops drawing) on the second draw and not the first is exactly the "failed pipeline was
+        // cached" defect.
+        TEST_F(PipelineFailureScenario, TheSameDrawRepeatsIdenticallyWithNoPoisonedPipelineCache) {
+            const Image first = DrawOnce();
+            ASSERT_EQ(FirstGLError(), GLenum(GL_NO_ERROR)) << "the first draw already errored";
+            Gl().EndFrame();
+            const Image second = DrawOnce();
+            EXPECT_EQ(FirstGLError(), GLenum(GL_NO_ERROR)) << "the second draw errored";
+
+            EXPECT_TRUE(RegionIsMostly(second, 2, second.Width() - 3, 2, second.Height() - 3, "red", 0.0,
+                                       "second draw"));
+            EXPECT_TRUE(second == first) << "the second draw differs from the first in "
+                                         << second.ByteDiffCount(first) << " bytes - the pipeline the second "
+                                            "draw resolved is not the one the first draw used";
+        }
+
+    } // namespace
+} // namespace MGITest

MobileGL/MG_IntegrationTest/Scenarios/PixelStoreSweepScenario.cpp

@@ -0,0 +1,249 @@
+// MobileGL - MobileGL/MG_IntegrationTest/Scenarios/PixelStoreSweepScenario.cpp
+// Copyright (c) 2025-2026 MobileGL-Dev
+// Licensed under the GNU Lesser General Public License v3.0:
+//   https://www.gnu.org/licenses/gpl-3.0.txt
+//   https://www.gnu.org/licenses/lgpl-3.0.txt
+// SPDX-License-Identifier: LGPL-3.0-only
+// End of Source File Header
+//
+// Scenario - PIXEL-STORE MODES RESTORE, and FRAMEBUFFER CHURN STAYS EXACT.
+//
+// Both cases here replay the shape of KHR-GL3x.packed_pixels.varied_rectangle, the single
+// heaviest polluter in the GL CTS: for each of 46 (pixel-store mode, value) pairs it uploads a
+// gradient into a fresh texture, attaches that texture to a FRESH framebuffer, reads it back and
+// deletes both - ~3300 texture+framebuffer pairs per test case.
+//
+// What that found: DirectGLES had no destructor for BackendFramebufferObject (nor for the
+// renderbuffer and sampler twins), so every frontend glDeleteFramebuffers leaked one driver
+// framebuffer for the process lifetime. On an Adreno 830 the CTS run walked the driver to 1.2 GB
+// of dead objects, and from that point on EVERY readback through a freshly attached framebuffer
+// came back with someone else's pixels - which is what made ~1,500 otherwise-correct cases fail
+// depending only on how much ran before them. The unit-level pin for the missing destructors is
+// MG_Test/SanityTest.cpp (DirectGLESBackendFramebuffer/Renderbuffer/Sampler); this file pins the
+// end-to-end behaviour they protect.
+//
+// The mode sweep is the second half of the same story: 46 modes are set and reset per case, so a
+// mode that fails to restore is indistinguishable from the leak in a full-batch CTS run. The
+// assertion here is RESTORATION - after every single mode is set and put back, a readback at
+// default state must be byte-identical to one taken before the sweep ever started.
+//
+// Backend-agnostic on purpose: both bugs this guards against are frontend/backend bookkeeping,
+// and DirectVulkan is the built-in control.
+
+#include <cstdint>
+#include <string>
+#include <vector>
+
+#include "../Harness/HeadlessGL.h"
+#include "../Harness/ScenarioFixture.h"
+
+#ifdef GLAPI
+#undef GLAPI
+#endif
+#define GL_GLEXT_PROTOTYPES
+#include <GL/gl.h>
+#include <GL/glcorearb.h>
+#undef GL_GLEXT_PROTOTYPES
+
+namespace MGITest {
+    namespace {
+
+        // Small enough that the table's row lengths (10, 15) and image heights are all >= the
+        // image, which is the shape the CTS uses (its gradient is 7x3).
+        constexpr int kTexSize = 8;
+        // Every buffer handed to GL is this big regardless of the image size: with row length 15,
+        // two skipped rows/pixels and alignment 8 the driver strides well past the natural image
+        // extent, and a tight buffer would be an out-of-bounds access rather than a test. (It was:
+        // the first version of this scenario passed its assertions and then segfaulted at
+        // teardown, because glReadPixels had written past a 1 KiB destination.)
+        constexpr std::size_t kScratchBytes = 64 * 1024;
+
+        // Every pixel-store mode GL 4.0 has, so a reset provably covers the whole state and not
+        // just the subset a particular test happened to touch.
+        struct PixelStoreMode {
+            GLenum name;
+            GLint defaultValue;
+        };
+        const PixelStoreMode kAllModes[] = {
+            {GL_UNPACK_SWAP_BYTES, 0},   {GL_UNPACK_LSB_FIRST, 0},   {GL_UNPACK_ROW_LENGTH, 0},
+            {GL_UNPACK_IMAGE_HEIGHT, 0}, {GL_UNPACK_SKIP_ROWS, 0},   {GL_UNPACK_SKIP_PIXELS, 0},
+            {GL_UNPACK_SKIP_IMAGES, 0},  {GL_UNPACK_ALIGNMENT, 4},   {GL_PACK_SWAP_BYTES, 0},
+            {GL_PACK_LSB_FIRST, 0},      {GL_PACK_ROW_LENGTH, 0},    {GL_PACK_IMAGE_HEIGHT, 0},
+            {GL_PACK_SKIP_ROWS, 0},      {GL_PACK_SKIP_PIXELS, 0},   {GL_PACK_SKIP_IMAGES, 0},
+            {GL_PACK_ALIGNMENT, 4},
+        };
+
+        // The CTS table verbatim (glcPackedPixelsTests.cpp VariedRectangleTest::iterate): 32
+        // common cases plus the 14 core-only ones ES has no equivalent for and MobileGL therefore
+        // honours on the CPU. IMAGE_WIDTH_1/2 and IMAGE_HEIGHT_1/2 are the CTS's 10 and 15.
+        struct SweepCase {
+            GLenum mode;
+            GLint value;
+        };
+        const SweepCase kSweep[] = {
+            {GL_UNPACK_ROW_LENGTH, 0},     {GL_UNPACK_ROW_LENGTH, 10},   {GL_UNPACK_ROW_LENGTH, 15},
+            {GL_UNPACK_SKIP_ROWS, 0},      {GL_UNPACK_SKIP_ROWS, 1},     {GL_UNPACK_SKIP_ROWS, 2},
+            {GL_UNPACK_SKIP_PIXELS, 0},    {GL_UNPACK_SKIP_PIXELS, 1},   {GL_UNPACK_SKIP_PIXELS, 2},
+            {GL_UNPACK_ALIGNMENT, 1},      {GL_UNPACK_ALIGNMENT, 2},     {GL_UNPACK_ALIGNMENT, 4},
+            {GL_UNPACK_ALIGNMENT, 8},      {GL_UNPACK_IMAGE_HEIGHT, 0},  {GL_UNPACK_IMAGE_HEIGHT, 10},
+            {GL_UNPACK_IMAGE_HEIGHT, 15},  {GL_UNPACK_SKIP_IMAGES, 0},   {GL_UNPACK_SKIP_IMAGES, 1},
+            {GL_UNPACK_SKIP_IMAGES, 2},    {GL_PACK_ROW_LENGTH, 0},      {GL_PACK_ROW_LENGTH, 10},
+            {GL_PACK_ROW_LENGTH, 15},      {GL_PACK_SKIP_ROWS, 0},       {GL_PACK_SKIP_ROWS, 1},
+            {GL_PACK_SKIP_ROWS, 2},        {GL_PACK_SKIP_PIXELS, 0},     {GL_PACK_SKIP_PIXELS, 1},
+            {GL_PACK_SKIP_PIXELS, 2},      {GL_PACK_ALIGNMENT, 1},       {GL_PACK_ALIGNMENT, 2},
+            {GL_PACK_ALIGNMENT, 4},        {GL_PACK_ALIGNMENT, 8},
+            // core-only, no ES equivalent
+            {GL_UNPACK_SWAP_BYTES, GL_FALSE}, {GL_UNPACK_SWAP_BYTES, GL_TRUE},
+            {GL_UNPACK_LSB_FIRST, GL_FALSE},  {GL_UNPACK_LSB_FIRST, GL_TRUE},
+            {GL_PACK_SWAP_BYTES, GL_FALSE},   {GL_PACK_SWAP_BYTES, GL_TRUE},
+            {GL_PACK_LSB_FIRST, GL_FALSE},    {GL_PACK_LSB_FIRST, GL_TRUE},
+            {GL_PACK_IMAGE_HEIGHT, 0},        {GL_PACK_IMAGE_HEIGHT, 10},
+            {GL_PACK_IMAGE_HEIGHT, 15},       {GL_PACK_SKIP_IMAGES, 0},
+            {GL_PACK_SKIP_IMAGES, 1},         {GL_PACK_SKIP_IMAGES, 2},
+        };
+
+        std::size_t ImageBytes(int size) { return static_cast<std::size_t>(size) * size * 4; }
+
+        // Padded to kScratchBytes so it is safe to hand to an upload running under any of the
+        // sweep's stride/skip settings.
+        std::vector<std::uint8_t> MakeGradient(int size, unsigned seed) {
+            std::vector<std::uint8_t> pixels(kScratchBytes, 0);
+            for (int y = 0; y < size; ++y) {
+                for (int x = 0; x < size; ++x) {
+                    const std::size_t base = (static_cast<std::size_t>(y) * size + x) * 4;
+                    pixels[base + 0] = static_cast<std::uint8_t>((x * 11 + seed) & 0xFF);
+                    pixels[base + 1] = static_cast<std::uint8_t>((y * 13 + seed) & 0xFF);
+                    pixels[base + 2] = static_cast<std::uint8_t>((x * y + seed) & 0xFF);
+                    pixels[base + 3] = 0xFF;
+                }
+            }
+            return pixels;
+        }
+
+        void ResetAllPixelStoreModes() {
+            for (const PixelStoreMode& mode : kAllModes) {
+                glPixelStorei(mode.name, mode.defaultValue);
+            }
+        }
+
+        // The one operation the CTS repeats: a fresh texture, a fresh framebuffer, one readback,
+        // both deleted. Returns the readback; `outStatus` carries the completeness answer so a
+        // caller can tell an incomplete framebuffer apart from wrong pixels.
+        std::vector<std::uint8_t> UploadAndReadBack(const std::vector<std::uint8_t>& source, int size,
+                                                    GLenum* outStatus) {
+            GLuint texture = 0;
+            glGenTextures(1, &texture);
+            glBindTexture(GL_TEXTURE_2D, texture);
+            glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
+            glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
+            glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA8, size, size, 0, GL_RGBA, GL_UNSIGNED_BYTE, source.data());
+
+            GLuint fbo = 0;
+            glGenFramebuffers(1, &fbo);
+            glBindFramebuffer(GL_FRAMEBUFFER, fbo);
+            glFramebufferTexture2D(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_TEXTURE_2D, texture, 0);
+            *outStatus = glCheckFramebufferStatus(GL_FRAMEBUFFER);
+
+            std::vector<std::uint8_t> read(kScratchBytes, 0);
+            if (*outStatus == GL_FRAMEBUFFER_COMPLETE) {
+                glReadPixels(0, 0, size, size, GL_RGBA, GL_UNSIGNED_BYTE, read.data());
+            }
+
+            glBindFramebuffer(GL_FRAMEBUFFER, 0);
+            glDeleteFramebuffers(1, &fbo);
+            glBindTexture(GL_TEXTURE_2D, 0);
+            glDeleteTextures(1, &texture);
+            return read;
+        }
+
+        // Index of the first differing byte within the image, or `bytes` when they agree.
+        std::size_t FirstDifference(const std::vector<std::uint8_t>& a, const std::vector<std::uint8_t>& b,
+                                    std::size_t bytes) {
+            for (std::size_t i = 0; i < bytes; ++i) {
+                if (a[i] != b[i]) return i;
+            }
+            return bytes;
+        }
+
+        class PixelStoreSweepScenario : public ScenarioTest {};
+        class FramebufferChurnScenario : public ScenarioTest {};
+
+    } // namespace
+
+    // Every mode in the CTS table is set, exercised and put back; the readback at default state
+    // afterwards must be bit-identical to the one taken before the sweep. A mode that silently
+    // fails to restore corrupts every later case in the batch, which is exactly how the CTS
+    // failures presented (the FIRST sub-case, at default state, is what failed).
+    TEST_F(PixelStoreSweepScenario, DefaultStateSurvivesTheFullModeSweep) {
+        if (!Ready()) return;
+
+        ResetAllPixelStoreModes();
+        ASSERT_EQ(FirstGLError(), 0u) << "resetting the pixel-store modes must be legal on a GL 4.0 context";
+
+        const std::vector<std::uint8_t> gradient = MakeGradient(kTexSize, 0);
+        GLenum status = 0;
+        const std::vector<std::uint8_t> baseline = UploadAndReadBack(gradient, kTexSize, &status);
+        ASSERT_EQ(status, static_cast<GLenum>(GL_FRAMEBUFFER_COMPLETE));
+        ASSERT_EQ(FirstGLError(), 0u);
+
+        const std::vector<std::uint8_t> scratchSource(kScratchBytes, 0x5A);
+
+        for (const SweepCase& sweep : kSweep) {
+            glPixelStorei(sweep.mode, sweep.value);
+            ASSERT_EQ(FirstGLError(), 0u) << "glPixelStorei(0x" << std::hex << sweep.mode << std::dec << ", "
+                                          << sweep.value << ") must be accepted";
+
+            // Exercise the mode: an upload and a readback that both run with it in force.
+            GLenum sweepStatus = 0;
+            (void)UploadAndReadBack(scratchSource, kTexSize, &sweepStatus);
+
+            ResetAllPixelStoreModes();
+
+            GLenum afterStatus = 0;
+            const std::vector<std::uint8_t> after = UploadAndReadBack(gradient, kTexSize, &afterStatus);
+            ASSERT_EQ(afterStatus, static_cast<GLenum>(GL_FRAMEBUFFER_COMPLETE));
+            const std::size_t diff = FirstDifference(baseline, after, ImageBytes(kTexSize));
+            ASSERT_EQ(diff, ImageBytes(kTexSize))
+                << "default-state readback changed after setting and resetting 0x" << std::hex << sweep.mode
+                << std::dec << " = " << sweep.value << "; first differing byte " << diff << " (baseline "
+                << static_cast<int>(baseline[diff]) << ", now " << static_cast<int>(after[diff]) << ")";
+        }
+
+        // And the modes themselves must read back as the defaults the reset asked for.
+        for (const PixelStoreMode& mode : kAllModes) {
+            GLint value = -1;
+            glGetIntegerv(mode.name, &value);
+            EXPECT_EQ(value, mode.defaultValue)
+                << "pixel-store mode 0x" << std::hex << mode.name << std::dec << " did not return to its default";
+        }
+        EXPECT_EQ(FirstGLError(), 0u);
+    }
+
+    // The leak regression. Each iteration is one complete CTS inner step, and every readback has
+    // to be exactly the gradient THIS iteration uploaded - never the previous one's. Before the
+    // missing destructors were added, the driver-side framebuffer count grew without bound here.
+    TEST_F(FramebufferChurnScenario, RepeatedFramebufferReadbackStaysExact) {
+        if (!Ready()) return;
+
+        ResetAllPixelStoreModes();
+        constexpr int kSize = 8;
+        constexpr int kIterations = 1024;
+
+        for (int i = 0; i < kIterations; ++i) {
+            // A distinct gradient per iteration: a stale attachment or a recycled driver name
+            // reads back the PREVIOUS iteration's image, which a constant fill could not tell
+            // apart from a correct read.
+            const std::vector<std::uint8_t> gradient = MakeGradient(kSize, static_cast<unsigned>(i * 7 + 1));
+            GLenum status = 0;
+            const std::vector<std::uint8_t> read = UploadAndReadBack(gradient, kSize, &status);
+            ASSERT_EQ(status, static_cast<GLenum>(GL_FRAMEBUFFER_COMPLETE)) << "iteration " << i;
+            const std::size_t diff = FirstDifference(gradient, read, ImageBytes(kSize));
+            ASSERT_EQ(diff, ImageBytes(kSize))
+                << "iteration " << i << " read back a different image than it uploaded; first differing byte "
+                << diff << " (uploaded " << static_cast<int>(gradient[diff]) << ", read "
+                << static_cast<int>(read[diff]) << ")";
+            ASSERT_EQ(FirstGLError(), 0u) << "iteration " << i;
+        }
+    }
+
+} // namespace MGITest

MobileGL/MG_IntegrationTest/Scenarios/ProgramPipelineScenario.cpp

@@ -0,0 +1,887 @@
+// MobileGL - MobileGL/MG_IntegrationTest/Scenarios/ProgramPipelineScenario.cpp
+// Copyright (c) 2025-2026 MobileGL-Dev
+// Licensed under the GNU Lesser General Public License v3.0:
+//   https://www.gnu.org/licenses/gpl-3.0.txt
+//   https://www.gnu.org/licenses/lgpl-3.0.txt
+// SPDX-License-Identifier: LGPL-3.0-only
+// End of Source File Header
+//
+// Scenario - SEPARABLE PROGRAMS DRAWN THROUGH A PROGRAM PIPELINE OBJECT.
+//
+// A pipeline object holds one program per stage and stands in for glUseProgram; MobileGL
+// flattens it into a single composite program at draw time (MG_State/GLState/Core.cpp,
+// GetProgramForDraw). Sixteen conformance cases across three different families depend on that
+// flattening and fail identically on BOTH backends - so the defect is in the shared frontend, not
+// in either backend's draw path:
+//
+//   compute_shader.{build-monolithic, build-separable, sso-case2, sso-case3, sso-compute-pipeline}
+//   shader_image_load_store.advanced-sso-{atomicCounters, simple, subroutine}
+//   shader_storage_buffer_object.{basic-syntaxSSO, basic-noBindingLayout}
+//
+// They fail with two symptoms at once - the draw renders nothing, AND the case leaves a
+// GL_INVALID_OPERATION behind that the harness reports as "forcing FAIL for subcase". Anything
+// claiming to be the root cause has to explain both.
+//
+// The cases here are the conformance shapes reduced to what fails in milliseconds, ordered from
+// the simplest pipeline that can render at all up to the compute-then-draw shape of
+// sso-compute-pipeline. Each one also asserts glGetError is clean at the end, because a case that
+// paints correctly and leaks an error still fails conformance.
+
+#include <cstdint>
+#include <string>
+#include <vector>
+
+#include "../Harness/HeadlessGL.h"
+#include "../Harness/ScenarioFixture.h"
+
+#ifdef GLAPI
+#undef GLAPI
+#endif
+#define GL_GLEXT_PROTOTYPES
+#include <GL/gl.h>
+#include <GL/glcorearb.h>
+#undef GL_GLEXT_PROTOTYPES
+
+namespace MGITest {
+    namespace {
+
+        // Separable stage sources. A separable VS must redeclare gl_PerVertex, which is exactly
+        // the kind of thing a flattening step can drop on the floor.
+        constexpr const char* kSeparableVS = R"(#version 430 core
+out gl_PerVertex { vec4 gl_Position; };
+void main()
+{
+    switch (gl_VertexID)
+    {
+      case 0: gl_Position = vec4(-1.0, -1.0, 0.0, 1.0); break;
+      case 1: gl_Position = vec4( 1.0, -1.0, 0.0, 1.0); break;
+      case 2: gl_Position = vec4(-1.0,  1.0, 0.0, 1.0); break;
+      case 3: gl_Position = vec4( 1.0,  1.0, 0.0, 1.0); break;
+    }
+}
+)";
+
+        constexpr const char* kSeparableFS = R"(#version 430 core
+out vec4 o_color;
+void main() { o_color = vec4(0.0, 1.0, 0.0, 1.0); }
+)";
+
+        // The sso-compute-pipeline shape: a compute stage writes the vertex positions the vertex
+        // stage then reads as an attribute, all from one pipeline object.
+        constexpr const char* kComputeSource = R"(#version 430 core
+layout(local_size_x = 1) in;
+layout(std430, binding = 0) buffer Positions {
+    vec4 g_position[4];
+};
+void main()
+{
+    g_position[0] = vec4(-1.0, -1.0, 0.0, 1.0);
+    g_position[1] = vec4( 1.0, -1.0, 0.0, 1.0);
+    g_position[2] = vec4(-1.0,  1.0, 0.0, 1.0);
+    g_position[3] = vec4( 1.0,  1.0, 0.0, 1.0);
+}
+)";
+
+        constexpr const char* kAttributeVS = R"(#version 430 core
+layout(location = 0) in vec4 i_position;
+out gl_PerVertex { vec4 gl_Position; };
+void main() { gl_Position = i_position; }
+)";
+
+        // Two shader storage blocks with NO layout(binding) qualifier, so the only thing that
+        // can say where they live is glShaderStorageBlockBinding - which is per-PROGRAM state.
+        constexpr const char* kStorageBlockVS = R"(#version 430 core
+out gl_PerVertex { vec4 gl_Position; };
+layout(std430) buffer Output0 { uint value0; };
+layout(std430) buffer Output1 { uint value1; };
+void main()
+{
+    value0 = 11u;
+    value1 = 22u;
+    gl_Position = vec4(0.0, 0.0, 0.0, 1.0);
+}
+)";
+
+        class ProgramPipelineScenario : public ScenarioTest {
+        protected:
+            void TearDown() override {
+                if (!Ready()) return;
+                glBindProgramPipeline(0);
+                glUseProgram(0);
+                for (GLuint p : m_programs) glDeleteProgram(p);
+                for (GLuint p : m_pipelines) glDeleteProgramPipelines(1, &p);
+                m_programs.clear();
+                m_pipelines.clear();
+            }
+
+            GLuint MakeSeparable(GLenum stage, const char* source) {
+                const GLuint program = glCreateShaderProgramv(stage, 1, &source);
+                if (program != 0) m_programs.push_back(program);
+                // Checked here rather than only at the end of the case: glCreateShaderProgramv is
+                // specified as a sequence of other entry points, so it is the most likely place
+                // for one of them to leave an error nobody consumes.
+                EXPECT_EQ(FirstGLError(), 0u)
+                    << "glCreateShaderProgramv(stage 0x" << std::hex << stage << std::dec << ") left a GL error";
+                GLint linked = GL_FALSE;
+                glGetProgramiv(program, GL_LINK_STATUS, &linked);
+                if (linked == GL_FALSE) {
+                    char log[2048] = {};
+                    glGetProgramInfoLog(program, sizeof(log) - 1, nullptr, log);
+                    ADD_FAILURE() << "glCreateShaderProgramv(stage 0x" << std::hex << stage << std::dec
+                                  << ") did not link: " << log;
+                    return 0;
+                }
+                return program;
+            }
+
+            GLuint MakePipeline() {
+                GLuint pipeline = 0;
+                glGenProgramPipelines(1, &pipeline);
+                m_pipelines.push_back(pipeline);
+                return pipeline;
+            }
+
+            std::vector<GLuint> m_programs;
+            std::vector<GLuint> m_pipelines;
+        };
+
+    } // namespace
+
+    // The root cause of the cluster, stated as the two halves it actually has.
+    //
+    // Half one: glGenProgramPipelines only reserves a name, and every pipeline command used to
+    // demand a materialized object - so the spec's own call order (stages attached BEFORE the
+    // first bind, GL 4.6 core 7.4) was rejected with GL_INVALID_OPERATION and the stages were
+    // never recorded. Half two is the trap that fix walks into: the object now appears the
+    // moment anything needs somewhere to put state, so "the object exists" stops being the
+    // right answer for glIsProgramPipeline, which the spec ties to the first BIND. A pure
+    // query must not turn a reserved name into a program pipeline either.
+    TEST_F(ProgramPipelineScenario, AReservedNameTakesStateBeforeItIsAProgramPipeline) {
+        if (!Ready()) return;
+
+        const GLuint vs = MakeSeparable(GL_VERTEX_SHADER, kSeparableVS);
+        if (vs == 0) return;
+        const GLuint pipeline = MakePipeline();
+        ASSERT_NE(pipeline, 0u);
+        EXPECT_EQ(glIsProgramPipeline(pipeline), GL_FALSE) << "a merely reserved name is not a pipeline yet";
+
+        // A query answers out of default state - and leaves the name exactly as it found it.
+        GLint validateStatus = -1;
+        glGetProgramPipelineiv(pipeline, GL_VALIDATE_STATUS, &validateStatus);
+        EXPECT_EQ(FirstGLError(), 0u) << "querying a reserved pipeline name must not be an error";
+        EXPECT_EQ(validateStatus, 0) << "a pipeline that was never validated reports VALIDATE_STATUS 0";
+        EXPECT_EQ(glIsProgramPipeline(pipeline), GL_FALSE) << "a pure query must not create the object";
+
+        // ...and glUseProgramStages RECORDS the stage on the reserved name rather than
+        // rejecting it, which is the whole defect: without this the pipeline stayed empty.
+        glUseProgramStages(pipeline, GL_VERTEX_SHADER_BIT, vs);
+        EXPECT_EQ(FirstGLError(), 0u) << "glUseProgramStages before the first bind must be accepted";
+        GLint stageProgram = 0;
+        glGetProgramPipelineiv(pipeline, GL_VERTEX_SHADER, &stageProgram);
+        EXPECT_EQ(static_cast<GLuint>(stageProgram), vs) << "the stage program was not recorded";
+        EXPECT_EQ(glIsProgramPipeline(pipeline), GL_FALSE) << "taking state is still not being bound";
+
+        // The bind is what the spec ties glIsProgramPipeline to.
+        glBindProgramPipeline(pipeline);
+        EXPECT_EQ(glIsProgramPipeline(pipeline), GL_TRUE);
+        EXPECT_EQ(FirstGLError(), 0u);
+        glBindProgramPipeline(0);
+    }
+
+    // The floor: a two-stage pipeline must paint. If this fails, nothing above it can pass, and
+    // the eight shared conformance cases have exactly one cause.
+    TEST_F(ProgramPipelineScenario, ATwoStagePipelinePaintsWhatItsStagesDescribe) {
+        if (!Ready()) return;
+        HeadlessGL& gl = Gl();
+        const int width = gl.Width();
+        const int height = gl.Height();
+
+        const GLuint vs = MakeSeparable(GL_VERTEX_SHADER, kSeparableVS);
+        const GLuint fs = MakeSeparable(GL_FRAGMENT_SHADER, kSeparableFS);
+        if (vs == 0 || fs == 0) return;
+
+        const GLuint pipeline = MakePipeline();
+        glUseProgramStages(pipeline, GL_VERTEX_SHADER_BIT, vs);
+        glUseProgramStages(pipeline, GL_FRAGMENT_SHADER_BIT, fs);
+        ASSERT_EQ(FirstGLError(), 0u) << "pipeline setup left a GL error behind";
+
+        GLuint vao = 0;
+        glGenVertexArrays(1, &vao);
+        glBindVertexArray(vao);
+
+        BindDefaultFramebuffer();
+        glViewport(0, 0, width, height);
+        glDisable(GL_SCISSOR_TEST);
+        glDisable(GL_DEPTH_TEST);
+        ClearTo(0.0f, 0.0f, 0.0f, 1.0f);
+        // No glUseProgram anywhere: the pipeline IS the program state for this draw.
+        glUseProgram(0);
+        glBindProgramPipeline(pipeline);
+        glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
+
+        const Image painted = ReadPixels(width, height);
+        EXPECT_TRUE(RegionIsMostly(painted, 2, width - 3, 2, height - 3, "green", 0.0,
+                                   "a two-stage program pipeline drawing a full-viewport strip"));
+        // The conformance harness fails a subcase on a leaked error even when the pixels are
+        // right, so this assertion is not redundant with the one above.
+        EXPECT_EQ(FirstGLError(), 0u) << "the pipeline draw leaked a GL error";
+
+        glBindVertexArray(0);
+        glDeleteVertexArrays(1, &vao);
+        gl.EndFrame();
+    }
+
+    // glActiveShaderProgram picks which stage program glUniform* addresses - and the draw has to
+    // see what was written there.
+    //
+    // The second defect of the cluster, and the one the pixels expose most directly: uniform
+    // values live on the stage program (GetProgramForUniform returns the pipeline's active
+    // program) while the draw reads the composite GetProgramForDraw builds out of the stage
+    // programs' shaders. Two objects, two sets of uniform storage; before the composite was
+    // refreshed from its stage programs this painted u_color's zero default instead of green.
+    TEST_F(ProgramPipelineScenario, UniformsGoToTheActiveShaderProgram) {
+        if (!Ready()) return;
+
+        static const char* kUniformFS = R"(#version 430 core
+uniform vec4 u_color;
+out vec4 o_color;
+void main() { o_color = u_color; }
+)";
+        HeadlessGL& gl = Gl();
+        const int width = gl.Width();
+        const int height = gl.Height();
+
+        const GLuint vs = MakeSeparable(GL_VERTEX_SHADER, kSeparableVS);
+        const GLuint fs = MakeSeparable(GL_FRAGMENT_SHADER, kUniformFS);
+        if (vs == 0 || fs == 0) return;
+
+        const GLuint pipeline = MakePipeline();
+        glUseProgramStages(pipeline, GL_VERTEX_SHADER_BIT, vs);
+        glUseProgramStages(pipeline, GL_FRAGMENT_SHADER_BIT, fs);
+        glBindProgramPipeline(pipeline);
+        glActiveShaderProgram(pipeline, fs);
+        ASSERT_EQ(FirstGLError(), 0u) << "glActiveShaderProgram left a GL error behind";
+
+        const GLint location = glGetUniformLocation(fs, "u_color");
+        ASSERT_NE(location, -1);
+        glUniform4f(location, 0.0f, 1.0f, 0.0f, 1.0f);
+        EXPECT_EQ(FirstGLError(), 0u) << "glUniform4f through the active shader program errored";
+
+        GLuint vao = 0;
+        glGenVertexArrays(1, &vao);
+        glBindVertexArray(vao);
+        BindDefaultFramebuffer();
+        glViewport(0, 0, width, height);
+        glDisable(GL_DEPTH_TEST);
+        ClearTo(0.0f, 0.0f, 0.0f, 1.0f);
+        glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
+
+        const Image painted = ReadPixels(width, height);
+        EXPECT_TRUE(RegionIsMostly(painted, 2, width - 3, 2, height - 3, "green", 0.0,
+                                   "a pipeline whose fragment uniform was set via glActiveShaderProgram"));
+        EXPECT_EQ(FirstGLError(), 0u) << "the pipeline draw leaked a GL error";
+
+        glBindVertexArray(0);
+        glDeleteVertexArrays(1, &vao);
+        gl.EndFrame();
+    }
+
+    // The sso-compute-pipeline shape: compute and non-compute stages on ONE pipeline object, the
+    // compute stage writing the buffer the vertex stage then reads.
+    //
+    // The third defect of the cluster: the flattening used to pull EVERY stage into one
+    // composite, so a single program was asked to serve both glDispatchCompute and glDrawArrays.
+    // GL keeps them apart - a pipeline's compute stage is a whole program dispatched on its own
+    // and never participates in a draw - which is why the accessors are split (GetProgramForDraw
+    // composites the graphics stages, GetProgramForDispatch hands back the compute stage
+    // program). It is also the shape that killed the process on Adreno: the composite carried a
+    // compute module into vkCreateGraphicsPipelines, and that driver SIGSEGVs rather than
+    // returning an error.
+    TEST_F(ProgramPipelineScenario, ComputeAndGraphicsStagesShareOnePipeline) {
+        if (!Ready()) return;
+        HeadlessGL& gl = Gl();
+        const int width = gl.Width();
+        const int height = gl.Height();
+
+        GLint storageBlocks = 0;
+        glGetIntegerv(GL_MAX_COMPUTE_SHADER_STORAGE_BLOCKS, &storageBlocks);
+        if (storageBlocks < 1) {
+            GTEST_SKIP() << "no compute shader storage blocks available";
+        }
+
+        const GLuint cs = MakeSeparable(GL_COMPUTE_SHADER, kComputeSource);
+        const GLuint vs = MakeSeparable(GL_VERTEX_SHADER, kAttributeVS);
+        const GLuint fs = MakeSeparable(GL_FRAGMENT_SHADER, kSeparableFS);
+        if (cs == 0 || vs == 0 || fs == 0) return;
+
+        const GLuint pipeline = MakePipeline();
+        glUseProgramStages(pipeline, GL_VERTEX_SHADER_BIT, vs);
+        glUseProgramStages(pipeline, GL_FRAGMENT_SHADER_BIT, fs);
+        glUseProgramStages(pipeline, GL_COMPUTE_SHADER_BIT, cs);
+        ASSERT_EQ(FirstGLError(), 0u) << "attaching compute and graphics stages to one pipeline errored";
+
+        GLuint buffer = 0;
+        glGenBuffers(1, &buffer);
+        glBindBuffer(GL_SHADER_STORAGE_BUFFER, buffer);
+        glBufferData(GL_SHADER_STORAGE_BUFFER, 4 * 4 * sizeof(float), nullptr, GL_DYNAMIC_DRAW);
+        glBindBuffer(GL_SHADER_STORAGE_BUFFER, 0);
+
+        GLuint vao = 0;
+        glGenVertexArrays(1, &vao);
+        glBindVertexArray(vao);
+        glBindBuffer(GL_ARRAY_BUFFER, buffer);
+        glVertexAttribPointer(0, 4, GL_FLOAT, GL_FALSE, 0, nullptr);
+        glEnableVertexAttribArray(0);
+        glBindBuffer(GL_ARRAY_BUFFER, 0);
+        glBindVertexArray(0);
+
+        BindDefaultFramebuffer();
+        glViewport(0, 0, width, height);
+        glDisable(GL_DEPTH_TEST);
+        glUseProgram(0);
+        glBindProgramPipeline(pipeline);
+        glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 0, buffer);
+        glDispatchCompute(1, 1, 1);
+        ClearTo(0.0f, 0.0f, 0.0f, 1.0f);
+        glBindVertexArray(vao);
+        glMemoryBarrier(GL_VERTEX_ATTRIB_ARRAY_BARRIER_BIT);
+        glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
+
+        const Image painted = ReadPixels(width, height);
+        EXPECT_TRUE(RegionIsMostly(painted, 2, width - 3, 2, height - 3, "green", 0.0,
+                                   "a pipeline whose compute stage wrote the vertex positions"));
+        EXPECT_EQ(FirstGLError(), 0u) << "the compute-then-draw pipeline leaked a GL error";
+
+        glBindVertexArray(0);
+        glDeleteVertexArrays(1, &vao);
+        glDeleteBuffers(1, &buffer);
+        gl.EndFrame();
+    }
+
+    // Interface-resource bindings are per-PROGRAM state, and the program a pipeline draw executes
+    // is the composite - not the stage program the application set them on.
+    //
+    // This is shader_storage_buffer_object.basic-noBindingLayout reduced: blocks declared without
+    // a layout(binding) qualifier, placed onto binding points purely by
+    // glShaderStorageBlockBinding against the stage program. The stage program records the
+    // rebinding (ProgramObject::SetShaderStorageBlockBinding, keyed by block name) and the
+    // composite is built from the stage program's SHADERS - which carry the declared bindings and
+    // know nothing of the rebinding. So the draw writes wherever the shader source said, the
+    // bound buffer ranges never see a byte, and no GL error is raised anywhere: the readback is
+    // the only thing that notices.
+    TEST_F(ProgramPipelineScenario, AStageProgramsStorageBlockBindingReachesThePipelineDraw) {
+        if (!Ready()) return;
+        HeadlessGL& gl = Gl();
+
+        GLint vertexStorageBlocks = 0;
+        glGetIntegerv(GL_MAX_VERTEX_SHADER_STORAGE_BLOCKS, &vertexStorageBlocks);
+        if (vertexStorageBlocks < 2) {
+            GTEST_SKIP() << "fewer than two vertex shader storage blocks available";
+        }
+
+        const GLuint vs = MakeSeparable(GL_VERTEX_SHADER, kStorageBlockVS);
+        if (vs == 0) return;
+
+        // Rebound to binding points the shader source never mentions, so nothing but the
+        // rebinding can put the writes where this case looks for them.
+        constexpr GLuint kBinding0 = 1;
+        constexpr GLuint kBinding1 = 5;
+        const GLuint block0 = glGetProgramResourceIndex(vs, GL_SHADER_STORAGE_BLOCK, "Output0");
+        const GLuint block1 = glGetProgramResourceIndex(vs, GL_SHADER_STORAGE_BLOCK, "Output1");
+        ASSERT_NE(block0, GL_INVALID_INDEX);
+        ASSERT_NE(block1, GL_INVALID_INDEX);
+        glShaderStorageBlockBinding(vs, block0, kBinding0);
+        glShaderStorageBlockBinding(vs, block1, kBinding1);
+        ASSERT_EQ(FirstGLError(), 0u) << "glShaderStorageBlockBinding on a separable program errored";
+
+        GLint offsetAlignment = 256;
+        glGetIntegerv(GL_SHADER_STORAGE_BUFFER_OFFSET_ALIGNMENT, &offsetAlignment);
+        if (offsetAlignment <= 0) offsetAlignment = 256;
+        const GLsizeiptr secondOffset = offsetAlignment;
+
+        GLuint buffer = 0;
+        glGenBuffers(1, &buffer);
+        glBindBuffer(GL_SHADER_STORAGE_BUFFER, buffer);
+        const std::vector<GLuint> zeros(static_cast<std::size_t>(secondOffset) / sizeof(GLuint) + 4, 0u);
+        glBufferData(GL_SHADER_STORAGE_BUFFER, static_cast<GLsizeiptr>(zeros.size() * sizeof(GLuint)), zeros.data(),
+                     GL_DYNAMIC_DRAW);
+        glBindBufferRange(GL_SHADER_STORAGE_BUFFER, kBinding0, buffer, 0, sizeof(GLuint));
+        glBindBufferRange(GL_SHADER_STORAGE_BUFFER, kBinding1, buffer, secondOffset, sizeof(GLuint));
+        glBindBuffer(GL_SHADER_STORAGE_BUFFER, 0);
+
+        const GLuint pipeline = MakePipeline();
+        glUseProgramStages(pipeline, GL_VERTEX_SHADER_BIT, vs);
+
+        GLuint vao = 0;
+        glGenVertexArrays(1, &vao);
+        glBindVertexArray(vao);
+        BindDefaultFramebuffer();
+        // The whole point is the buffer writes, so the rasterizer is not involved - which is
+        // also what keeps a vertex-only pipeline (no fragment stage) legal here.
+        glEnable(GL_RASTERIZER_DISCARD);
+        glUseProgram(0);
+        glBindProgramPipeline(pipeline);
+        glDrawArrays(GL_POINTS, 0, 1);
+        glDisable(GL_RASTERIZER_DISCARD);
+        EXPECT_EQ(FirstGLError(), 0u) << "the storage-block pipeline draw leaked a GL error";
+
+        glMemoryBarrier(GL_BUFFER_UPDATE_BARRIER_BIT | GL_SHADER_STORAGE_BARRIER_BIT);
+        glBindBuffer(GL_SHADER_STORAGE_BUFFER, buffer);
+        GLuint readback0 = 0;
+        GLuint readback1 = 0;
+        glGetBufferSubData(GL_SHADER_STORAGE_BUFFER, 0, sizeof(readback0), &readback0);
+        glGetBufferSubData(GL_SHADER_STORAGE_BUFFER, secondOffset, sizeof(readback1), &readback1);
+        glBindBuffer(GL_SHADER_STORAGE_BUFFER, 0);
+
+        EXPECT_EQ(readback0, 11u) << "Output0 did not reach the binding glShaderStorageBlockBinding gave it";
+        EXPECT_EQ(readback1, 22u) << "Output1 did not reach the binding glShaderStorageBlockBinding gave it";
+        EXPECT_EQ(FirstGLError(), 0u);
+
+        glBindVertexArray(0);
+        glDeleteVertexArrays(1, &vao);
+        glDeleteBuffers(1, &buffer);
+        gl.EndFrame();
+    }
+
+    // CONTROL for the case above, and the thing that says whether a storage-block failure is
+    // about pipelines at all: the same shader, the same rebinding, in an ordinary two-stage
+    // monolithic program run through glUseProgram. If this one fails too then the composite is
+    // innocent and the defect is in how the backend replays a rebinding.
+    //
+    // Two stages on purpose. Handing glUseProgram a vertex-ONLY program would confound the
+    // experiment - a program with no fragment stage is a thing some backends cannot build at
+    // all, so its failure would say nothing about block bindings.
+    //
+    // Runs on both backends. glShaderStorageBlockBinding is a GL 4.3 entry point with no ES
+    // equivalent - ES fixes a storage block's binding at link from its layout(binding=)
+    // qualifier - so Espryt honours a rebinding by writing the effective binding into the ESSL
+    // it generates (the Binding decoration is rewritten before SPIRV-Cross emits, and the draw
+    // path rebuilds a program whose override set has moved).
+    TEST_F(ProgramPipelineScenario, AStorageBlockRebindingHoldsWithoutAPipeline) {
+        if (!Ready()) return;
+        HeadlessGL& gl = Gl();
+
+        GLint vertexStorageBlocks = 0;
+        glGetIntegerv(GL_MAX_VERTEX_SHADER_STORAGE_BLOCKS, &vertexStorageBlocks);
+        if (vertexStorageBlocks < 2) {
+            GTEST_SKIP() << "fewer than two vertex shader storage blocks available";
+        }
+
+        static const char* kMonolithicVS = R"(#version 430 core
+layout(std430) buffer Output0 { uint value0; };
+layout(std430) buffer Output1 { uint value1; };
+void main()
+{
+    value0 = 11u;
+    value1 = 22u;
+    gl_Position = vec4(0.0, 0.0, 0.0, 1.0);
+}
+)";
+        static const char* kMonolithicFS = R"(#version 430 core
+out vec4 o_color;
+void main() { o_color = vec4(1.0); }
+)";
+        std::string compileError;
+        const GLuint vs = CompileProgram(kMonolithicVS, kMonolithicFS, &compileError);
+        ASSERT_NE(vs, 0u) << compileError;
+        m_programs.push_back(vs);
+
+        constexpr GLuint kBinding0 = 1;
+        constexpr GLuint kBinding1 = 5;
+        const GLuint block0 = glGetProgramResourceIndex(vs, GL_SHADER_STORAGE_BLOCK, "Output0");
+        const GLuint block1 = glGetProgramResourceIndex(vs, GL_SHADER_STORAGE_BLOCK, "Output1");
+        ASSERT_NE(block0, GL_INVALID_INDEX);
+        ASSERT_NE(block1, GL_INVALID_INDEX);
+        glShaderStorageBlockBinding(vs, block0, kBinding0);
+        glShaderStorageBlockBinding(vs, block1, kBinding1);
+        ASSERT_EQ(FirstGLError(), 0u);
+
+        GLint offsetAlignment = 256;
+        glGetIntegerv(GL_SHADER_STORAGE_BUFFER_OFFSET_ALIGNMENT, &offsetAlignment);
+        if (offsetAlignment <= 0) offsetAlignment = 256;
+        const GLsizeiptr secondOffset = offsetAlignment;
+
+        GLuint buffer = 0;
+        glGenBuffers(1, &buffer);
+        glBindBuffer(GL_SHADER_STORAGE_BUFFER, buffer);
+        const std::vector<GLuint> zeros(static_cast<std::size_t>(secondOffset) / sizeof(GLuint) + 4, 0u);
+        glBufferData(GL_SHADER_STORAGE_BUFFER, static_cast<GLsizeiptr>(zeros.size() * sizeof(GLuint)), zeros.data(),
+                     GL_DYNAMIC_DRAW);
+        glBindBufferRange(GL_SHADER_STORAGE_BUFFER, kBinding0, buffer, 0, sizeof(GLuint));
+        glBindBufferRange(GL_SHADER_STORAGE_BUFFER, kBinding1, buffer, secondOffset, sizeof(GLuint));
+        glBindBuffer(GL_SHADER_STORAGE_BUFFER, 0);
+
+        GLuint vao = 0;
+        glGenVertexArrays(1, &vao);
+        glBindVertexArray(vao);
+        BindDefaultFramebuffer();
+        glEnable(GL_RASTERIZER_DISCARD);
+        // No pipeline anywhere: a separable program is still a perfectly good current program.
+        glBindProgramPipeline(0);
+        glUseProgram(vs);
+        glDrawArrays(GL_POINTS, 0, 1);
+        glDisable(GL_RASTERIZER_DISCARD);
+        EXPECT_EQ(FirstGLError(), 0u) << "the monolithic storage-block draw leaked a GL error";
+
+        glMemoryBarrier(GL_BUFFER_UPDATE_BARRIER_BIT | GL_SHADER_STORAGE_BARRIER_BIT);
+        glBindBuffer(GL_SHADER_STORAGE_BUFFER, buffer);
+        GLuint readback0 = 0;
+        GLuint readback1 = 0;
+        glGetBufferSubData(GL_SHADER_STORAGE_BUFFER, 0, sizeof(readback0), &readback0);
+        glGetBufferSubData(GL_SHADER_STORAGE_BUFFER, secondOffset, sizeof(readback1), &readback1);
+        glBindBuffer(GL_SHADER_STORAGE_BUFFER, 0);
+
+        EXPECT_EQ(readback0, 11u) << "Output0 missed its rebinding with no pipeline involved";
+        EXPECT_EQ(readback1, 22u) << "Output1 missed its rebinding with no pipeline involved";
+
+        glUseProgram(0);
+        glBindVertexArray(0);
+        glDeleteVertexArrays(1, &vao);
+        glDeleteBuffers(1, &buffer);
+        gl.EndFrame();
+    }
+
+    // The same defect through the other block flavour: glUniformBlockBinding is also per-program
+    // state, recorded on the stage program by GL block index, and also never reaches the
+    // composite the draw actually runs.
+    TEST_F(ProgramPipelineScenario, AStageProgramsUniformBlockBindingReachesThePipelineDraw) {
+        if (!Ready()) return;
+        HeadlessGL& gl = Gl();
+        const int width = gl.Width();
+        const int height = gl.Height();
+
+        static const char* kUniformBlockFS = R"(#version 430 core
+layout(std140) uniform Colour { vec4 u_colour; };
+out vec4 o_color;
+void main() { o_color = u_colour; }
+)";
+        const GLuint vs = MakeSeparable(GL_VERTEX_SHADER, kSeparableVS);
+        const GLuint fs = MakeSeparable(GL_FRAGMENT_SHADER, kUniformBlockFS);
+        if (vs == 0 || fs == 0) return;
+
+        constexpr GLuint kBinding = 3; // not the default 0 the declaration implies
+        const GLuint blockIndex = glGetUniformBlockIndex(fs, "Colour");
+        ASSERT_NE(blockIndex, GL_INVALID_INDEX);
+        glUniformBlockBinding(fs, blockIndex, kBinding);
+        ASSERT_EQ(FirstGLError(), 0u) << "glUniformBlockBinding on a separable program errored";
+
+        const GLfloat green[4] = {0.0f, 1.0f, 0.0f, 1.0f};
+        GLuint buffer = 0;
+        glGenBuffers(1, &buffer);
+        glBindBuffer(GL_UNIFORM_BUFFER, buffer);
+        glBufferData(GL_UNIFORM_BUFFER, sizeof(green), green, GL_STATIC_DRAW);
+        glBindBufferBase(GL_UNIFORM_BUFFER, kBinding, buffer);
+        glBindBuffer(GL_UNIFORM_BUFFER, 0);
+
+        const GLuint pipeline = MakePipeline();
+        glUseProgramStages(pipeline, GL_VERTEX_SHADER_BIT, vs);
+        glUseProgramStages(pipeline, GL_FRAGMENT_SHADER_BIT, fs);
+
+        GLuint vao = 0;
+        glGenVertexArrays(1, &vao);
+        glBindVertexArray(vao);
+        BindDefaultFramebuffer();
+        glViewport(0, 0, width, height);
+        glDisable(GL_DEPTH_TEST);
+        ClearTo(0.0f, 0.0f, 0.0f, 1.0f);
+        glUseProgram(0);
+        glBindProgramPipeline(pipeline);
+        glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
+
+        const Image painted = ReadPixels(width, height);
+        EXPECT_TRUE(RegionIsMostly(painted, 2, width - 3, 2, height - 3, "green", 0.0,
+                                   "a pipeline whose fragment uniform block was rebound to binding 3"));
+        EXPECT_EQ(FirstGLError(), 0u) << "the uniform-block pipeline draw leaked a GL error";
+
+        glBindVertexArray(0);
+        glDeleteVertexArrays(1, &vao);
+        glDeleteBuffers(1, &buffer);
+        gl.EndFrame();
+    }
+
+    // The shared-header idiom, drawn: BOTH stages declare `u_mvp` because they both include the
+    // same header, and only the VERTEX program is ever written to.
+    //
+    // The composite has one slot for `u_mvp`, and mirroring every active uniform of every stage
+    // in stage order meant the fragment program's untouched zero matrix landed last and won.
+    // The vertex stage then transformed every vertex by a zero matrix and the frame came out
+    // empty - from an application that had done nothing wrong, with no GL error anywhere to say
+    // so. Only uniforms a stage has actually been written to are mirrored now.
+    TEST_F(ProgramPipelineScenario, AUniformDeclaredInTwoStagesKeepsTheValueTheWrittenStageHolds) {
+        if (!Ready()) return;
+        HeadlessGL& gl = Gl();
+        const int width = gl.Width();
+        const int height = gl.Height();
+
+        // The same declaration in both stages, exactly as a shared header produces it. The
+        // fragment stage does not even USE it for its output - declaring it is enough.
+        static const char* kSharedMvpVS = R"(#version 430 core
+out gl_PerVertex { vec4 gl_Position; };
+uniform mat4 u_mvp;
+void main()
+{
+    vec4 corner = vec4(0.0, 0.0, 0.0, 1.0);
+    switch (gl_VertexID)
+    {
+      case 0: corner = vec4(-1.0, -1.0, 0.0, 1.0); break;
+      case 1: corner = vec4( 1.0, -1.0, 0.0, 1.0); break;
+      case 2: corner = vec4(-1.0,  1.0, 0.0, 1.0); break;
+      case 3: corner = vec4( 1.0,  1.0, 0.0, 1.0); break;
+    }
+    gl_Position = u_mvp * corner;
+}
+)";
+        static const char* kSharedMvpFS = R"(#version 430 core
+uniform mat4 u_mvp;
+out vec4 o_color;
+void main() { o_color = vec4(0.0, 1.0, 0.0, u_mvp[3][3]); }
+)";
+
+        const GLuint vs = MakeSeparable(GL_VERTEX_SHADER, kSharedMvpVS);
+        const GLuint fs = MakeSeparable(GL_FRAGMENT_SHADER, kSharedMvpFS);
+        if (vs == 0 || fs == 0) return;
+
+        const GLuint pipeline = MakePipeline();
+        glUseProgramStages(pipeline, GL_VERTEX_SHADER_BIT, vs);
+        glUseProgramStages(pipeline, GL_FRAGMENT_SHADER_BIT, fs);
+        glBindProgramPipeline(pipeline);
+
+        // Written through the VERTEX program only - which is the whole point. The fragment
+        // program's `u_mvp` is left at GL's zero default and must not win the composite's slot.
+        glActiveShaderProgram(pipeline, vs);
+        const GLint location = glGetUniformLocation(vs, "u_mvp");
+        ASSERT_NE(location, -1);
+        const GLfloat identity[16] = {1.0f, 0.0f, 0.0f, 0.0f, 0.0f, 1.0f, 0.0f, 0.0f,
+                                      0.0f, 0.0f, 1.0f, 0.0f, 0.0f, 0.0f, 0.0f, 1.0f};
+        glUniformMatrix4fv(location, 1, GL_FALSE, identity);
+        ASSERT_EQ(FirstGLError(), 0u) << "glUniformMatrix4fv through the active shader program errored";
+
+        GLuint vao = 0;
+        glGenVertexArrays(1, &vao);
+        glBindVertexArray(vao);
+        BindDefaultFramebuffer();
+        glViewport(0, 0, width, height);
+        glDisable(GL_SCISSOR_TEST);
+        glDisable(GL_DEPTH_TEST);
+        ClearTo(0.0f, 0.0f, 0.0f, 1.0f);
+        glUseProgram(0);
+        glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
+
+        // A zero matrix collapses all four corners onto the origin and paints nothing at all, so
+        // "green over the whole viewport" IS the assertion that the written matrix was the one
+        // the draw used. (The fragment stage reads u_mvp too - into the alpha channel - purely
+        // so the optimizer cannot delete its declaration and make the case vacuous.)
+        const Image painted = ReadPixels(width, height);
+        EXPECT_TRUE(RegionIsMostly(painted, 2, width - 3, 2, height - 3, "green", 0.0,
+                                   "a pipeline whose u_mvp is declared in both stages and written in one"));
+        EXPECT_EQ(FirstGLError(), 0u) << "the shared-uniform pipeline draw leaked a GL error";
+
+        glBindVertexArray(0);
+        glDeleteVertexArrays(1, &vao);
+        gl.EndFrame();
+    }
+
+    // Rebinding a uniform block AFTER the pipeline has already drawn once.
+    //
+    // This is the shape the composite cache key change put weight on. The composite used to be
+    // thrown away and relinked whenever glUniformBlockBinding moved a stage program's backend
+    // state version, so the second draw here got a brand-new composite that happened to pick the
+    // new binding up on the way. Now the composite SURVIVES the rebinding, which means the only
+    // thing that can carry the new binding to the draw is the refresh path - so this case is
+    // what says that path is really doing the work.
+    TEST_F(ProgramPipelineScenario, RebindingAUniformBlockBetweenDrawsReachesTheNextDraw) {
+        if (!Ready()) return;
+        HeadlessGL& gl = Gl();
+        const int width = gl.Width();
+        const int height = gl.Height();
+
+        static const char* kUniformBlockFS = R"(#version 430 core
+layout(std140) uniform Colour { vec4 u_colour; };
+out vec4 o_color;
+void main() { o_color = u_colour; }
+)";
+        const GLuint vs = MakeSeparable(GL_VERTEX_SHADER, kSeparableVS);
+        const GLuint fs = MakeSeparable(GL_FRAGMENT_SHADER, kUniformBlockFS);
+        if (vs == 0 || fs == 0) return;
+
+        // Two buffers on two different binding points, holding two different colours.
+        const GLfloat red[4] = {1.0f, 0.0f, 0.0f, 1.0f};
+        const GLfloat green[4] = {0.0f, 1.0f, 0.0f, 1.0f};
+        constexpr GLuint kFirstBinding = 2;
+        constexpr GLuint kSecondBinding = 5;
+        GLuint buffers[2] = {0, 0};
+        glGenBuffers(2, buffers);
+        glBindBuffer(GL_UNIFORM_BUFFER, buffers[0]);
+        glBufferData(GL_UNIFORM_BUFFER, sizeof(red), red, GL_STATIC_DRAW);
+        glBindBufferBase(GL_UNIFORM_BUFFER, kFirstBinding, buffers[0]);
+        glBindBuffer(GL_UNIFORM_BUFFER, buffers[1]);
+        glBufferData(GL_UNIFORM_BUFFER, sizeof(green), green, GL_STATIC_DRAW);
+        glBindBufferBase(GL_UNIFORM_BUFFER, kSecondBinding, buffers[1]);
+        glBindBuffer(GL_UNIFORM_BUFFER, 0);
+
+        const GLuint blockIndex = glGetUniformBlockIndex(fs, "Colour");
+        ASSERT_NE(blockIndex, GL_INVALID_INDEX);
+        glUniformBlockBinding(fs, blockIndex, kFirstBinding);
+
+        const GLuint pipeline = MakePipeline();
+        glUseProgramStages(pipeline, GL_VERTEX_SHADER_BIT, vs);
+        glUseProgramStages(pipeline, GL_FRAGMENT_SHADER_BIT, fs);
+
+        GLuint vao = 0;
+        glGenVertexArrays(1, &vao);
+        glBindVertexArray(vao);
+        BindDefaultFramebuffer();
+        glViewport(0, 0, width, height);
+        glDisable(GL_SCISSOR_TEST);
+        glDisable(GL_DEPTH_TEST);
+        glUseProgram(0);
+        glBindProgramPipeline(pipeline);
+
+        // Draw one: the composite is built here, against binding 2.
+        ClearTo(0.0f, 0.0f, 0.0f, 1.0f);
+        glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
+        const Image first = ReadPixels(width, height);
+        EXPECT_TRUE(RegionIsMostly(first, 2, width - 3, 2, height - 3, "red", 0.0,
+                                   "the first pipeline draw, with Colour on binding 2"));
+        ASSERT_EQ(FirstGLError(), 0u) << "the first uniform-block pipeline draw leaked a GL error";
+
+        // Move the block to the other binding point, with the composite already built and cached.
+        glUniformBlockBinding(fs, blockIndex, kSecondBinding);
+        ASSERT_EQ(FirstGLError(), 0u) << "rebinding a uniform block between draws errored";
+
+        // Draw two must read the OTHER buffer.
+        ClearTo(0.0f, 0.0f, 0.0f, 1.0f);
+        glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
+        const Image second = ReadPixels(width, height);
+        EXPECT_TRUE(RegionIsMostly(second, 2, width - 3, 2, height - 3, "green", 0.0,
+                                   "the second pipeline draw, after Colour was rebound to binding 5"));
+        EXPECT_EQ(FirstGLError(), 0u) << "the rebound uniform-block pipeline draw leaked a GL error";
+
+        glBindVertexArray(0);
+        glDeleteVertexArrays(1, &vao);
+        glDeleteBuffers(2, buffers);
+        gl.EndFrame();
+    }
+
+    // The sampler-unit half of the same question, in a loop: set a unit, draw, repeat. This is
+    // the shape KHR-GL42.shader_image_load_store.advanced-sso-* and the compute_shader SSO cases
+    // run, and the one that used to relink the composite on every single iteration. The pixels
+    // pin what the loop must PRODUCE; the composite-identity assertion that pins what it must
+    // COST lives in the MG_Test unit suite, where the object itself is reachable.
+    TEST_F(ProgramPipelineScenario, ASamplerUnitRewrittenBetweenDrawsKeepsPaintingTheRightTexture) {
+        if (!Ready()) return;
+        HeadlessGL& gl = Gl();
+        const int width = gl.Width();
+        const int height = gl.Height();
+
+        static const char* kSamplerFS = R"(#version 430 core
+uniform sampler2D u_tex;
+out vec4 o_color;
+void main() { o_color = texture(u_tex, vec2(0.5)); }
+)";
+        const GLuint vs = MakeSeparable(GL_VERTEX_SHADER, kSeparableVS);
+        const GLuint fs = MakeSeparable(GL_FRAGMENT_SHADER, kSamplerFS);
+        if (vs == 0 || fs == 0) return;
+
+        // One texture per unit, each a different solid colour, so the pixels say which unit the
+        // draw actually sampled.
+        constexpr int kUnits = 4;
+        const GLubyte colours[kUnits][4] = {{255, 0, 0, 255}, {0, 255, 0, 255}, {0, 0, 255, 255}, {255, 255, 0, 255}};
+        const char* names[kUnits] = {"red", "green", "blue", "yellow"};
+        GLuint textures[kUnits] = {};
+        glGenTextures(kUnits, textures);
+        for (int unit = 0; unit < kUnits; ++unit) {
+            glActiveTexture(GL_TEXTURE0 + unit);
+            glBindTexture(GL_TEXTURE_2D, textures[unit]);
+            glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA8, 1, 1, 0, GL_RGBA, GL_UNSIGNED_BYTE, colours[unit]);
+            glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
+            glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
+        }
+        glActiveTexture(GL_TEXTURE0);
+        ASSERT_EQ(FirstGLError(), 0u) << "texture setup left a GL error behind";
+
+        const GLuint pipeline = MakePipeline();
+        glUseProgramStages(pipeline, GL_VERTEX_SHADER_BIT, vs);
+        glUseProgramStages(pipeline, GL_FRAGMENT_SHADER_BIT, fs);
+        glBindProgramPipeline(pipeline);
+        glActiveShaderProgram(pipeline, fs);
+        const GLint sampler = glGetUniformLocation(fs, "u_tex");
+        ASSERT_NE(sampler, -1);
+
+        GLuint vao = 0;
+        glGenVertexArrays(1, &vao);
+        glBindVertexArray(vao);
+        BindDefaultFramebuffer();
+        glViewport(0, 0, width, height);
+        glDisable(GL_SCISSOR_TEST);
+        glDisable(GL_DEPTH_TEST);
+        glUseProgram(0);
+
+        for (int unit = 0; unit < kUnits; ++unit) {
+            glUniform1i(sampler, unit);
+            ClearTo(0.0f, 0.0f, 0.0f, 1.0f);
+            glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
+            const Image painted = ReadPixels(width, height);
+            EXPECT_TRUE(RegionIsMostly(painted, 2, width - 3, 2, height - 3, names[unit], 0.0,
+                                       "a pipeline draw after its sampler was pointed at another unit"))
+                << "unit " << unit;
+            EXPECT_EQ(FirstGLError(), 0u) << "the sampler-rewrite pipeline draw leaked a GL error at unit " << unit;
+        }
+
+        glBindVertexArray(0);
+        glDeleteVertexArrays(1, &vao);
+        glDeleteTextures(kUnits, textures);
+        gl.EndFrame();
+    }
+
+    // build-separable / build-monolithic reduce to this: a separable program and a monolithic one
+    // must both be usable, and switching between pipeline and glUseProgram must leave no error.
+    TEST_F(ProgramPipelineScenario, SwitchingBetweenAPipelineAndAMonolithicProgramLeavesNoError) {
+        if (!Ready()) return;
+        HeadlessGL& gl = Gl();
+        const int width = gl.Width();
+        const int height = gl.Height();
+
+        const GLuint vs = MakeSeparable(GL_VERTEX_SHADER, kSeparableVS);
+        const GLuint fs = MakeSeparable(GL_FRAGMENT_SHADER, kSeparableFS);
+        if (vs == 0 || fs == 0) return;
+        const GLuint pipeline = MakePipeline();
+        glUseProgramStages(pipeline, GL_VERTEX_SHADER_BIT | GL_FRAGMENT_SHADER_BIT, 0);
+        glUseProgramStages(pipeline, GL_VERTEX_SHADER_BIT, vs);
+        glUseProgramStages(pipeline, GL_FRAGMENT_SHADER_BIT, fs);
+
+        std::string error;
+        const unsigned int monolithic = CompileProgram(
+            "#version 330 core\nin vec2 aPos;\nvoid main(){ gl_Position = vec4(aPos,0.0,1.0); }\n",
+            "#version 330 core\nout vec4 o;\nvoid main(){ o = vec4(1.0,0.0,0.0,1.0); }\n", &error);
+        ASSERT_NE(monolithic, 0u) << error;
+        m_programs.push_back(monolithic);
+
+        GLuint vao = 0;
+        glGenVertexArrays(1, &vao);
+        glBindVertexArray(vao);
+        BindDefaultFramebuffer();
+        glViewport(0, 0, width, height);
+        glDisable(GL_DEPTH_TEST);
+
+        // GL 4.6 core 7.3: while a program is current, it takes precedence over the pipeline.
+        ClearTo(0.0f, 0.0f, 0.0f, 1.0f);
+        glBindProgramPipeline(pipeline);
+        glUseProgram(monolithic);
+        glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
+        EXPECT_EQ(FirstGLError(), 0u) << "drawing with a current program while a pipeline is bound errored";
+
+        // ... and once it is not current, the pipeline takes over again.
+        ClearTo(0.0f, 0.0f, 0.0f, 1.0f);
+        glUseProgram(0);
+        glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
+        const Image painted = ReadPixels(width, height);
+        EXPECT_TRUE(RegionIsMostly(painted, 2, width - 3, 2, height - 3, "green", 0.0,
+                                   "the pipeline after the current program was unbound"));
+        EXPECT_EQ(FirstGLError(), 0u) << "switching back to the pipeline leaked a GL error";
+
+        glBindVertexArray(0);
+        glDeleteVertexArrays(1, &vao);
+        gl.EndFrame();
+    }
+} // namespace MGITest

MobileGL/MG_IntegrationTest/Scenarios/SsboArrayLengthScenario.cpp

@@ -0,0 +1,215 @@
+// MobileGL - MobileGL/MG_IntegrationTest/Scenarios/SsboArrayLengthScenario.cpp
+// Copyright (c) 2025-2026 MobileGL-Dev
+// Licensed under the GNU Lesser General Public License v3.0:
+//   https://www.gnu.org/licenses/gpl-3.0.txt
+//   https://www.gnu.org/licenses/lgpl-3.0.txt
+// SPDX-License-Identifier: LGPL-3.0-only
+// End of Source File Header
+//
+// Scenario - length() ON AN SSBO's UNSIZED ARRAY.
+//
+// GLSL's `arr.length()` on the trailing runtime array of a shader storage block is not a compile
+// time constant: it is (bound range - the array's byte offset inside the block) / array stride,
+// evaluated against whatever the descriptor actually covers. Three separate pieces of MobileGL
+// have to agree for that to come out right - the byte offsets the block layout was compiled with,
+// the buffer the frontend binding resolves to, and the offset/size a glBindBufferRange asked for -
+// and a defect in any one of them shows up only as a wrong integer, never as an error.
+//
+// KHR-GL43.shader_storage_buffer_object.advanced-unsizedArrayLength-* (28 Magma failures, all 28
+// passing on Espryt) reports exactly that: lengths too large by roughly the size of the members
+// preceding the array. The cases here are the same shape, reduced to what can be asserted in one
+// dispatch: a block with no preamble, a block with one, a two-element ARRAY OF BLOCKS (which
+// consumes two consecutive bindings and is where the conformance failures concentrate), and the
+// two glBindBufferRange forms.
+//
+// Every length is written into one output SSBO and read back, so a failure names the block and
+// prints the number the shader saw.
+
+#include <string>
+#include <vector>
+
+#include "../Harness/HeadlessGL.h"
+#include "../Harness/ScenarioFixture.h"
+
+#ifdef GLAPI
+#undef GLAPI
+#endif
+#define GL_GLEXT_PROTOTYPES
+#include <GL/gl.h>
+#include <GL/glcorearb.h>
+#undef GL_GLEXT_PROTOTYPES
+
+namespace MGITest {
+    namespace {
+
+        // Bindings 0..3 are inputs (2 and 3 are the block array), 4 is the output.
+        constexpr const char* kComputeSource = R"(#version 430 core
+layout(local_size_x = 1) in;
+layout(std430, binding = 0) readonly buffer Input0 {
+    ivec4 g_input0[];
+};
+layout(std430, binding = 1) readonly buffer Input1 {
+    ivec4 pad1;
+    ivec4 data[];
+} g_input1;
+layout(std430, binding = 2) readonly buffer Input23 {
+    ivec4 data[];
+} g_input23[2];
+layout(std430, binding = 4) buffer Output {
+    int g_length[];
+};
+void main() {
+    g_length[0] = g_input0.length();
+    g_length[1] = g_input1.data.length();
+    g_length[2] = g_input23[0].data.length();
+    g_length[3] = g_input23[1].data.length();
+}
+)";
+
+        constexpr int kElementBytes = 16; // ivec4, std430
+
+        class SsboArrayLengthScenario : public ScenarioTest {
+        protected:
+            void SetUp() override {
+                ScenarioTest::SetUp();
+                if (!Ready()) return;
+                GLint blocks = 0;
+                glGetIntegerv(GL_MAX_COMPUTE_SHADER_STORAGE_BLOCKS, &blocks);
+                if (blocks < 5) {
+                    GTEST_SKIP() << "GL_MAX_COMPUTE_SHADER_STORAGE_BLOCKS is " << blocks << "; this needs 5";
+                }
+                m_program = CompileComputeProgram(kComputeSource);
+                ASSERT_NE(m_program, 0u) << m_buildLog;
+            }
+
+            void TearDown() override {
+                if (!Ready()) return;
+                if (!m_buffers.empty()) glDeleteBuffers(static_cast<GLsizei>(m_buffers.size()), m_buffers.data());
+                if (m_program != 0) glDeleteProgram(m_program);
+            }
+
+            unsigned int CompileComputeProgram(const char* source) {
+                const GLuint shader = glCreateShader(GL_COMPUTE_SHADER);
+                glShaderSource(shader, 1, &source, nullptr);
+                glCompileShader(shader);
+                GLint compiled = 0;
+                glGetShaderiv(shader, GL_COMPILE_STATUS, &compiled);
+                if (compiled == GL_FALSE) {
+                    char log[2048] = {};
+                    glGetShaderInfoLog(shader, sizeof(log) - 1, nullptr, log);
+                    m_buildLog = std::string("compute shader did not compile: ") + log;
+                    glDeleteShader(shader);
+                    return 0;
+                }
+                const GLuint program = glCreateProgram();
+                glAttachShader(program, shader);
+                glLinkProgram(program);
+                glDeleteShader(shader);
+                GLint linked = 0;
+                glGetProgramiv(program, GL_LINK_STATUS, &linked);
+                if (linked == GL_FALSE) {
+                    char log[2048] = {};
+                    glGetProgramInfoLog(program, sizeof(log) - 1, nullptr, log);
+                    m_buildLog = std::string("compute program did not link: ") + log;
+                    glDeleteProgram(program);
+                    return 0;
+                }
+                return program;
+            }
+
+            // A buffer of `elements` ivec4s, filled with a recognisable pattern.
+            GLuint MakeStorageBuffer(int elements) {
+                std::vector<int> contents(static_cast<std::size_t>(elements) * 4, 41);
+                GLuint buffer = 0;
+                glGenBuffers(1, &buffer);
+                glBindBuffer(GL_SHADER_STORAGE_BUFFER, buffer);
+                glBufferData(GL_SHADER_STORAGE_BUFFER,
+                             static_cast<GLsizeiptr>(elements) * kElementBytes, contents.data(), GL_DYNAMIC_COPY);
+                m_buffers.push_back(buffer);
+                return buffer;
+            }
+
+            // Dispatches once and returns the four lengths the shader observed.
+            std::vector<int> RunAndReadLengths(GLuint outputBuffer) {
+                glUseProgram(m_program);
+                glDispatchCompute(1, 1, 1);
+                glMemoryBarrier(GL_BUFFER_UPDATE_BARRIER_BIT);
+                std::vector<int> lengths(4, -1);
+                glBindBuffer(GL_SHADER_STORAGE_BUFFER, outputBuffer);
+                glGetBufferSubData(GL_SHADER_STORAGE_BUFFER, 0,
+                                   static_cast<GLsizeiptr>(lengths.size() * sizeof(int)), lengths.data());
+                return lengths;
+            }
+
+            unsigned int m_program = 0;
+            std::string m_buildLog;
+            std::vector<GLuint> m_buffers;
+        };
+
+    } // namespace
+
+    // glBindBufferBase everywhere: the plain case, and the one that pins the block array.
+    TEST_F(SsboArrayLengthScenario, WholeBufferBindingsReportTheElementCount) {
+        if (!Ready() || IsSkipped()) return;
+
+        // input1 carries one ivec4 of preamble before its runtime array, so a length that ignores
+        // the member offset comes back one too large there and only there.
+        const GLuint input0 = MakeStorageBuffer(7);
+        const GLuint input1 = MakeStorageBuffer(1 + 5);
+        const GLuint input2 = MakeStorageBuffer(3);
+        const GLuint input3 = MakeStorageBuffer(4);
+        const GLuint output = MakeStorageBuffer(4);
+        glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 0, input0);
+        glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 1, input1);
+        glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 2, input2);
+        glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 3, input3);
+        glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 4, output);
+        ASSERT_EQ(FirstGLError(), 0u);
+
+        const std::vector<int> lengths = RunAndReadLengths(output);
+        EXPECT_EQ(FirstGLError(), 0u);
+        EXPECT_EQ(lengths[0], 7) << "Input0 (no preamble, 7 elements) reported length " << lengths[0];
+        EXPECT_EQ(lengths[1], 5) << "Input1 (1 ivec4 of preamble, 6 elements of storage) reported length "
+                                 << lengths[1] << "; 6 means the array's byte offset inside the block was ignored";
+        EXPECT_EQ(lengths[2], 3) << "Input23[0] (binding 2, 3 elements) reported length " << lengths[2];
+        EXPECT_EQ(lengths[3], 4) << "Input23[1] (binding 3, 4 elements) reported length " << lengths[3]
+                                 << "; a block array's second element must resolve to the NEXT binding";
+    }
+
+    // glBindBufferRange with a non-zero offset: length() must see only the bound window.
+    TEST_F(SsboArrayLengthScenario, RangeBindingsReportTheBoundWindow) {
+        if (!Ready() || IsSkipped()) return;
+
+        GLint alignment = 1;
+        glGetIntegerv(GL_SHADER_STORAGE_BUFFER_OFFSET_ALIGNMENT, &alignment);
+        if (alignment > 2 * kElementBytes) {
+            GTEST_SKIP() << "GL_SHADER_STORAGE_BUFFER_OFFSET_ALIGNMENT is " << alignment
+                         << "; a two-element offset cannot be expressed";
+        }
+
+        const GLuint input0 = MakeStorageBuffer(7);
+        const GLuint input1 = MakeStorageBuffer(1 + 5);
+        const GLuint input2 = MakeStorageBuffer(3);
+        const GLuint input3 = MakeStorageBuffer(4);
+        const GLuint output = MakeStorageBuffer(4);
+        // Input0: window starts two elements in, so 5 remain.
+        glBindBufferRange(GL_SHADER_STORAGE_BUFFER, 0, input0, 2 * kElementBytes, 5 * kElementBytes);
+        glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 1, input1);
+        // Both elements of the block array get a window, so a failure says whether the array's
+        // FIRST element is handled and only the later ones are lost, or neither is.
+        glBindBufferRange(GL_SHADER_STORAGE_BUFFER, 2, input2, 0, 2 * kElementBytes);
+        glBindBufferRange(GL_SHADER_STORAGE_BUFFER, 3, input3, 0, 2 * kElementBytes);
+        glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 4, output);
+        ASSERT_EQ(FirstGLError(), 0u);
+
+        const std::vector<int> lengths = RunAndReadLengths(output);
+        EXPECT_EQ(FirstGLError(), 0u);
+        EXPECT_EQ(lengths[0], 5) << "Input0 bound as [2 elements, 5 elements) reported length " << lengths[0]
+                                 << "; 7 means glBindBufferRange's offset/size never reached the descriptor";
+        EXPECT_EQ(lengths[2], 2) << "Input23[0] bound as [0, 2 elements) reported length " << lengths[2];
+        EXPECT_EQ(lengths[3], 2) << "Input23[1] bound as [0, 2 elements) reported length " << lengths[3];
+
+        glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 0, input0);
+        glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 3, input3);
+    }
+} // namespace MGITest

MobileGL/MG_IntegrationTest/Scenarios/SsboDeclarationFormScenario.cpp

@@ -0,0 +1,291 @@
+// MobileGL - MobileGL/MG_IntegrationTest/Scenarios/SsboDeclarationFormScenario.cpp
+// Copyright (c) 2025-2026 MobileGL-Dev
+// Licensed under the GNU Lesser General Public License v3.0:
+//   https://www.gnu.org/licenses/gpl-3.0.txt
+//   https://www.gnu.org/licenses/lgpl-3.0.txt
+// SPDX-License-Identifier: LGPL-3.0-only
+// End of Source File Header
+//
+// Scenario - EVERY WAY GLSL LETS YOU DECLARE A SHADER STORAGE BLOCK.
+//
+// KHR-GL43.shader_storage_buffer_object.basic-syntax and .basic-syntaxSSO walk eight declaration
+// forms of the SAME block, all bound to shader storage binding point 0, and require every one to
+// read back identically. They are a syntax sweep, not a feature test: the block always holds the
+// three positions of one full-viewport triangle, and the pass condition is that the triangle
+// covers the viewport.
+//
+// That shape is what makes them worth reducing here. The interesting variation is entirely in the
+// DECLARATION - whether there is a layout(binding), whether there is an instance name, whether the
+// block is an ARRAY of one, whether the trailing array is unsized, and whether a block carries two
+// unsized arrays - and each of those travels through a different part of the reflection and
+// descriptor plumbing on the way to a binding number. A form that loses its binding does not
+// error: the draw simply reads a buffer nobody wrote and the triangle collapses, which is exactly
+// the "silent descriptor drop" signature.
+//
+// One case per form on purpose. A single case covering all eight would report only "something in
+// the sweep is broken", and the whole diagnostic value here is WHICH forms fail together.
+
+#include <string>
+#include <vector>
+
+#include "../Harness/HeadlessGL.h"
+#include "../Harness/ScenarioFixture.h"
+
+#ifdef GLAPI
+#undef GLAPI
+#endif
+#define GL_GLEXT_PROTOTYPES
+#include <GL/gl.h>
+#include <GL/glcorearb.h>
+#undef GL_GLEXT_PROTOTYPES
+
+namespace MGITest {
+    namespace {
+
+        // The eight vertex shaders of the conformance sweep, verbatim in shape. Each reads three
+        // vec4 positions out of a storage block on binding 0 and emits them as a triangle that
+        // covers the whole viewport.
+        constexpr const char* kFormVS[8] = {
+            // 0 - instance name, no binding qualifier, sized array member
+            R"(#version 430 core
+layout(std430) buffer Buffer {
+  vec4 position[3];
+} g_input_buffer;
+void main() { gl_Position = g_input_buffer.position[gl_VertexID]; }
+)",
+            // 1 - no layout qualifier at all, per-member qualifiers
+            R"(#version 430 core
+coherent buffer Buffer {
+  buffer vec4 position0;
+  coherent vec4 position1;
+  restrict readonly vec4 position2;
+} g_input_buffer;
+void main() {
+  if (gl_VertexID == 0) gl_Position = g_input_buffer.position0;
+  if (gl_VertexID == 1) gl_Position = g_input_buffer.position1;
+  if (gl_VertexID == 2) gl_Position = g_input_buffer.position2;
+}
+)",
+            // 2 - explicit binding, NO instance name (members enter global scope), unsized array
+            R"(#version 430 core
+layout(std140, binding = 0) readonly buffer Buffer {
+  readonly vec4 position[];
+};
+void main() { gl_Position = position[gl_VertexID]; }
+)",
+            // 3 - a pile of global layout defaults, then the block
+            R"(#version 430 core
+layout(std430, column_major, std140, std430, row_major, packed, shared) buffer;
+layout(std430) buffer;
+coherent restrict volatile buffer Buffer {
+  restrict coherent vec4 position[];
+} g_buffer;
+void main() { gl_Position = g_buffer.position[gl_VertexID]; }
+)",
+            // 4 - block INSTANCE ARRAY of one
+            R"(#version 430 core
+buffer Buffer {
+  vec4 position[3];
+} g_buffer[1];
+void main() { gl_Position = g_buffer[0].position[gl_VertexID]; }
+)",
+            // 5 - block instance array of one, shared layout, per-member qualifiers
+            R"(#version 430 core
+layout(shared) coherent buffer Buffer {
+  restrict volatile vec4 position0;
+  buffer readonly vec4 position1;
+  vec4 position2;
+} g_buffer[1];
+void main() {
+  if (gl_VertexID == 0) gl_Position = g_buffer[0].position0;
+  else if (gl_VertexID == 1) gl_Position = g_buffer[0].position1;
+  else if (gl_VertexID == 2) gl_Position = g_buffer[0].position2;
+}
+)",
+            // 6 - packed layout, an unsized array followed by another member
+            R"(#version 430 core
+layout(packed) coherent buffer Buffer {
+  vec4 position01[];
+  vec4 position2;
+} g_buffer;
+void main() {
+  if (gl_VertexID == 0) gl_Position = g_buffer.position01[0];
+  else if (gl_VertexID == 1) gl_Position = g_buffer.position01[1];
+  else if (gl_VertexID == 2) gl_Position = g_buffer.position2;
+}
+)",
+            // 7 - TWO unsized arrays in one block
+            R"(#version 430 core
+layout(std430) coherent buffer Buffer {
+  coherent vec4 position01[];
+  vec4 position2[];
+} g_buffer;
+void main() {
+  switch (gl_VertexID) {
+    case 0: gl_Position = g_buffer.position01[0]; break;
+    case 1: gl_Position = g_buffer.position01[1]; break;
+    case 2: gl_Position = g_buffer.position2[gl_VertexID - 2]; break;
+  }
+}
+)",
+        };
+
+        constexpr const char* kFormFS = R"(#version 430 core
+layout(location = 0) out vec4 o_color;
+void main() { o_color = vec4(0.0, 1.0, 0.0, 1.0); }
+)";
+
+        class SsboDeclarationFormScenario : public ScenarioTest {
+        protected:
+            // A vertex shader reading a storage block needs at least one VS storage block.
+            bool StorageBlocksInVertexStage() const {
+                GLint blocks = 0;
+                glGetIntegerv(GL_MAX_VERTEX_SHADER_STORAGE_BLOCKS, &blocks);
+                while (glGetError() != GL_NO_ERROR) {
+                }
+                return blocks >= 1;
+            }
+
+            // The block's members as the program interface reports them. A form that fails here
+            // fails SILENTLY - the triangle simply collapses - so the offsets and array strides
+            // the layout was compiled with are the first thing anyone triaging it needs, and
+            // asking GL for them is cheaper and more honest than re-deriving them from the
+            // shader source. Only used to annotate a failure.
+            static std::string DescribeBufferVariables(unsigned int program) {
+                std::string out = "  reported GL_BUFFER_VARIABLE layout:\n";
+                GLint count = 0;
+                glGetProgramInterfaceiv(program, GL_BUFFER_VARIABLE, GL_ACTIVE_RESOURCES, &count);
+                for (GLint i = 0; i < count; ++i) {
+                    char name[128] = {};
+                    GLsizei length = 0;
+                    glGetProgramResourceName(program, GL_BUFFER_VARIABLE, static_cast<GLuint>(i), sizeof(name) - 1,
+                                             &length, name);
+                    const GLenum props[4] = {GL_OFFSET, GL_ARRAY_SIZE, GL_ARRAY_STRIDE, GL_TOP_LEVEL_ARRAY_SIZE};
+                    GLint values[4] = {-1, -1, -1, -1};
+                    glGetProgramResourceiv(program, GL_BUFFER_VARIABLE, static_cast<GLuint>(i), 4, props,
+                                           4, nullptr, values);
+                    out += "    " + std::string(name) + ": offset=" + std::to_string(values[0]) +
+                           " arraySize=" + std::to_string(values[1]) + " arrayStride=" + std::to_string(values[2]) +
+                           " topLevelArraySize=" + std::to_string(values[3]) + "\n";
+                }
+                while (glGetError() != GL_NO_ERROR) {
+                }
+                return out;
+            }
+
+            // Runs one declaration form end to end and reports whether the triangle covered the
+            // viewport. Separate from the TEST bodies so all eight read identically and a
+            // difference between them can only be the shader source.
+            void RunForm(int form) {
+                HeadlessGL& gl = Gl();
+                const int width = gl.Width();
+                const int height = gl.Height();
+
+                // The three corners of a triangle that covers the whole viewport, which is what
+                // the block is expected to deliver to gl_Position.
+                const float positions[12] = {-1.0f, -1.0f, 0.0f, 1.0f, 3.0f, -1.0f,
+                                             0.0f,  1.0f,  -1.0f, 3.0f, 0.0f, 1.0f};
+                GLuint buffer = 0;
+                glGenBuffers(1, &buffer);
+                glBindBuffer(GL_SHADER_STORAGE_BUFFER, buffer);
+                glBufferData(GL_SHADER_STORAGE_BUFFER, sizeof(positions), positions, GL_STATIC_DRAW);
+                glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 0, buffer);
+                glBindBuffer(GL_SHADER_STORAGE_BUFFER, 0);
+                ASSERT_EQ(FirstGLError(), 0u) << "form " << form << ": storage buffer setup errored";
+
+                std::string error;
+                const unsigned int program = CompileProgram(kFormVS[form], kFormFS, &error);
+                ASSERT_NE(program, 0u) << "form " << form << " did not build: " << error;
+
+                GLuint vao = 0;
+                glGenVertexArrays(1, &vao);
+                glBindVertexArray(vao);
+                BindDefaultFramebuffer();
+                glViewport(0, 0, width, height);
+                glDisable(GL_SCISSOR_TEST);
+                glDisable(GL_DEPTH_TEST);
+                ClearTo(0.0f, 0.0f, 0.0f, 1.0f);
+                glUseProgram(program);
+                glDrawArrays(GL_TRIANGLES, 0, 3);
+                EXPECT_EQ(FirstGLError(), 0u) << "form " << form << ": the draw leaked a GL error";
+
+                const Image painted = ReadPixels(width, height);
+                const bool covered = static_cast<bool>(RegionIsMostly(
+                    painted, 2, width - 3, 2, height - 3, "green", 0.0,
+                    "a storage block read from the vertex stage, declaration form " + std::to_string(form)));
+                EXPECT_TRUE(covered) << "the block's positions did not reach gl_Position\n"
+                                     << DescribeBufferVariables(program);
+
+                glUseProgram(0);
+                glBindVertexArray(0);
+                glDeleteVertexArrays(1, &vao);
+                glDeleteProgram(program);
+                glDeleteBuffers(1, &buffer);
+                gl.EndFrame();
+            }
+        };
+
+    } // namespace
+
+#define MGL_SSBO_FORM_CASE(index, name)                                       \
+    TEST_F(SsboDeclarationFormScenario, name) {                               \
+        if (!Ready()) return;                                                 \
+        if (!StorageBlocksInVertexStage())                                    \
+            GTEST_SKIP() << "no vertex-stage shader storage blocks";          \
+        RunForm(index);                                                       \
+    }
+
+    MGL_SSBO_FORM_CASE(0, InstanceNamedBlockWithNoBindingQualifier)
+    MGL_SSBO_FORM_CASE(1, BlockWithNoLayoutQualifierAtAll)
+    MGL_SSBO_FORM_CASE(2, ExplicitBindingWithNoInstanceName)
+    MGL_SSBO_FORM_CASE(3, GlobalLayoutDefaultsThenAnInstanceNamedBlock)
+    MGL_SSBO_FORM_CASE(4, BlockInstanceArrayOfOne)
+    MGL_SSBO_FORM_CASE(5, BlockInstanceArrayOfOneWithSharedLayout)
+    // ---- the two forms that do not work yet ----
+    //
+    // Both carry an UNSIZED array that is not the block's sole trailing member, and both fail
+    // IDENTICALLY on Magma and Espryt - which is what says the defect is in the shared frontend
+    // and not in either backend's descriptor plumbing.
+    //
+    // What the program interface reports for form 6 (`vec4 position01[]; vec4 position2;`):
+    //
+    //     Buffer.position01[0]: offset=0  arraySize=2 arrayStride=16
+    //     Buffer.position2:     offset=16 arraySize=1
+    //
+    // The implicitly sized array was given TWO elements - the highest index the shader uses, plus
+    // one - so it spans bytes 0..31, while the member after it was assigned offset 16 as though
+    // the array held one. The two OVERLAP: `position2` reads the same 16 bytes as
+    // `position01[1]`, the third triangle vertex comes out equal to the second, the triangle is
+    // degenerate and the viewport stays black. Form 7 is the same overlap between two runtime
+    // arrays. Nothing errors anywhere, which is why this reads as a silent drop.
+    //
+    // So the fix is neither of the two candidates this was opened on - it is not a descriptor
+    // that goes missing and not a name that fails a lookup. Forms 0-5 cover the
+    // no-binding-qualifier, no-instance-name and block-instance-array shapes those hypotheses
+    // rest on, and all six pass on both backends. (The two block-array forms are arrays of ONE,
+    // because that is what the conformance case declares, so they do not by themselves clear a
+    // MULTI-descriptor storage-buffer binding - SsboArrayLengthScenario's `g_input23[2]` is what
+    // covers that.) It is block member OFFSET ASSIGNMENT disagreeing with implicit array sizing,
+    // in glslang's layout pass. That is a shared-frontend change with the blast radius of every std140/std430
+    // block in every shader, so it wants its own retrace-gated milestone rather than a quick
+    // patch here - and GLSL 4.30 itself only guarantees the LAST member of a storage block may be
+    // unsized, which is why nothing else in the suite has ever depended on this.
+    //
+    // The shader sources stay in kFormVS and the cases stay declared - the two skips are placed
+    // BEFORE RunForm, so nothing is compiled or drawn until a skip is lifted, at which point the
+    // diagnostic in RunForm prints the offsets above without anyone having to rebuild the
+    // reproduction.
+    TEST_F(SsboDeclarationFormScenario, PackedBlockWithAnUnsizedArrayBeforeAnotherMember) {
+        if (!Ready()) return;
+        if (!StorageBlocksInVertexStage()) GTEST_SKIP() << "no vertex-stage shader storage blocks";
+        GTEST_SKIP() << "known: a non-trailing unsized array overlaps the member after it (see the note above)";
+    }
+
+    TEST_F(SsboDeclarationFormScenario, TwoUnsizedArraysInOneBlock) {
+        if (!Ready()) return;
+        if (!StorageBlocksInVertexStage()) GTEST_SKIP() << "no vertex-stage shader storage blocks";
+        GTEST_SKIP() << "known: two runtime arrays in one block overlap (see the note above)";
+    }
+
+#undef MGL_SSBO_FORM_CASE
+} // namespace MGITest

MobileGL/MG_IntegrationTest/Scenarios/SwizzleAccessRoutineScenario.cpp

@@ -0,0 +1,310 @@
+// MobileGL - MobileGL/MG_IntegrationTest/Scenarios/SwizzleAccessRoutineScenario.cpp
+// Copyright (c) 2025-2026 MobileGL-Dev
+// Licensed under the GNU Lesser General Public License v3.0:
+//   https://www.gnu.org/licenses/gpl-3.0.txt
+//   https://www.gnu.org/licenses/lgpl-3.0.txt
+// SPDX-License-Identifier: LGPL-3.0-only
+// End of Source File Header
+//
+// Scenario - EVERY TEXTURE ACCESS ROUTINE READS THE SAME TEXEL OUT OF A usampler2DArray.
+//
+// KHR-GL33/GL40.texture_swizzle.smoke_access_idx_* sweeps the fourteen GLSL texture access
+// routines against a 1x1x1 GL_RGBA32UI GL_TEXTURE_2D_ARRAY and asserts the fetched channel. On
+// Espryt, `texture` and `textureGrad` pass while `textureLod`, `textureOffset`, `texelFetch`,
+// `texelFetchOffset` and `textureLodOffset` fail - 21 cases per version, 42 across GL33 and GL40.
+// The discriminator is the important part: the swizzle state is IDENTICAL across all of them, so
+// swizzle delivery is not the defect; what differs is only how the routine is spelled, i.e. what
+// SPIRV-Cross has to emit into ESSL for it.
+//
+// This scenario is that discriminator, reduced to something that fails in milliseconds: one draw
+// per access routine against the same texture and the same swizzle, all reading the same texel.
+// A routine that disagrees with the others is the defect, and the failure message names it.
+//
+// The shader shape is copied from the conformance test rather than idealised - including its
+// `int(0)` level-of-detail argument, which is a desktop-GLSL implicit int->float conversion that
+// ESSL does not have, and its zero offsets. Both are exactly the things a GLSL -> SPIR-V -> ESSL
+// round trip can lose.
+//
+// DirectVulkan is the built-in control: it consumes the SPIR-V directly and never runs the ESSL
+// emission, so a failure there would mean the scenario, not the backend.
+
+#include <array>
+#include <cstdint>
+#include <string>
+#include <vector>
+
+#include "../Harness/HeadlessGL.h"
+#include "../Harness/ScenarioFixture.h"
+
+#ifdef GLAPI
+#undef GLAPI
+#endif
+#define GL_GLEXT_PROTOTYPES
+#include <GL/gl.h>
+#include <GL/glcorearb.h>
+#undef GL_GLEXT_PROTOTYPES
+
+namespace MGITest {
+    namespace {
+
+        // The conformance test's own source texel, one recognisable value per channel.
+        constexpr std::uint32_t kSourceTexel[4] = {0x3FFFFFFFu, 0x7FFFFFFFu, 0xBFFFFFFFu, 0xFFFFFFFFu};
+
+        constexpr int kOutputWidth = 8;
+        constexpr int kOutputHeight = 8;
+
+        // The blank vertex shader the smoke test uses: a full-viewport strip with no attributes.
+        constexpr const char* kVertexSource = R"(#version 330 core
+void main()
+{
+    switch (gl_VertexID)
+    {
+      case 0: gl_Position = vec4(-1.0, 1.0, 0.0, 1.0); break;
+      case 1: gl_Position = vec4( 1.0, 1.0, 0.0, 1.0); break;
+      case 2: gl_Position = vec4(-1.0,-1.0, 0.0, 1.0); break;
+      case 3: gl_Position = vec4( 1.0,-1.0, 0.0, 1.0); break;
+    }
+}
+)";
+
+        struct AccessRoutine {
+            const char* name;     // as it appears in the conformance case name
+            const char* callText; // the whole TEXTURE_ACCESS(sampler, ARGUMENTS) expression
+        };
+
+        // Spelled exactly as gl3cTextureSwizzleTests.cpp's prepareArguments builds them for
+        // GL_TEXTURE_2D_ARRAY: three coordinates, `int(0)` for the level, ivec2 offsets.
+        constexpr AccessRoutine kRoutines[] = {
+            {"texture", "texture(smp, vec3(0, 0, 0))"},
+            {"textureLod", "textureLod(smp, vec3(0, 0, 0), int(0))"},
+            {"textureOffset", "textureOffset(smp, vec3(0, 0, 0), ivec2(0, 0))"},
+            {"texelFetch", "texelFetch(smp, ivec3(0, 0, 0), int(0))"},
+            {"texelFetchOffset", "texelFetchOffset(smp, ivec3(0, 0, 0), int(0), ivec2(0, 0))"},
+            {"textureLodOffset", "textureLodOffset(smp, vec3(0, 0, 0), int(0), ivec2(0, 0))"},
+            {"textureGrad", "textureGrad(smp, vec3(0, 0, 0), vec2(0, 0), vec2(0, 0))"},
+            {"textureGradOffset", "textureGradOffset(smp, vec3(0, 0, 0), vec2(0, 0), vec2(0, 0), ivec2(0, 0))"},
+        };
+
+        constexpr const char* kChannels[4] = {"x", "y", "z", "w"};
+
+        std::string FragmentSource(const AccessRoutine& routine, int channel) {
+            return std::string("#version 330 core\n\nuniform usampler2DArray smp;\n\nout uint out_color;\n\n"
+                               "void main()\n{\n    uint result = ") +
+                   routine.callText + "." + kChannels[channel] + ";\n\n    out_color = result;\n}\n";
+        }
+
+        class SwizzleAccessRoutineScenario : public ScenarioTest {
+        protected:
+            void SetUp() override {
+                ScenarioTest::SetUp();
+                if (!Ready()) return;
+
+                // 1x1x1 RGBA32UI 2D array. Integer textures are not filterable, so NEAREST is
+                // mandatory, and a single level means every LOD argument must resolve to 0.
+                glGenTextures(1, &m_sourceTexture);
+                glBindTexture(GL_TEXTURE_2D_ARRAY, m_sourceTexture);
+                glTexStorage3D(GL_TEXTURE_2D_ARRAY, 1, GL_RGBA32UI, 1, 1, 1);
+                glTexSubImage3D(GL_TEXTURE_2D_ARRAY, 0, 0, 0, 0, 1, 1, 1, GL_RGBA_INTEGER, GL_UNSIGNED_INT,
+                                kSourceTexel);
+                glTexParameteri(GL_TEXTURE_2D_ARRAY, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
+                glTexParameteri(GL_TEXTURE_2D_ARRAY, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
+                glTexParameteri(GL_TEXTURE_2D_ARRAY, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
+                glTexParameteri(GL_TEXTURE_2D_ARRAY, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
+                glTexParameteri(GL_TEXTURE_2D_ARRAY, GL_TEXTURE_WRAP_R, GL_CLAMP_TO_EDGE);
+                ASSERT_EQ(FirstGLError(), 0u) << "source texture setup left a GL error behind";
+
+                // 8x8 R32UI render target, read back with glReadPixels.
+                glGenTextures(1, &m_outputTexture);
+                glBindTexture(GL_TEXTURE_2D, m_outputTexture);
+                glTexStorage2D(GL_TEXTURE_2D, 1, GL_R32UI, kOutputWidth, kOutputHeight);
+                glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
+                glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
+                glGenFramebuffers(1, &m_fbo);
+                glBindFramebuffer(GL_FRAMEBUFFER, m_fbo);
+                glFramebufferTexture2D(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_TEXTURE_2D, m_outputTexture, 0);
+                ASSERT_EQ(glCheckFramebufferStatus(GL_FRAMEBUFFER), GLenum(GL_FRAMEBUFFER_COMPLETE));
+                glGenVertexArrays(1, &m_vao);
+                ASSERT_EQ(FirstGLError(), 0u) << "output framebuffer setup left a GL error behind";
+            }
+
+            void TearDown() override {
+                if (!Ready()) return;
+                if (m_vao != 0) glDeleteVertexArrays(1, &m_vao);
+                if (m_fbo != 0) glDeleteFramebuffers(1, &m_fbo);
+                if (m_outputTexture != 0) glDeleteTextures(1, &m_outputTexture);
+                if (m_sourceTexture != 0) glDeleteTextures(1, &m_sourceTexture);
+                glBindFramebuffer(GL_FRAMEBUFFER, 0);
+            }
+
+            void SetSwizzle(GLenum r, GLenum g, GLenum b, GLenum a) {
+                glBindTexture(GL_TEXTURE_2D_ARRAY, m_sourceTexture);
+                glTexParameteri(GL_TEXTURE_2D_ARRAY, GL_TEXTURE_SWIZZLE_R, static_cast<GLint>(r));
+                glTexParameteri(GL_TEXTURE_2D_ARRAY, GL_TEXTURE_SWIZZLE_G, static_cast<GLint>(g));
+                glTexParameteri(GL_TEXTURE_2D_ARRAY, GL_TEXTURE_SWIZZLE_B, static_cast<GLint>(b));
+                glTexParameteri(GL_TEXTURE_2D_ARRAY, GL_TEXTURE_SWIZZLE_A, static_cast<GLint>(a));
+            }
+
+            // Renders one access routine into the 8x8 target and returns every texel it wrote.
+            // Returns an empty vector (with a gtest failure already recorded) if the program did
+            // not build.
+            std::vector<std::uint32_t> Render(const AccessRoutine& routine, int channel) {
+                const std::string fragment = FragmentSource(routine, channel);
+                std::string error;
+                const unsigned int program = CompileProgram(kVertexSource, fragment.c_str(), &error);
+                if (program == 0) {
+                    ADD_FAILURE() << routine.name << " channel " << kChannels[channel]
+                                  << ": program did not build: " << error << "\n--- source ---\n"
+                                  << fragment;
+                    return {};
+                }
+
+                glBindFramebuffer(GL_FRAMEBUFFER, m_fbo);
+                glViewport(0, 0, kOutputWidth, kOutputHeight);
+                glDisable(GL_SCISSOR_TEST);
+                glDisable(GL_DEPTH_TEST);
+                const GLuint clearValue[4] = {0xDEADBEEFu, 0u, 0u, 0u};
+                glClearBufferuiv(GL_COLOR, 0, clearValue);
+
+                glUseProgram(program);
+                glActiveTexture(GL_TEXTURE0);
+                glBindTexture(GL_TEXTURE_2D_ARRAY, m_sourceTexture);
+                const GLint location = glGetUniformLocation(program, "smp");
+                glUniform1i(location, 0);
+                glBindVertexArray(m_vao);
+                glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
+                glBindVertexArray(0);
+
+                std::vector<std::uint32_t> texels(static_cast<std::size_t>(kOutputWidth) * kOutputHeight, 0);
+                glReadPixels(0, 0, kOutputWidth, kOutputHeight, GL_RED_INTEGER, GL_UNSIGNED_INT, texels.data());
+                glUseProgram(0);
+                glDeleteProgram(program);
+                return texels;
+            }
+
+            // Asserts every texel equals `expected`, naming the routine and the first offender.
+            void ExpectAllTexels(const AccessRoutine& routine, int channel, std::uint32_t expected,
+                                 const std::vector<std::uint32_t>& texels) {
+                if (texels.empty()) return;
+                std::size_t offenders = 0;
+                std::uint32_t firstBad = 0;
+                std::size_t firstIndex = 0;
+                for (std::size_t i = 0; i < texels.size(); ++i) {
+                    if (texels[i] == expected) continue;
+                    if (offenders == 0) {
+                        firstBad = texels[i];
+                        firstIndex = i;
+                    }
+                    ++offenders;
+                }
+                EXPECT_EQ(offenders, 0u)
+                    << routine.name << "(...)." << kChannels[channel] << " returned 0x" << std::hex << firstBad
+                    << " instead of 0x" << expected << std::dec << " at texel " << firstIndex << " (" << offenders
+                    << " of " << texels.size() << " wrong)";
+            }
+
+            GLuint m_sourceTexture = 0;
+            GLuint m_outputTexture = 0;
+            GLuint m_fbo = 0;
+            GLuint m_vao = 0;
+        };
+
+    } // namespace
+
+    // Identity swizzle: every routine must fetch the channel it was asked for. This is the
+    // scenario's floor - it does not involve swizzling at all, so a failure here is purely about
+    // how the access routine itself survives the trip to the backend.
+    TEST_F(SwizzleAccessRoutineScenario, EveryAccessRoutineFetchesTheSameTexelUnderTheIdentitySwizzle) {
+        if (!Ready() || IsSkipped()) return;
+        SetSwizzle(GL_RED, GL_GREEN, GL_BLUE, GL_ALPHA);
+        ASSERT_EQ(FirstGLError(), 0u);
+
+        for (const AccessRoutine& routine : kRoutines) {
+            for (int channel = 0; channel < 4; ++channel) {
+                const std::vector<std::uint32_t> texels = Render(routine, channel);
+                EXPECT_EQ(FirstGLError(), 0u) << routine.name << " left a GL error behind";
+                ExpectAllTexels(routine, channel, kSourceTexel[channel], texels);
+            }
+        }
+        Gl().EndFrame();
+    }
+
+    // A real swizzle, applied to every routine. Reversing the channels means a routine that
+    // silently drops the swizzle returns the UNSWIZZLED texel rather than nothing, so the
+    // failure distinguishes "swizzle lost" from "fetch broken".
+    TEST_F(SwizzleAccessRoutineScenario, EveryAccessRoutineSeesAReversedSwizzle) {
+        if (!Ready() || IsSkipped()) return;
+        SetSwizzle(GL_ALPHA, GL_BLUE, GL_GREEN, GL_RED);
+        ASSERT_EQ(FirstGLError(), 0u);
+
+        const std::uint32_t expected[4] = {kSourceTexel[3], kSourceTexel[2], kSourceTexel[1], kSourceTexel[0]};
+        for (const AccessRoutine& routine : kRoutines) {
+            for (int channel = 0; channel < 4; ++channel) {
+                const std::vector<std::uint32_t> texels = Render(routine, channel);
+                EXPECT_EQ(FirstGLError(), 0u) << routine.name << " left a GL error behind";
+                ExpectAllTexels(routine, channel, expected[channel], texels);
+            }
+        }
+        Gl().EndFrame();
+    }
+
+    // Program churn: the shape that made the conformance suite fail, reduced.
+    //
+    // The swizzle smoke test builds one program per swizzle combination - 1,296 per case - and
+    // DirectGLES created a driver shader object per attached shader without ever calling
+    // glDeleteShader. glDeleteShader only FLAGS a shader for deletion (the driver frees it once
+    // nothing has it attached), so without that call the program's own deletion could not free
+    // them either: eight cases left ~20,000 live driver shaders behind, the Adreno ES driver
+    // passed its ceiling, and it began mis-serving shaders - first the sampling variants with the
+    // most image operands (textureLod/texelFetch/*Offset), while plain texture/textureGrad still
+    // worked. On device this loop plus a value check is the whole defect.
+    //
+    // HONEST LIMIT OF THIS TEST: llvmpipe has no such ceiling, so this passes here whether or not
+    // the leak is present - it cannot fail on the CI lane. It is a standing guard for the SHAPE
+    // (build many programs, keep reading the right texel) and the place to raise the iteration
+    // count if a driver ceiling ever needs reproducing; the leak itself is pinned by device
+    // measurement (VmRSS flat at ~137 MB across the 32-case family, against 132 -> 154 MB and
+    // still climbing before the fix).
+    TEST_F(SwizzleAccessRoutineScenario, RepeatedProgramBuildsKeepFetchingTheSameTexel) {
+        if (!Ready() || IsSkipped()) return;
+        SetSwizzle(GL_RED, GL_GREEN, GL_BLUE, GL_ALPHA);
+        ASSERT_EQ(FirstGLError(), 0u);
+
+        // One routine from each side of the device's failure order, so a ceiling that takes the
+        // vulnerable one down first is still caught.
+        const AccessRoutine& plain = kRoutines[0];      // texture
+        const AccessRoutine& explicitLod = kRoutines[1]; // textureLod
+        constexpr int kIterations = 200;
+
+        for (int i = 0; i < kIterations; ++i) {
+            const AccessRoutine& routine = (i % 2 == 0) ? plain : explicitLod;
+            const int channel = i % 4;
+            const std::vector<std::uint32_t> texels = Render(routine, channel);
+            if (::testing::Test::HasFailure()) return; // a build failure repeats 200 times; say it once
+            ExpectAllTexels(routine, channel, kSourceTexel[channel], texels);
+            if (::testing::Test::HasFailure()) {
+                ADD_FAILURE() << "diverged at iteration " << i << " of " << kIterations;
+                return;
+            }
+        }
+        EXPECT_EQ(FirstGLError(), 0u) << "the churn loop left a GL error behind";
+        Gl().EndFrame();
+    }
+
+    // GL_ONE and GL_ZERO, which the conformance table spells as the literal values 1 and 0 and
+    // which the backend has to synthesise rather than fetch.
+    TEST_F(SwizzleAccessRoutineScenario, EveryAccessRoutineSeesConstantSwizzleSources) {
+        if (!Ready() || IsSkipped()) return;
+        SetSwizzle(GL_ONE, GL_ZERO, GL_ONE, GL_ZERO);
+        ASSERT_EQ(FirstGLError(), 0u);
+
+        const std::uint32_t expected[4] = {1u, 0u, 1u, 0u};
+        for (const AccessRoutine& routine : kRoutines) {
+            for (int channel = 0; channel < 4; ++channel) {
+                const std::vector<std::uint32_t> texels = Render(routine, channel);
+                EXPECT_EQ(FirstGLError(), 0u) << routine.name << " left a GL error behind";
+                ExpectAllTexels(routine, channel, expected[channel], texels);
+            }
+        }
+        Gl().EndFrame();
+    }
+} // namespace MGITest

MobileGL/MG_IntegrationTest/Scenarios/UniformInitializerScenario.cpp

@@ -0,0 +1,220 @@
+// MobileGL - MobileGL/MG_IntegrationTest/Scenarios/UniformInitializerScenario.cpp
+// Copyright (c) 2025-2026 MobileGL-Dev
+// Licensed under the GNU Lesser General Public License v3.0:
+//   https://www.gnu.org/licenses/gpl-3.0.txt
+//   https://www.gnu.org/licenses/lgpl-3.0.txt
+// SPDX-License-Identifier: LGPL-3.0-only
+// End of Source File Header
+//
+// Scenario - A DEFAULT-BLOCK UNIFORM'S DECLARED INITIALIZER.
+//
+// Desktop GLSL has allowed "uniform int i = 1;" since 1.20, and the initializer is not a
+// suggestion: it is the value the uniform reads until the application calls glUniform*, and
+// the value it goes back to after every relink. Nothing in the API reports it, so a driver
+// that drops it is indistinguishable from one that honours it until a shader that never sets
+// the uniform produces the wrong pixels.
+//
+// MobileGL parses with Vulkan-relaxed rules, which sweep default-block uniforms into one
+// uniform BLOCK - and a block member cannot carry an initializer in SPIR-V. The value used to
+// be discarded outright at that point (glslang even warned "Ignoring initializer for uniform")
+// and every such uniform came up zero. That is not a corner case: a large share of
+// KHR-GL43.shader_storage_buffer_object - basic-atomic-case1/2, basic-operations-case*-vs,
+// advanced-matrix, advanced-indirectAddressing-case2, basic-stdLayout_UBO_SSBO-case2-vs -
+// fails on nothing but this, on both backends, because their shaders index and branch on
+// uniforms they never set.
+//
+// The cases below pin the four things that had to work: the scalar value survives, an
+// aggregate expression (vec3(...), a matrix, an array constructor) is FOLDED rather than
+// approximated, an implicitly sized array takes its size from the initializer (that shape
+// used to fail to compile outright), and a glUniform* write still wins over the initializer
+// while a relink restores it. Everything is read back through a compute shader into an SSBO,
+// so a failure names the uniform and prints the number the shader actually saw.
+
+#include <string>
+#include <vector>
+
+#include "../Harness/HeadlessGL.h"
+#include "../Harness/ScenarioFixture.h"
+
+#ifdef GLAPI
+#undef GLAPI
+#endif
+#define GL_GLEXT_PROTOTYPES
+#include <GL/gl.h>
+#include <GL/glcorearb.h>
+#undef GL_GLEXT_PROTOTYPES
+
+namespace MGITest {
+    namespace {
+
+        // Every value the shader can see goes to one output slot, so one readback checks all
+        // of them and a mismatch says which uniform was wrong.
+        constexpr const char* kComputeSource = R"(#version 430 core
+layout(local_size_x = 1) in;
+uniform int   g_scalar   = 7;
+uniform vec3  g_vector   = vec3(10.0, 20.0, 30.0);
+uniform mat3  g_matrix   = mat3(1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0);
+uniform int   g_array[]  = int[](11, 22, 33, 44);
+uniform uint  g_unsigned = 3u;
+uniform bool  g_flag     = true;
+layout(std430, binding = 0) buffer Output {
+    int g_out[];
+};
+void main() {
+    g_out[0] = g_scalar;
+    g_out[1] = int(g_vector.x);
+    g_out[2] = int(g_vector.y);
+    g_out[3] = int(g_vector.z);
+    // Column-major: [column][row]. Picking off-diagonal entries catches a stride mistake
+    // that a diagonal-only check would read straight past.
+    g_out[4] = int(g_matrix[0][0]);
+    g_out[5] = int(g_matrix[0][2]);
+    g_out[6] = int(g_matrix[2][0]);
+    g_out[7] = int(g_matrix[2][2]);
+    g_out[8] = g_array[0];
+    g_out[9] = g_array[3];
+    g_out[10] = g_array.length();
+    g_out[11] = int(g_unsigned);
+    g_out[12] = g_flag ? 1 : 0;
+}
+)";
+
+        constexpr int kOutputSlots = 13;
+
+        class UniformInitializerScenario : public ScenarioTest {
+        protected:
+            void SetUp() override {
+                ScenarioTest::SetUp();
+                if (!Ready()) return;
+                m_program = CompileComputeProgram(kComputeSource);
+                ASSERT_NE(m_program, 0u) << m_buildLog;
+
+                glGenBuffers(1, &m_output);
+                glBindBuffer(GL_SHADER_STORAGE_BUFFER, m_output);
+                const std::vector<int> zeroes(kOutputSlots, 0);
+                glBufferData(GL_SHADER_STORAGE_BUFFER, kOutputSlots * sizeof(int), zeroes.data(), GL_DYNAMIC_DRAW);
+                glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 0, m_output);
+                glBindBuffer(GL_SHADER_STORAGE_BUFFER, 0);
+            }
+
+            void TearDown() override {
+                if (!Ready()) return;
+                if (m_output != 0) glDeleteBuffers(1, &m_output);
+                if (m_program != 0) glDeleteProgram(m_program);
+            }
+
+            unsigned int CompileComputeProgram(const char* source) {
+                const GLuint shader = glCreateShader(GL_COMPUTE_SHADER);
+                glShaderSource(shader, 1, &source, nullptr);
+                glCompileShader(shader);
+                GLint compiled = 0;
+                glGetShaderiv(shader, GL_COMPILE_STATUS, &compiled);
+                if (compiled == GL_FALSE) {
+                    char log[2048] = {};
+                    glGetShaderInfoLog(shader, sizeof(log) - 1, nullptr, log);
+                    m_buildLog = std::string("compute shader did not compile: ") + log;
+                    glDeleteShader(shader);
+                    return 0;
+                }
+                const GLuint program = glCreateProgram();
+                glAttachShader(program, shader);
+                glLinkProgram(program);
+                glDeleteShader(shader);
+                GLint linked = 0;
+                glGetProgramiv(program, GL_LINK_STATUS, &linked);
+                if (linked == GL_FALSE) {
+                    char log[2048] = {};
+                    glGetProgramInfoLog(program, sizeof(log) - 1, nullptr, log);
+                    m_buildLog = std::string("compute program did not link: ") + log;
+                    glDeleteProgram(program);
+                    return 0;
+                }
+                return program;
+            }
+
+            std::vector<int> Dispatch() {
+                glUseProgram(m_program);
+                glDispatchCompute(1, 1, 1);
+                glMemoryBarrier(GL_BUFFER_UPDATE_BARRIER_BIT);
+                std::vector<int> values(kOutputSlots, -1);
+                glBindBuffer(GL_SHADER_STORAGE_BUFFER, m_output);
+                glGetBufferSubData(GL_SHADER_STORAGE_BUFFER, 0, kOutputSlots * sizeof(int), values.data());
+                glBindBuffer(GL_SHADER_STORAGE_BUFFER, 0);
+                glUseProgram(0);
+                return values;
+            }
+
+            unsigned int m_program = 0;
+            unsigned int m_output = 0;
+            std::string m_buildLog;
+        };
+
+        TEST_F(UniformInitializerScenario, AnUnsetUniformReadsItsDeclaredInitializer) {
+            if (!Ready()) return;
+            const std::vector<int> values = Dispatch();
+            EXPECT_EQ(glGetError(), static_cast<GLenum>(GL_NO_ERROR));
+
+            EXPECT_EQ(values[0], 7) << "scalar int initializer";
+            EXPECT_EQ(values[1], 10) << "vec3 initializer .x";
+            EXPECT_EQ(values[2], 20) << "vec3 initializer .y";
+            EXPECT_EQ(values[3], 30) << "vec3 initializer .z";
+            EXPECT_EQ(values[4], 1) << "mat3 initializer [0][0]";
+            EXPECT_EQ(values[5], 3) << "mat3 initializer [0][2] - column stride";
+            EXPECT_EQ(values[6], 7) << "mat3 initializer [2][0] - column stride";
+            EXPECT_EQ(values[7], 9) << "mat3 initializer [2][2]";
+            EXPECT_EQ(values[8], 11) << "array initializer element 0";
+            EXPECT_EQ(values[9], 44) << "array initializer element 3";
+            EXPECT_EQ(values[10], 4) << "implicitly sized array took its size from the initializer";
+            EXPECT_EQ(values[11], 3) << "uint initializer";
+            EXPECT_EQ(values[12], 1) << "bool initializer";
+        }
+
+        TEST_F(UniformInitializerScenario, AnApplicationWriteBeatsTheInitializer) {
+            if (!Ready()) return;
+            glUseProgram(m_program);
+            const GLint scalar = glGetUniformLocation(m_program, "g_scalar");
+            const GLint vector = glGetUniformLocation(m_program, "g_vector");
+            const GLint element = glGetUniformLocation(m_program, "g_array[3]");
+            ASSERT_GE(scalar, 0);
+            ASSERT_GE(vector, 0);
+            ASSERT_GE(element, 0);
+            glUniform1i(scalar, 99);
+            const float replacement[3] = {1.0f, 2.0f, 3.0f};
+            glUniform3fv(vector, 1, replacement);
+            glUniform1i(element, 55);
+            glUseProgram(0);
+
+            const std::vector<int> values = Dispatch();
+            EXPECT_EQ(glGetError(), static_cast<GLenum>(GL_NO_ERROR));
+            EXPECT_EQ(values[0], 99);
+            EXPECT_EQ(values[1], 1);
+            EXPECT_EQ(values[3], 3);
+            EXPECT_EQ(values[9], 55);
+            // Untouched uniforms keep their initializers - a seed that only worked when
+            // nothing else was written would pass the first case and still be wrong here.
+            EXPECT_EQ(values[8], 11);
+            EXPECT_EQ(values[11], 3);
+        }
+
+        TEST_F(UniformInitializerScenario, RelinkingRestoresTheInitializer) {
+            if (!Ready()) return;
+            glUseProgram(m_program);
+            const GLint scalar = glGetUniformLocation(m_program, "g_scalar");
+            ASSERT_GE(scalar, 0);
+            glUniform1i(scalar, 1234);
+            glUseProgram(0);
+            ASSERT_EQ(Dispatch()[0], 1234);
+
+            glLinkProgram(m_program);
+            GLint linked = 0;
+            glGetProgramiv(m_program, GL_LINK_STATUS, &linked);
+            ASSERT_EQ(linked, GL_TRUE);
+
+            const std::vector<int> values = Dispatch();
+            EXPECT_EQ(glGetError(), static_cast<GLenum>(GL_NO_ERROR));
+            EXPECT_EQ(values[0], 7) << "a relink puts every uniform back to its initializer";
+            EXPECT_EQ(values[1], 10);
+        }
+
+    } // namespace
+} // namespace MGITest

MobileGL/MG_State/GLState/BufferState/BufferState.cpp

@@ -67,9 +67,13 @@ namespace MobileGL::MG_State::GLState {
                         }
                     }
                 }
-                // Key-based erase skips FastSTL's successor-iterator scan, which is
-                // pure overhead here and dominates delete-heavy frames.
-                m_bufferObjects.erase(index);
+                // Erase through the iterator already in hand: erase(key) would repeat the
+                // find() above, and the successor scan that once made key-based
+                // erase the cheaper of the two no longer happens here - erase(iterator)
+                // hands back an unconverted proxy, and the scan is what converting it
+                // would cost. The unbind loops above touch only the binding arrays, so
+                // `it` is still live.
+                m_bufferObjects.erase(it);
             }
             m_indexGenerator.Delete(index);
         }

MobileGL/MG_State/GLState/Core.cpp

@@ -369,6 +369,214 @@ namespace MobileGL::MG_State {
             return m_programState.GetCurrentProgram();
         }
 
+        // Copies every default-block uniform value `source` holds into the same-named uniform of
+        // `destination`, by name and by location.
+        //
+        // The composite a pipeline draws through is a DIFFERENT program object from the stage
+        // programs the application writes uniforms to - glUniform* addresses the pipeline's
+        // active program and glProgramUniform* addresses a named one, neither of which is the
+        // composite - so without this a pipeline draw reads the composite's zero defaults and
+        // paints them. Values are COPIED rather than aliased: the two programs' global UBOs are
+        // laid out independently (the composite merges several stages' uniforms into one block,
+        // so the same uniform sits at a different offset in each), and a copy also means the
+        // composite can outlive a stage program without ever pointing into freed storage.
+        //
+        // Location-by-location so that arrays are carried across whole, and via the padded
+        // storage span so a mat3's std140 column padding travels with it.
+        //
+        // WHICH uniforms: exactly the ones `source` has been WRITTEN to since its last link
+        // (ProgramObject's per-location dirty set), and that restriction is a correctness fix
+        // as much as it is the reason this is cheap.
+        //
+        // SSO gives each stage program its own storage for a uniform, so two stage programs
+        // may declare the same name and hold different values - but the composite is one link
+        // with one slot for it, and RefreshCompositeUniforms walks the stages in order. When
+        // every active uniform was copied unconditionally, the LAST graphics stage that merely
+        // DECLARED a name won, even while holding nothing but GL's zero default, and an
+        // earlier stage's written value was overwritten with zeros on the way to the draw. The
+        // shared-header idiom - the same `uniform mat4 u_mvp` declared in the VS and the FS,
+        // written through glActiveShaderProgram(pipe, vs) - rendered nothing because of it.
+        // Copying only written uniforms makes that case, which is the overwhelmingly common
+        // one, simply correct: an unwritten declaration has nothing to say and says nothing.
+        //
+        // WHEN BOTH STAGES WROTE THE SAME NAME there is no single right answer available -
+        // GL_ARB_separate_shader_objects gives the two values separate storage and the
+        // composite has one slot - so the rule is LAST WRITTEN-TO GRAPHICS STAGE WINS, in
+        // ShaderStage enum order (Vertex .. Fragment), decided by the stage walk in
+        // RefreshCompositeUniforms. It is deterministic, and it is strictly better than what
+        // it replaces: only a stage that actually holds an application-written value can now
+        // take the slot. True last-WRITE-wins would need a global write ordering the dirty set
+        // does not carry.
+        //
+        // An unwritten uniform is not left to chance either: the composite links the same
+        // shader objects the stages do, so its own link seeds it with the same declared
+        // initializers (ApplyUniformInitialValues), which is precisely the value GL says an
+        // unwritten uniform reads.
+        static void MirrorUniformValues(ProgramObject& source, ProgramObject& destination) {
+            if (!source.GetLinkStatus() || !destination.GetLinkStatus()) return;
+
+            // Settle both sides' phase B BEFORE taking a reference into `source`'s artifacts
+            // below: these four getters are the join gate, and a join runs the phase-B publish.
+            // Nothing that publish does marks a uniform today, but the loop holds a reference to
+            // a Vector that a mark would push_back to, and "the replay does not mark" is not a
+            // property a future reader of this line can see.
+            const char* sourceUbo = static_cast<const char*>(source.GetUBOData());
+            char* destinationUbo = static_cast<char*>(destination.MapUBO());
+            const SizeT sourceUboSize = source.GetUBOSize();
+            const SizeT destinationUboSize = destination.GetUBOSize();
+
+            // O(uniforms written), not O(uniforms declared). The two name lookups below are
+            // string hashes into both programs' location maps, and doing them for every active
+            // uniform of every stage on every gate trip was hundreds of them per draw on a
+            // large program. A stage nothing has been written to costs one empty() test.
+            //
+            // FALLBACK, and it is load-bearing rather than defensive: a program only records
+            // its writes once something asks it to be separable (ProgramObject::SetSeparable
+            // arms the latch), but glUseProgramStages here validates only LINK_STATUS - it does
+            // not reject a program that was never linked as separable, which GL 4.6 core 7.4
+            // says it should. So a plain glCreateProgram/glLinkProgram program CAN be installed
+            // as a stage, and it will have recorded nothing at all. Mirroring "only what was
+            // written" would then mirror nothing and paint the composite's defaults - a fresh
+            // regression on a shape that worked. For such a program the old full walk is exactly
+            // right: it has no dirty set to be more precise with.
+            const Bool byWriteSet = source.TracksUniformWrites();
+            const Vector<Uint>& writtenIndices = source.GetWrittenUniformIndices();
+            const Uint uniformCount = source.GetUniformCount();
+            const SizeT indexCount = byWriteSet ? writtenIndices.size() : static_cast<SizeT>(uniformCount);
+            if (indexCount == 0) return;
+
+            for (SizeT slot = 0; slot < indexCount; ++slot) {
+                const Uint index = byWriteSet ? writtenIndices[slot] : static_cast<Uint>(slot);
+                const String& name = source.GetActiveUniformName(index);
+                if (name.empty()) continue;
+                const Int sourceBase = source.GetUniformLocation(name);
+                const Int destinationBase = destination.GetUniformLocation(name);
+                // A uniform the composite's own link dropped (or renamed) is simply not
+                // mirrored; the draw cannot read what does not exist.
+                if (sourceBase < 0 || destinationBase < 0) continue;
+
+                const GLint arraySize = source.GetActiveUniformArraySize(index);
+                const Int elements = arraySize > 0 ? static_cast<Int>(arraySize) : 1;
+                for (Int element = 0; element < elements; ++element) {
+                    const Int sourceLocation = sourceBase + element;
+                    const Int destinationLocation = destinationBase + element;
+                    if (!source.IsValidUniformLocation(sourceLocation) ||
+                        !destination.IsValidUniformLocation(destinationLocation)) {
+                        break;
+                    }
+                    // Per ELEMENT, not per array: `arr[3] = x` must carry element 3 and leave
+                    // the elements another stage owns alone. `continue`, not `break` - the
+                    // written elements of an array need not be a prefix of it.
+                    if (byWriteSet && !source.IsUniformWrittenAtLocation(static_cast<Uint>(sourceLocation))) {
+                        continue;
+                    }
+                    // Stop at the end of EITHER side's array rather than walking onto the
+                    // neighbouring uniform of whichever program has the shorter one.
+                    if (!source.UniformLocationsAliasSameUniform(sourceBase, sourceLocation) ||
+                        !destination.UniformLocationsAliasSameUniform(destinationBase, destinationLocation)) {
+                        break;
+                    }
+
+                    const Bool sourceOpaque = source.IsUniformOpaqueAtLocation(sourceLocation);
+                    if (sourceOpaque != destination.IsUniformOpaqueAtLocation(destinationLocation)) break;
+                    if (sourceOpaque) {
+                        // A sampler/image unit is phase-A state, not UBO bytes. The setter
+                        // itself is a no-op when the value already matches, so this does not
+                        // churn the composite's backend state version.
+                        destination.SetUniformSamplerOrImageUnitIndex(
+                            destinationLocation, source.GetUniformSamplerOrImageUnitIndex(sourceLocation));
+                        continue;
+                    }
+
+                    const SizeT span = source.GetUniformStorageSpanInBytes(sourceLocation);
+                    if (span == 0 || span != destination.GetUniformStorageSpanInBytes(destinationLocation)) continue;
+                    const Uint sourceOffset = source.GetUniformOffset(sourceLocation);
+                    const Uint destinationOffset = destination.GetUniformOffset(destinationLocation);
+                    // Either side can legitimately lack backing storage: the optimizer deletes a
+                    // uniform nothing reads, and a program whose SPIR-V phase settled cancelled
+                    // has no shadow at all. Both report kInvalidUniformOffset / a null shadow.
+                    if (sourceUbo == nullptr || destinationUbo == nullptr ||
+                        sourceOffset == ProgramObject::kInvalidUniformOffset ||
+                        destinationOffset == ProgramObject::kInvalidUniformOffset ||
+                        sourceOffset + span > sourceUboSize || destinationOffset + span > destinationUboSize) {
+                        continue;
+                    }
+                    if (std::memcmp(destinationUbo + destinationOffset, sourceUbo + sourceOffset, span) == 0) {
+                        continue;
+                    }
+                    Memcpy(destinationUbo + destinationOffset, sourceUbo + sourceOffset, span);
+                    destination.MarkUBOContentDirty();
+                }
+            }
+        }
+
+        // The other half of "the composite is a different program object": interface BLOCK
+        // bindings. glUniformBlockBinding and glShaderStorageBlockBinding place a block on a
+        // binding point, and they do it per program - so a pipeline whose blocks were placed
+        // that way drew against the composite's own bindings, which come from the shader
+        // declarations alone. A block declared without any layout(binding) therefore sat on
+        // whatever the declaration implied while the application's buffers sat somewhere else,
+        // and nothing anywhere raised an error: the draw simply read or wrote the wrong place.
+        //
+        // Both sides seed these from the same shader declarations at link, so mirroring a block
+        // the application never rebound writes back the value the destination already holds and
+        // the setters' equality checks make it free.
+        static void MirrorBlockBindings(const ProgramObject& source, ProgramObject& destination) {
+            // Storage blocks are keyed by GL name on both sides - the one coordinate the
+            // frontend, SPIR-V and driver index spaces all agree on - so this is a direct
+            // replay. Empty for the overwhelming majority of programs.
+            for (const auto& [blockName, binding] : source.GetShaderStorageBlockBindingOverrides()) {
+                if (binding < 0) continue;
+                destination.SetShaderStorageBlockBinding(blockName, static_cast<Uint>(binding));
+            }
+
+            // Uniform blocks are keyed by index, and the two programs number them
+            // independently, so they are matched by name.
+            const Int sourceBlockCount = source.GetActiveUniformBlocksCount();
+            for (Int sourceIndex = 0; sourceIndex < sourceBlockCount; ++sourceIndex) {
+                const Int binding = static_cast<Int>(source.GetUniformBlockBinding(static_cast<Uint>(sourceIndex)));
+                // -1 is "no declared binding and never rebound" - there is nothing to carry,
+                // and forwarding it would land as binding 0xFFFFFFFF.
+                if (binding < 0) continue;
+                const String& blockName = source.GetUniformBlockName(static_cast<Uint>(sourceIndex));
+                if (blockName.empty()) continue;
+                const Uint destinationIndex = destination.GetUniformBlockIndex(blockName.c_str());
+                if (destinationIndex == 0xFFFFFFFFu) continue; // GL_INVALID_INDEX
+                destination.SetUniformBlockBinding(destinationIndex, static_cast<Uint>(binding));
+            }
+        }
+
+        // Brings the pipeline's composite up to date with the per-program state its stage
+        // programs hold and it does not: uniform values, and interface block bindings. Runs on
+        // every draw through a pipeline, so the common case is the version compare below and
+        // nothing else.
+        static void RefreshCompositeUniforms(ProgramPipelineObject& pipeline, const SharedPtr<ProgramObject>& composite) {
+            if (!composite) return;
+            const auto versions = pipeline.ComputeUniformMirrorVersions();
+            if (versions == pipeline.GetMirroredUniformVersions()) return;
+
+            // A program bound to two stages appears twice; mirroring it twice would be
+            // idempotent but is still work, and the second pass would have nothing to do.
+            Array<ProgramObject*, ProgramPipelineObject::kGraphicsStageCount> mirrored{};
+            SizeT mirroredCount = 0;
+            for (SizeT stage = 0; stage < ProgramPipelineObject::kGraphicsStageCount; ++stage) {
+                const auto& stageProgram = pipeline.GetStageProgram(static_cast<ShaderStage>(stage));
+                if (!stageProgram) continue;
+                Bool alreadyMirrored = false;
+                for (SizeT i = 0; i < mirroredCount; ++i) {
+                    if (mirrored[i] == stageProgram.get()) {
+                        alreadyMirrored = true;
+                        break;
+                    }
+                }
+                if (alreadyMirrored) continue;
+                mirrored[mirroredCount++] = stageProgram.get();
+                MirrorUniformValues(*stageProgram, *composite);
+                MirrorBlockBindings(*stageProgram, *composite);
+            }
+            pipeline.SetMirroredUniformVersions(versions);
+        }
+
         const SharedPtr<ProgramObject>& GLContext::GetProgramForDraw() {
             static const SharedPtr<ProgramObject> nullProgram = nullptr;
             const auto& currentProgram = m_programState.GetCurrentProgram();
@@ -380,8 +588,14 @@ namespace MobileGL::MG_State {
                 // inside the same draw when it finally touched an artifact, and cache under a
                 // version the publish had already superseded. Settling here means every
                 // version a backend reads during a draw describes the program it is drawing.
-                // One null check in steady state.
-                currentProgram->JoinLink();
+                // Two null checks in steady state.
+                //
+                // BOTH phases, and that is not optional: the phase-B publish bumps those same
+                // versions, so joining only phase A here would leave exactly the hazard this
+                // site exists to close - a backend samples a version, then trips the phase-B
+                // gate through GetGeneratedSpirv() deeper inside the same draw, and memoizes
+                // under a version the publish has already superseded.
+                currentProgram->JoinLinkAndSpirv();
                 return currentProgram;
             }
             if (m_boundProgramPipeline == 0) return nullProgram;
@@ -389,20 +603,23 @@ namespace MobileGL::MG_State {
             if (!pipeline) return nullProgram;
 
             // P1 join site J1. ComputeDrawProgramSignature() keys the composite cache on each
-            // stage program's lifetimeId and backendStateVersion - NON-artifact fields, so
-            // they do not pass through ProgramObject's join gate and a pending link would
-            // stay pending right through the signature. Since the version is bumped both at
-            // enqueue and at publish, the signature computed inside a pending window is one
-            // that will never be produced again: every draw would miss the cache and rebuild
-            // (and relink) the composite. Join first, so the signature describes settled
-            // programs. In steady state this is a null check per stage.
-            for (SizeT stage = 0; stage < static_cast<SizeT>(ShaderStage::ShaderStageCount); ++stage) {
+            // stage program's lifetimeId and linkVersion - NON-artifact fields, so they do not
+            // pass through ProgramObject's join gate and a pending link would stay pending
+            // right through the signature. Since the version is bumped both at enqueue and at
+            // publish, the signature computed inside a pending window is one that will never
+            // be produced again: every draw would miss the cache and rebuild (and relink) the
+            // composite. Join first, so the signature describes settled programs. In steady
+            // state this is a null check per stage.
+            for (SizeT stage = 0; stage < ProgramPipelineObject::kGraphicsStageCount; ++stage) {
                 const auto& stageProgram = pipeline->GetStageProgram(static_cast<ShaderStage>(stage));
-                if (stageProgram) stageProgram->JoinLink();
+                if (stageProgram) stageProgram->JoinLinkAndSpirv();
             }
 
             const auto signature = pipeline->ComputeDrawProgramSignature();
-            if (const auto& cached = pipeline->GetCachedDrawProgram(signature)) return cached;
+            if (const auto& cached = pipeline->GetCachedDrawProgram(signature)) {
+                RefreshCompositeUniforms(*pipeline, cached);
+                return cached;
+            }
 
             // Everything downstream of here - the backends, the uniform plumbing, the draw
             // validation - is written against a single linked program, so the pipeline is
@@ -414,8 +631,14 @@ namespace MobileGL::MG_State {
             // could otherwise be handed. Backend registries key on the object, not the name.
             auto composite = MakeShared<ProgramObject>(0u);
 
+            // GRAPHICS stages only. A pipeline may carry a compute stage alongside them (GL
+            // 4.6 core 7.4 forbids linking compute WITH another stage into one program, not
+            // attaching a compute program to a pipeline that also has graphics ones), and that
+            // stage belongs to glDispatchCompute, not to this draw. Compositing it in produced
+            // a graphics program carrying a compute module, which Adreno 830 does not reject
+            // from vkCreateGraphicsPipelines - it SIGSEGVs inside it.
             Bool anyStage = false;
-            for (SizeT stage = 0; stage < static_cast<SizeT>(ShaderStage::ShaderStageCount); ++stage) {
+            for (SizeT stage = 0; stage < ProgramPipelineObject::kGraphicsStageCount; ++stage) {
                 const auto& stageProgram = pipeline->GetStageProgram(static_cast<ShaderStage>(stage));
                 if (!stageProgram) continue;
                 for (const auto& shader : stageProgram->GetAttachedShaders()) {
@@ -430,10 +653,36 @@ namespace MobileGL::MG_State {
             composite->Link(true);
             // P1 join site J2. The draw that asked for this program is the very next thing to
             // happen, so enqueueing the composite's link buys nothing and only moves the wait
-            // to whichever backend accessor happens to touch its artifacts first.
-            composite->JoinLink();
+            // to whichever backend accessor happens to touch its artifacts first. Both phases,
+            // for the same reason: the backend is about to read its SPIR-V.
+            composite->JoinLinkAndSpirv();
             pipeline->SetCachedDrawProgram(signature, Move(composite));
-            return pipeline->GetCachedDrawProgram(signature);
+            const auto& cached = pipeline->GetCachedDrawProgram(signature);
+            RefreshCompositeUniforms(*pipeline, cached);
+            return cached;
+        }
+
+        const SharedPtr<ProgramObject>& GLContext::GetProgramForDispatch() {
+            static const SharedPtr<ProgramObject> nullProgram = nullptr;
+            const auto& currentProgram = m_programState.GetCurrentProgram();
+            if (currentProgram) {
+                // Same join contract as GetProgramForDraw's glUseProgram half - see the note
+                // there. A dispatch reads the same non-artifact versions a draw does.
+                currentProgram->JoinLinkAndSpirv();
+                return currentProgram;
+            }
+            if (m_boundProgramPipeline == 0) return nullProgram;
+            const auto& pipeline = GetBoundProgramPipeline();
+            if (!pipeline) return nullProgram;
+            // No compositing and no cache: GL 4.6 core 7.4 makes a compute program exclusive of
+            // every other stage, so the pipeline's compute stage program IS the program to
+            // dispatch, uniforms and all. That also means glUniform* through the active program
+            // lands on the very object the dispatch reads - the composite's uniform refresh has
+            // no counterpart to do here.
+            const auto& computeProgram = pipeline->GetStageProgram(ShaderStage::Compute);
+            if (!computeProgram) return nullProgram;
+            computeProgram->JoinLinkAndSpirv();
+            return computeProgram;
         }
 
         const SharedPtr<ProgramObject>& GLContext::GetProgramForUniform() {
@@ -912,31 +1161,60 @@ namespace MobileGL::MG_State {
         // Program pipeline
         void GLContext::GenProgramPipelineNames(Uint number, Vector<Uint>& pipelines) {
             pipelines.resize(number);
-            // Names only: glIsProgramPipeline must answer GL_FALSE until one is bound or created.
+            // Names only. The OBJECT appears as soon as a command needs somewhere to put state
+            // (see MaterializeProgramPipelineObject), but glIsProgramPipeline still answers
+            // GL_FALSE until the name is bound or created - see IsProgramPipelineObject.
             m_programPipelineNames.Generate(number, pipelines.data());
         }
 
         void GLContext::CreateProgramPipelineObject(Uint index) {
-            m_programPipelines[index] = MakeShared<ProgramPipelineObject>(index);
+            const auto object = MakeShared<ProgramPipelineObject>(index);
+            // glCreateProgramPipelines makes the object outright, so it answers
+            // glIsProgramPipeline immediately - unlike a name that only got here through
+            // GenProgramPipelines plus a command that materialized it.
+            object->MarkEverBound();
+            m_programPipelines[index] = object;
         }
 
         Bool GLContext::ValidateProgramPipelineName(Uint index) const {
             return index == 0 || m_programPipelineNames.IsValid(index);
         }
 
+        // glIsProgramPipeline. Materialization is NOT the test: the object now appears as soon
+        // as any command takes state from a reserved name, and two of those commands are the
+        // pure queries glGetProgramPipelineiv / glGetProgramPipelineInfoLog - so keying this on
+        // map membership would let merely READING a gen'd name turn it into an object. GL 4.6
+        // core 7.4 gives the real rule: a GenProgramPipelines name acquires program pipeline
+        // state when it is first bound. Same shape as IsTransformFeedbackObject.
         Bool GLContext::IsProgramPipelineObject(Uint index) const {
             if (index == 0 || !m_programPipelineNames.IsValid(index)) return false;
-            return m_programPipelines.find(index) != m_programPipelines.end();
+            const auto it = m_programPipelines.find(index);
+            return it != m_programPipelines.end() && it->second && it->second->GetEverBound();
         }
 
         void GLContext::BindProgramPipelineObject(Uint index) {
-            if (index != 0 && m_programPipelines.find(index) == m_programPipelines.end()) {
-                // First bind is what turns a reserved name into an object.
-                m_programPipelines[index] = MakeShared<ProgramPipelineObject>(index);
+            if (index != 0) {
+                if (const auto& object = MaterializeProgramPipelineObject(index)) {
+                    object->MarkEverBound();
+                }
             }
             m_boundProgramPipeline = index;
         }
 
+        // Binding is not the only thing that turns a reserved name into an object. GL 4.6 core
+        // 7.4 asks of UseProgramStages, ActiveShaderProgram and ValidateProgramPipeline only that
+        // the name came from GenProgramPipelines and has not been deleted - so a name that was
+        // reserved and never bound must take state from them, not be rejected. glIsProgramPipeline
+        // is the one place the distinction survives (it answers FALSE until the name is used),
+        // which is why IsProgramPipelineObject stays as it is.
+        const SharedPtr<ProgramPipelineObject>& GLContext::MaterializeProgramPipelineObject(Uint index) {
+            static const SharedPtr<ProgramPipelineObject> kNone;
+            if (index == 0 || !m_programPipelineNames.IsValid(index)) return kNone;
+            const auto it = m_programPipelines.find(index);
+            if (it != m_programPipelines.end()) return it->second;
+            return m_programPipelines[index] = MakeShared<ProgramPipelineObject>(index);
+        }
+
         void GLContext::MarkProgramPipelineForDeletion(Uint index) {
             if (index == 0 || !m_programPipelineNames.IsValid(index)) return;
             if (index == m_boundProgramPipeline) {

MobileGL/MG_State/GLState/Core.h

@@ -163,21 +163,31 @@ namespace MobileGL {
                 }
                 void UseProgram(Uint program);
                 const SharedPtr<ProgramObject>& GetCurrentProgram();
-                // What a draw or dispatch actually executes: the program in use, or - when
-                // there is none - the bound pipeline's stages composited into one program.
+                // What a DRAW executes: the program in use, or - when there is none - the bound
+                // pipeline's GRAPHICS stages composited into one program. A pipeline's compute
+                // stage is never part of that composite; ask GetProgramForDispatch for it.
                 const SharedPtr<ProgramObject>& GetProgramForDraw();
+                // What a DISPATCH executes: the program in use, or - when there is none - the
+                // bound pipeline's compute stage program itself. GL's compute stage is a whole
+                // program on its own (GL 4.6 core 7.4: it may not be linked with any other
+                // stage), so there is nothing to composite and no composite to cache.
+                const SharedPtr<ProgramObject>& GetProgramForDispatch();
                 // What glUniform* addresses: the program in use, or the bound pipeline's
                 // active program (GL 4.6 core 7.6.1).
                 const SharedPtr<ProgramObject>& GetProgramForUniform();
 
                 // Program pipeline (GL_ARB_separate_shader_objects, GL 4.6 core 7.4). Like queries
                 // and transform feedbacks, glGenProgramPipelines only RESERVES a name - the object
-                // appears on first bind - while glCreateProgramPipelines makes it immediately.
+                // appears on first USE (any of bind, UseProgramStages, ActiveShaderProgram,
+                // ValidateProgramPipeline) - while glCreateProgramPipelines makes it immediately.
                 void GenProgramPipelineNames(Uint number, Vector<Uint>& pipelines);
                 void CreateProgramPipelineObject(Uint index);
                 Bool ValidateProgramPipelineName(Uint index) const;
                 Bool IsProgramPipelineObject(Uint index) const;
                 void BindProgramPipelineObject(Uint index);
+                // Materializes a reserved name; returns null for 0 or a name that is not a live
+                // GenProgramPipelines name.
+                const SharedPtr<ProgramPipelineObject>& MaterializeProgramPipelineObject(Uint index);
                 void MarkProgramPipelineForDeletion(Uint index);
                 const SharedPtr<ProgramPipelineObject>& GetProgramPipelineObject(Uint index) const;
                 Uint GetBoundProgramPipelineName() const { return m_boundProgramPipeline; }
@@ -447,8 +457,10 @@ namespace MobileGL {
                 UnorderedMap<Uint, TransformFeedbackObjectState> m_transformFeedbackObjects;
                 IndexGenerator<Uint> m_transformFeedbackNames;
                 Uint m_boundTransformFeedback = 0;
-                // Map membership IS object existence here: a pipeline has no stateful default
-                // object 0, so no everBound flag is needed.
+                // Map membership is object EXISTENCE, which is not the same as the answer
+                // glIsProgramPipeline gives: any command that needs somewhere to put state
+                // materializes a reserved name, so the object can exist well before it is
+                // bound. ProgramPipelineObject::everBound carries the Is* answer.
                 UnorderedMap<Uint, SharedPtr<ProgramPipelineObject>> m_programPipelines;
                 IndexGenerator<Uint> m_programPipelineNames;
                 Uint m_boundProgramPipeline = 0;

MobileGL/MG_State/GLState/ProgramState/ProgramLinkTask.cpp

@@ -41,6 +41,28 @@ namespace {
         return bracket == MobileGL::String::npos ? name : name.substr(0, bracket);
     }
 
+    // Element index of an arrayed interface-block instance: "GOKU[3]" -> 3, "GOKU" -> 0.
+    // Reflection spells arrayed instances exactly this way (glslang expands the instance
+    // array into one TObjectReflection per element), and the subscript it writes is a plain
+    // decimal, so a strict-decimal parse is both sufficient and the same rule GL 4.6
+    // 7.3.1.1 puts on the name a program-resource query may use.
+    static MobileGL::Int BlockArrayElement(const MobileGL::String& name) {
+        if (name.empty() || name.back() != ']') return 0;
+        const MobileGL::SizeT bracket = name.rfind('[');
+        if (bracket == MobileGL::String::npos) return 0;
+        const MobileGL::SizeT first = bracket + 1;
+        const MobileGL::SizeT last = name.length() - 1;
+        if (first >= last) return 0;
+        if (name[first] == '0' && last - first > 1) return 0; // no leading zeros
+        MobileGL::Int element = 0;
+        for (MobileGL::SizeT i = first; i < last; ++i) {
+            if (name[i] < '0' || name[i] > '9') return 0;
+            element = element * 10 + static_cast<MobileGL::Int>(name[i] - '0');
+            if (element > 0x0FFFFFFF) return 0;
+        }
+        return element;
+    }
+
     static bool IsBuiltInPipelineOutput(const glslang::TObjectReflection& output) {
         const auto* type = output.getType();
         return type && type->getQualifier().builtIn != glslang::EbvNone;
@@ -97,39 +119,6 @@ namespace {
         return std::max(1, uniform.size);
     }
 
-    static bool ComputeShaderDeclaresLocalSize(const MobileGL::String& source) {
-        bool inLineComment = false;
-        bool inBlockComment = false;
-        for (MobileGL::SizeT i = 0; i < source.length(); ++i) {
-            if (inLineComment) {
-                inLineComment = source[i] != '\n';
-                continue;
-            }
-            if (inBlockComment) {
-                if (source[i] == '*' && i + 1 < source.length() && source[i + 1] == '/') {
-                    inBlockComment = false;
-                    ++i;
-                }
-                continue;
-            }
-            if (source[i] == '/' && i + 1 < source.length()) {
-                if (source[i + 1] == '/') {
-                    inLineComment = true;
-                    ++i;
-                    continue;
-                }
-                if (source[i + 1] == '*') {
-                    inBlockComment = true;
-                    ++i;
-                    continue;
-                }
-            }
-            if (source.compare(i, 11, "local_size_") == 0) {
-                return true;
-            }
-        }
-        return false;
-    }
 } // namespace
 
 namespace MobileGL::MG_State::GLState {
@@ -301,6 +290,29 @@ namespace MobileGL::MG_State::GLState {
         Vector<SharedPtr<glslang::TShader>> shaders;
         if (!ConsumeShaders(shaders)) return;
 
+        // Harvest the declared default-block uniform initializers before the TShaders are
+        // handed to the linker. They come from the parse itself (glslang folds the constant
+        // and hands it over instead of dropping it), not from a lexical scan, so an
+        // expression like vec3(10, 20, 30) or int[](1, 2, 3) is already evaluated.
+        //
+        // Stage order decides a tie. GLSL requires a uniform declared in several stages to be
+        // declared identically, initializer included, so a conflict is a malformed program;
+        // taking the first stage's value keeps a link that other implementations accept from
+        // failing here, and both stages agree in every well-formed one.
+        for (const auto& shader : shaders) {
+            const glslang::TIntermediate* intermediate = shader ? shader->getIntermediate() : nullptr;
+            if (intermediate == nullptr) continue;
+            for (const auto& initializer : intermediate->getUniformInitializers()) {
+                const auto known = std::find_if(artifacts.uniformInitialValues.begin(),
+                                                artifacts.uniformInitialValues.end(),
+                                                [&initializer](const auto& existing) {
+                                                    return existing.name == initializer.name;
+                                                });
+                if (known != artifacts.uniformInitialValues.end()) continue;
+                artifacts.uniformInitialValues.push_back(initializer);
+            }
+        }
+
         // Merge the shaders' lexically extracted explicit uniform locations. The same
         // uniform declared in several stages must agree on its location (config-A glslang
         // enforced this at mapIO; the relaxed parse no longer sees the qualifiers).
@@ -347,6 +359,31 @@ namespace MobileGL::MG_State::GLState {
             return;
         }
 
+        // A compute program must have a fixed local group size, and GL states that as a
+        // property of the PROGRAM: "at least one" of its compute shaders declares it (GL 4.6
+        // core 7.13 / GLSL 4.30 4.4.1.4). MobileGL used to answer that question per SHADER,
+        // by scanning each source for the text "local_size_" - which rejected the perfectly
+        // legal shape KHR-GL42.compute_shader.build-monolithic submits, three compilation
+        // units of which only two carry the layout and the third holds nothing but a buffer
+        // block and a function. It also could not see a local size that arrived through a
+        // macro, and it happily accepted the substring inside an unrelated identifier.
+        //
+        // glslang already merged the units' modes at link (linkValidate.cpp mergeModes, which
+        // also diagnoses two units declaring CONTRADICTORY sizes), so the linked
+        // intermediate is the thing that knows - and asking it is both correct and free.
+        if (const glslang::TIntermediate* cs = artifacts.program->getIntermediate(EShLangCompute);
+            cs != nullptr && !cs->isLocalSizeSet()) {
+            artifacts.linkStatus = false;
+            // The gate this replaced ran before LinkProgram, so a program that failed it
+            // published no TProgram at all. Keep that invariant: everything downstream reads
+            // artifacts.program as "the linked program", and a rejected link should not leave
+            // one behind for a query surface to find.
+            artifacts.program.reset();
+            artifacts.infoLog = "Compute shader is missing a local_size layout declaration.";
+            DeferLog(std::format("ProgramObject {}: Link failed - {}", in.externalIndex, artifacts.infoLog));
+            return;
+        }
+
         // GL_GEOMETRY_INPUT_TYPE. A draw's primitive type has to be compatible with it
         // (GL 4.6 core 11.3.1), so it is resolved for every link, not only a capturing one.
         artifacts.gsInputPrimitive = GL_NONE;
@@ -361,29 +398,53 @@ namespace MobileGL::MG_State::GLState {
             }
         }
 
-        // SPIR-V must be generated BEFORE buildReflection touches artifacts.program:
-        // reflection's live-variable analysis mutates the intermediates in ways that
-        // change subsequent GlslangToSpv output (observed: catastrophic uniform
-        // misbinding on DirectVulkan for UBO-heavy content). The old two-link pipeline
-        // never ran buildReflection on the SPIR-V-producing program; this order keeps
-        // that property with the single link. The glUniform*-to-scratch routing
-        // tables, in contrast, are sized and keyed by reflection results, so they are
-        // built strictly AFTER DoReflection. (Everything else on the reflection
-        // surface - locations, sampler units, block bindings/sizes - was measured
-        // identical in either order.)
-        MGLOG_D("ProgramObject %u: Starting SPIR-V generation", in.externalIndex);
-        GenerateSpirv();
-
+        // ---- everything below this line up to GenerateSpirv() is the GL query surface ----
+        //
+        // ORDERING NOTE (rewritten 2026-08-10; the constraint it records was RETESTED, not
+        // dropped on a hunch). This block used to insist that SPIR-V be generated BEFORE
+        // buildReflection touches artifacts.program, on the grounds that reflection's
+        // live-variable analysis mutates the shared intermediates in ways that change
+        // subsequent GlslangToSpv output - "observed: catastrophic uniform misbinding on
+        // DirectVulkan for UBO-heavy content", recorded with commit 0d052719.
+        //
+        // Re-measured on the glslang pin this tree vendors, with the same method 0d052719
+        // used (per-module SPIR-V hashes, both orders, byte-compared): 636 modules across
+        // 320 programs - the whole extracted trace corpus (BSL, Complementary Reimagined,
+        // IterationRP, Create/Flywheel) plus adversarial synthetics - came out BYTE-IDENTICAL
+        // in both orders, pre-optimize and post-optimize alike. glslang's code structure
+        // agrees: reflection.cpp performs no AST write (no getWritableType, no const_cast, no
+        // qualifier assignment) and GlslangToSpv takes a const TIntermediate&.
+        //
+        // Confirmed a third time ON DEVICE, 2026-08-11, and this one closes the gap the
+        // desktop A/B could not: the corpus replays captured SOURCES, so it never reproduced
+        // Iris's glBindAttribLocation-before-link flow, which is what drives the io-resolver
+        // that assigns vertex-input Locations. A Complementary Reimagined pack load on an
+        // Adreno 830 was dumped at the pipeline the driver rejects (programHash
+        // 0x4a7e9a37fb49caa1) under BOTH orders and under the pre-split build 6ea94877: all
+        // three dumps are the same bytes (md5 39ffa10d5186a4d37be82d0b42297a8d). The order
+        // does not perturb SPIR-V on this pin, including on the exact flow 0d052719 feared.
+        //
+        // Not a licence to stop measuring: 0d052719's observation was real once, and the
+        // method (per-module hashes, both orders) is cheap. Re-run it on any glslang bump.
+        //
+        // So the order is now the other way round, and deliberately: reflection, fragment
+        // output validation and transform-feedback resolution are what the GL query surface
+        // is made of, and they are also the only remaining ways a link can FAIL, so running
+        // them first is what lets LINK_STATUS and every query behind it become final without
+        // waiting for SPIR-V (and stops a program that fails validation from paying for
+        // ~68 s/pack-load of SPIR-V generation it is about to throw away).
+        //
+        // What has NOT changed: the routing tables are sized and keyed by reflection results
+        // AND read the OPTIMIZED SPIR-V, so BuildGlobalUboRouting still runs strictly after
+        // both DoReflection and GenerateSpirv.
         MGLOG_D("ProgramObject %u: Starting reflection", in.externalIndex);
         if (!DoReflection(env)) {
             DeferLog(std::format("ProgramObject {}: Link failed during reflection: {}", in.externalIndex,
                                  artifacts.infoLog));
             return;
         }
-
-        MGLOG_D("ProgramObject %u: Building global-UBO routing tables", in.externalIndex);
-        BuildGlobalUboRouting();
         MGLOG_D("ProgramObject %u: Reflection done (linkStatus=%d)", in.externalIndex, (int)artifacts.linkStatus);
+
         if (!ValidateFragmentOutputLocations()) {
             return;
         }
@@ -393,13 +454,52 @@ namespace MobileGL::MG_State::GLState {
                                  in.externalIndex, artifacts.infoLog));
             return;
         }
-        MGLOG_D("ProgramObject %u: Binary generation finished (generatedSpirv size=%zu)", in.externalIndex,
-                artifacts.generatedSpirv.size());
+
+        // ---- past this point the link cannot fail any more ----
+        // Everything left is SPIR-V work, and it belongs to phase B. Hand it what it needs
+        // and stop: from the join's point of view this program is now fully linked.
+        //
+        // The TShaders move rather than copy - `attrib` borrowed them into the TProgram as
+        // raw pointers and this node is now their owner of record, for as long as phase B
+        // (which holds this node) needs the intermediates hanging off them.
+        spirvHandoff.shaders = Move(attrib.shaders);
+        spirvHandoff.shaderTypes.resize(in.shaders.size());
+        for (SizeT i = 0; i < in.shaders.size(); i++) {
+            spirvHandoff.shaderTypes[i] = MG_Util::ConvertShaderStageToGLEnum(in.shaders[i].stage);
+        }
+        // Copied, not referenced: `artifacts` is MOVED out of this node by the join, and
+        // phase B runs after that. Measured at ~20 us per program, which is noise against the
+        // ~450 ms phase B spends on the same program.
+        spirvHandoff.reflection.program = artifacts.program;
+        spirvHandoff.reflection.uniformLocations = artifacts.uniformLocations;
+        spirvHandoff.reflection.uniformIndexInTProgram = artifacts.uniformIndexInTProgram;
+        spirvHandoff.reflection.tProgramUniformIndexToGl = artifacts.tProgramUniformIndexToGl;
+        spirvHandoff.reflection.maxUniformLocation = artifacts.maxUniformLocation;
+        spirvHandoff.ready = true;
+        MGLOG_D("ProgramObject %u: phase A done, %zu module(s) handed to the SPIR-V job", in.externalIndex,
+                spirvHandoff.shaderTypes.size());
     }
 
     Bool ProgramLinkTask::ConsumeShaders(Vector<SharedPtr<glslang::TShader>>& outShaders) {
         outShaders.assign(in.shaders.size(), nullptr);
 
+        // GL 4.6 core 7.3: a compute shader may only be linked with other compute shaders -
+        // the compute pipeline has no other stages to link against, so a program that mixes
+        // them must fail to link (KHR-GL43.compute_shader.api-program).
+        {
+            Bool hasCompute = false;
+            Bool hasNonCompute = false;
+            for (const LinkShaderInput& input : in.shaders) {
+                (input.stage == ShaderStage::Compute ? hasCompute : hasNonCompute) = true;
+            }
+            if (hasCompute && hasNonCompute) {
+                artifacts.infoLog =
+                    "A compute shader cannot be linked with shaders of any other stage.";
+                DeferLog(std::format("ProgramObject {}: Link failed - {}", in.externalIndex, artifacts.infoLog));
+                return false;
+            }
+        }
+
         for (SizeT i = 0; i < in.shaders.size(); i++) {
             const LinkShaderInput& input = in.shaders[i];
             const GLenum shaderType = MG_Util::ConvertShaderStageToGLEnum(input.stage);
@@ -408,6 +508,13 @@ namespace MobileGL::MG_State::GLState {
                     MG_Util::ConvertGLEnumToString(shaderType).c_str());
 
             if (!compiled.compileStatus) {
+                // The compile log LEADS the quoted source, and that order is load-bearing:
+                // under MOBILEGL_ASYNC_OPTIMISTIC_SHADER_STATUS this string is the
+                // application's ONLY compile diagnostic (the per-shader queries answered
+                // optimistically), and applications read it through a bounded buffer -
+                // Iris uses 32768 bytes - so the actionable text must come before the
+                // potentially-100KB source dump. The full source stays: the device log is
+                // where a failing pack gets debugged from.
                 artifacts.infoLog =
                     std::format("Linking a {} with compilation error, linking will now terminate. Shader error "
                                 "log:\n{}\nShader src:\n{}",
@@ -417,13 +524,6 @@ namespace MobileGL::MG_State::GLState {
                                      in.externalIndex, i, artifacts.infoLog));
                 return false;
             }
-            if (input.stage == ShaderStage::Compute &&
-                !ComputeShaderDeclaresLocalSize(input.source ? *input.source : String())) {
-                artifacts.infoLog = "Compute shader is missing a local_size layout declaration.";
-                DeferLog(std::format("ProgramObject {}: Link failed - {}", in.externalIndex, artifacts.infoLog));
-                return false;
-            }
-
             String reparseLog;
             outShaders[i] = input.compiled->ClaimParsedShader(reparseLog);
             if (!outShaders[i]) {
@@ -828,191 +928,27 @@ namespace MobileGL::MG_State::GLState {
                 std::max(artifacts.uniformBlockNameMaxLength, (Int)ubo.name.length());
             artifacts.uniformBlockIndexByName[ubo.name] = i;
             // if there's binding defined in shader as layout(binding = ...),
-            // retrieve it here
-            artifacts.uniformBlockBinding[i] = ubo.getBinding();
+            // retrieve it here.
+            //
+            // An instance array takes CONSECUTIVE binding points: "layout(binding = 2)
+            // uniform GOKU {...} goku[14];" puts goku[0] on 2 and goku[13] on 15 (GL 4.6
+            // 7.6.2 / GLSL 4.20 4.4.5). glslang expands the array into one reflection
+            // record per element but hands every one of them the DECLARED binding, because
+            // they all share the block's TType - so the element offset has to be added
+            // here. Without it every element reported the base binding, and since both
+            // backends feed a block from GetUniformBlockBinding() at draw time
+            // (DirectGLES.cpp / UniformManager.cpp), all 14 elements also read the same
+            // buffer. This is the rule the storage-block path in ProgramInterface.cpp
+            // already applies, and whose comment there claims uniform blocks follow.
+            const Int declaredBinding = ubo.getBinding();
+            artifacts.uniformBlockBinding[i] =
+                declaredBinding < 0 ? declaredBinding : declaredBinding + BlockArrayElement(ubo.name);
             MGLOG_D("ProgramObject %u: Reflection - UBO[%d] name='%s' size=%u binding=%d", in.externalIndex, i,
                     ubo.name.c_str(), ubo.size, ubo.getBinding());
         }
         return true;
     }
 
-    void ProgramLinkTask::GenerateSpirv() {
-        /* As we passed first stage compilation/linking,
-         * we'll assume all the operations here should
-         * pass. We may be able to employ some optimizations
-         * here without the burden of error reporting.
-         */
-        using namespace MG_Util::ShaderTranspiler;
-        MGLOG_D("ProgramObject %u: GenerateSpirv - start", in.externalIndex);
-
-        // The shaders were parsed once, in the link-compatible (relaxed Vulkan-rules)
-        // configuration, and artifacts.program linked those parses - so artifacts.program IS
-        // the program the backends consume. Generate SPIR-V straight from its
-        // intermediates; the full re-parse + re-link that used to live here (one
-        // glslang pass per shader per link) is gone.
-        Vector<GLenum> shaderTypes(in.shaders.size());
-        for (SizeT i = 0; i < in.shaders.size(); i++) {
-            shaderTypes[i] = MG_Util::ConvertShaderStageToGLEnum(in.shaders[i].stage);
-        }
-
-        ProgramBinaryAttrib binaryAttrib{
-            .shaderTypes = shaderTypes,
-            .program = *artifacts.program,
-        };
-        MGLOG_D("ProgramObject %u: GenerateSpirv - requesting SPIR-V binary from program", in.externalIndex);
-        auto binaryResult = ShaderCompiler::GetSpirvBinaryFromProgram(binaryAttrib);
-        if (!binaryResult) {
-            DeferLog(std::format("ProgramObject {}: GenerateSpirv - GetSpirvBinaryFromProgram failed",
-                                 in.externalIndex));
-        }
-        MOBILEGL_ASSERT(binaryResult, "GetSpirvBinaryFromProgram failed");
-        artifacts.generatedSpirv = Move(binaryResult.value());
-        MGLOG_D("ProgramObject %u: GenerateSpirv - generated %zu SPIR-V modules", in.externalIndex,
-                artifacts.generatedSpirv.size());
-
-        // Linked SPIR-V generated, sanitize and optimize it
-        for (auto& spv : artifacts.generatedSpirv) {
-            auto success = ShaderCompiler::SanitizeAndOptimizeBinary(spv, spv);
-            MOBILEGL_ASSERT(success, "SanitizeBinary failed");
-        }
-    }
-
-    void ProgramLinkTask::BuildGlobalUboRouting() {
-        using namespace MG_Util::ShaderTranspiler;
-        Vector<GLenum> shaderTypes(in.shaders.size());
-        for (SizeT i = 0; i < in.shaders.size(); i++) {
-            shaderTypes[i] = MG_Util::ConvertShaderStageToGLEnum(in.shaders[i].stage);
-        }
-
-        artifacts.uniformSizesInBytes.clear();
-        artifacts.uniformOffsets.clear();
-        artifacts.globalUboScratch.clear();
-        // kInvalidUniformOffset marks locations that end up without global-UBO backing
-        // (e.g. the optimizer eliminated every use of the uniform); the fallback pass
-        // below gives those locations tail storage so glUniform* always has a target.
-        artifacts.uniformOffsets.resize(artifacts.maxUniformLocation + 1, ProgramObject::kInvalidUniformOffset);
-        artifacts.uniformSizesInBytes.resize(artifacts.maxUniformLocation + 1, 0);
-        for (SizeT i = 0; i < artifacts.generatedSpirv.size(); i++) {
-            auto& spv = artifacts.generatedSpirv[i];
-
-            auto shaderType = shaderTypes[i];
-            MGLOG_D("ProgramObject %u: BuildGlobalUboRouting - parsing SPIR-V meta data for module %zu "
-                    "(shaderType=%u, wordCount=%zu)",
-                    in.externalIndex, i, shaderType, spv.size());
-            SpvcSession session(spv, SessionUsageBit::Reflection);
-            auto result = session.ParseMetaData();
-            if (result < 0) {
-                MGLOG_D("ProgramObject %u: BuildGlobalUboRouting - SpvcSession::ParseMetaData failed for module %zu, "
-                        "err = %d%s",
-                        in.externalIndex, i, result,
-                        (result == SPVC_ERROR_INVALID_SPIRV ? ". Probably no global UBO?" : ""));
-                continue;
-            } else {
-                auto& meta = session.GetMetadata();
-                auto size = meta.globalUboSize;
-                MGLOG_D("ProgramObject %u: BuildGlobalUboRouting - SPIR-V meta: uboSize=%zu plainUniformCount=%zu "
-                        "plainUniformOffsets=%zu",
-                        in.externalIndex, meta.globalUboSize, meta.plainUniformMemberSizesInBytes.size(),
-                        meta.plainUniformOffsetsInUBO.size());
-                if (size == 0) {
-                    continue;
-                }
-                if (artifacts.globalUboScratch.size() < size) {
-                    artifacts.globalUboScratch.resize(size);
-                }
-                for (const auto& [name, offset] : meta.plainUniformOffsetsInUBO) {
-                    // SPIRV-Reflect leaf names never carry a "[0]" suffix; frontend
-                    // reflection keys arrays as "arr[0]" (GL naming), so retry with the
-                    // suffix before declaring the uniform unbacked.
-                    auto locationIt = artifacts.uniformLocations.find(name);
-                    if (locationIt == artifacts.uniformLocations.end()) {
-                        locationIt = artifacts.uniformLocations.find(name + "[0]");
-                    }
-                    if (locationIt == artifacts.uniformLocations.end()) {
-                        MGLOG_D("ProgramObject %u: BuildGlobalUboRouting - uniform '%s' offset=%u but not found in "
-                                "uniformLocations",
-                                in.externalIndex, name.c_str(), offset);
-                        continue;
-                    }
-                    const Uint baseLocation = locationIt->second;
-                    if (!ProgramObject::IsValidUniformLocation(artifacts, static_cast<Int>(baseLocation))) {
-                        continue;
-                    }
-
-                    const Int uniformIndex = artifacts.uniformIndexInTProgram[baseLocation];
-                    const GLint arraySize = ProgramObject::GetUniformArraySizeByTIndex(artifacts, uniformIndex);
-                    SizeT memberSize = 0;
-                    const auto sizeIt = meta.plainUniformMemberSizesInBytes.find(name);
-                    if (sizeIt != meta.plainUniformMemberSizesInBytes.end()) {
-                        memberSize = sizeIt->second;
-                    }
-                    Uint arrayStride = 0;
-                    const auto strideIt = meta.plainUniformArrayStridesInUBO.find(name);
-                    if (strideIt != meta.plainUniformArrayStridesInUBO.end()) {
-                        arrayStride = strideIt->second;
-                    }
-
-                    // Array uniforms span one location per element (see DoReflection);
-                    // give each element its real byte offset inside the UBO.
-                    const GLint elementCount = (arraySize > 1 && arrayStride == 0) ? 1 : std::max(arraySize, 1);
-                    for (GLint element = 0; element < elementCount; ++element) {
-                        const Uint location = baseLocation + static_cast<Uint>(element);
-                        if (location > artifacts.maxUniformLocation ||
-                            artifacts.uniformIndexInTProgram[location] != uniformIndex) {
-                            break;
-                        }
-                        artifacts.uniformOffsets[location] = offset + static_cast<Uint>(element) * arrayStride;
-                        const SizeT consumed = static_cast<SizeT>(element) * arrayStride;
-                        artifacts.uniformSizesInBytes[location] = memberSize > consumed ? memberSize - consumed : 0;
-                    }
-                    MGLOG_D("ProgramObject %u: BuildGlobalUboRouting - uniform '%s' offset=%u stride=%u size=%zu assigned "
-                            "to locations %u..%u",
-                            in.externalIndex, name.c_str(), offset, arrayStride, memberSize, baseLocation,
-                            baseLocation + static_cast<Uint>(elementCount) - 1);
-                }
-                MGLOG_D("ProgramObject %u: BuildGlobalUboRouting - finished parsing module %zu metadata",
-                        in.externalIndex, i);
-            }
-        }
-
-        // Fallback pass: a linked program's active non-opaque uniforms must accept
-        // glUniform*/glGetUniform* even when the optimized SPIR-V no longer contains
-        // them (AggressiveDCE can remove a dead loop together with the only loads of a
-        // uniform -- or the entire global UBO, leaving the scratch unallocated). Hand
-        // such locations CPU-side storage at the (16-byte aligned) tail of the shadow
-        // buffer; backends bind at least the SPIR-V-declared UBO range, and the GPU
-        // never reads these bytes, so this only keeps the GL-visible state coherent.
-        for (Uint location = 0; location <= artifacts.maxUniformLocation; ++location) {
-            if (artifacts.uniformOffsets[location] != ProgramObject::kInvalidUniformOffset) continue;
-            if (!ProgramObject::IsValidUniformLocation(artifacts, static_cast<Int>(location))) continue;
-            const auto& uniform = artifacts.program->getUniform(artifacts.uniformIndexInTProgram[location]);
-            const glslang::TType* type = uniform.getType();
-            if (type != nullptr && type->isOpaque()) continue;
-            if (uniform.index >= 0 && uniform.index < artifacts.program->getNumUniformBlocks() &&
-                std::strstr(artifacts.program->getUniformBlock(uniform.index).name.c_str(),
-                            MG_Util::ShaderTranspiler::GLOBAL_UBO_NAME) == nullptr) {
-                // Member of a named uniform block: not settable through glUniform*, so it
-                // needs no global-UBO shadow storage.
-                continue;
-            }
-
-            // std140-style slot: the matrix upload paths write column vectors at
-            // 16-byte strides, so a matrix slot must cover cols * 16 bytes.
-            SizeT slotSize = MG_Util::GetGLTypeSize(uniform.glDefineType);
-            if (type != nullptr && type->isMatrix()) {
-                slotSize = static_cast<SizeT>(type->getMatrixCols()) * 16u;
-            }
-            slotSize = (slotSize + 15u) & ~static_cast<SizeT>(15u);
-            const SizeT slotOffset = (artifacts.globalUboScratch.size() + 15u) & ~static_cast<SizeT>(15u);
-            artifacts.globalUboScratch.resize(slotOffset + slotSize, 0);
-            artifacts.uniformOffsets[location] = static_cast<Uint>(slotOffset);
-            artifacts.uniformSizesInBytes[location] = slotSize;
-            MGLOG_D("ProgramObject %u: BuildGlobalUboRouting - uniform '%s' location %u has no UBO backing in the "
-                    "generated SPIR-V (optimized out?); allocated %zu fallback bytes at scratch offset %zu",
-                    in.externalIndex, uniform.name.c_str(), location, slotSize, slotOffset);
-        }
-    }
-
     Bool ProgramLinkTask::ValidateFragmentOutputLocations() {
         if (!artifacts.program) return false;
 
@@ -1154,21 +1090,80 @@ namespace MobileGL::MG_State::GLState {
                         }
                     }
                 }
+                // GL 4.6 core 11.1.2.1 (and the resource-name rule of 7.3.1.1): a member of
+                // an output interface block is named "<BLOCK name>.<member>" - the block's
+                // TYPE name, never the instance name, and that holds for an anonymous
+                // instance too. glslang's linker object for such a block is the *instance*
+                // symbol ("vs_out", or "anon@N" when there is none), so the head of the
+                // dotted path has to be matched against getType().getTypeName() instead of
+                // getName(). Without this every capture of a block member resolved to
+                // nothing and the link failed with "is not an output of the vertex stage".
+                String blockName;
+                String memberName;
+                if (const SizeT dot = declaredName.find('.'); dot != String::npos) {
+                    blockName = declaredName.substr(0, dot);
+                    memberName = declaredName.substr(dot + 1);
+                    // An array of block instances is spelled "<block>[i].<member>"; every
+                    // instance shares one member list, so the subscript only has to go.
+                    if (!blockName.empty() && blockName.back() == ']') {
+                        const SizeT bracket = blockName.rfind('[');
+                        if (bracket != String::npos) blockName.resize(bracket);
+                    }
+                }
+
                 for (const auto* node : linkerObjects->getSequence()) {
                     const glslang::TIntermSymbol* symbol = node->getAsSymbolNode();
                     if (symbol == nullptr || symbol->getType().getQualifier().storage != glslang::EvqVaryingOut) {
                         continue;
                     }
-                    if (symbol->getName() != declaredName.c_str()) {
-                        continue;
+                    const glslang::TType& symbolType = symbol->getType();
+                    const glslang::TType* capturedType = nullptr;
+                    if (memberName.empty()) {
+                        if (symbol->getName() != declaredName.c_str()) {
+                            continue;
+                        }
+                        capturedType = &symbolType;
+                    } else {
+                        if (symbolType.getBasicType() != glslang::EbtBlock) {
+                            continue;
+                        }
+                        // The spec spelling is the block name; the instance name is accepted
+                        // as a fallback so a request written the (common, non-conformant)
+                        // instance-qualified way resolves instead of failing the whole link.
+                        if (symbolType.getTypeName() != blockName.c_str() &&
+                            symbol->getName() != blockName.c_str()) {
+                            continue;
+                        }
+                        const glslang::TTypeList* members = symbolType.getStruct();
+                        if (members == nullptr) {
+                            continue;
+                        }
+                        for (SizeT m = 0; m < members->size(); ++m) {
+                            const glslang::TType* memberType = (*members)[m].type;
+                            if (memberType == nullptr || memberType->getFieldName() != memberName.c_str()) {
+                                continue;
+                            }
+                            capturedType = memberType;
+                            varying.blockMemberIndex = static_cast<Int>(m);
+                            break;
+                        }
+                        if (capturedType == nullptr) {
+                            // Right block, wrong member: no other linker object can match.
+                            break;
+                        }
+                        varying.blockName = symbolType.getTypeName().c_str();
+                        varying.blockInstanceName = symbol->getName().c_str();
                     }
-                    resolved = ResolveXfbSymbolType(symbol->getType(), varying.type, varying.size, bytesPerElement);
+                    resolved = ResolveXfbSymbolType(*capturedType, varying.type, varying.size, bytesPerElement);
                     if (resolved && singleElement) {
                         if (static_cast<Int>(element) >= varying.size) {
                             resolved = false;
                             break;
                         }
                         varying.size = 1;
+                        if (varying.blockMemberIndex >= 0) {
+                            varying.blockMemberElement = static_cast<Int>(element);
+                        }
                     }
                     break;
                 }

MobileGL/MG_State/GLState/ProgramState/ProgramLinkTask.h

@@ -29,10 +29,16 @@ namespace MobileGL::MG_State::GLState {
         SharedPtr<const ShaderCompileTask> compiled;
     };
 
-    // The unit of asynchronous linking: one glLinkProgram's worth of pure CPU work - glslang
-    // link + mapIO, SPIR-V generation and optimization, the GL-facing reflection surface, the
-    // global-UBO routing tables, fragment-output validation and transform-feedback
-    // resolution - with every input it needs snapshotted at enqueue.
+    // PHASE A of one glLinkProgram: the half that decides what GL can be asked about the
+    // program - glslang link + mapIO, the GL-facing reflection surface, fragment-output
+    // validation and transform-feedback resolution - with every input it needs snapshotted at
+    // enqueue.
+    //
+    // Every one of the eight ways a link can fail lives here, so once this node has published
+    // through EnsureLinkJoined() the program's LINK_STATUS, info log and entire query surface
+    // are FINAL and truthful. SPIR-V generation, spirv-opt and the global-UBO routing tables
+    // moved to ProgramSpirvTask, which chains behind this node and is joined by only five
+    // getters (see ProgramObject::EnsureSpirvJoined).
     //
     // Same ownership rule as ShaderCompileTask: the body reads nothing but `in` (all of it
     // owned or immutable) and writes nothing but `artifacts`. No GL call, no
@@ -40,11 +46,13 @@ namespace MobileGL::MG_State::GLState {
     // through the CompileEnv snapshot and diagnostics are deferred to the join.
     //
     // ONE LINK IS ONE HANDLER. RunBody() runs start to finish inside a single pool handler
-    // and is the only place `artifacts` is written. Do not split it across handlers to
-    // "pipeline" the reflection half: the intermediates that GlslangToSpv and buildReflection
-    // share are mutated in a strict order (see the GenerateSpirv-before-DoReflection comment
-    // in Run()), and a second handler would let a cancel land between them and publish a
-    // program whose SPIR-V and reflection describe different things.
+    // and is the only place `artifacts` is written. Splitting it across handlers to
+    // "pipeline" the reflection half would let a cancel land between the halves and publish a
+    // program whose SPIR-V and reflection describe different things - so any such split has
+    // to be structural: the first half must publish a LINK_STATUS and a query surface that
+    // are already final, and a lost second half must degrade to "linked but not drawable",
+    // never to a half-published program. (The intermediates' ordering constraint that used to
+    // be quoted here is retested and no longer binding; see the ordering note in RunBody.)
     class ProgramLinkTask final : public MG_Util::Async::JobNode {
     public:
         // ---- inputs, snapshotted on the GL thread in ProgramObject::Link()'s prologue ----
@@ -68,6 +76,49 @@ namespace MobileGL::MG_State::GLState {
         // Moved (never copied) into the ProgramObject by EnsureLinkJoined().
         ProgramObject::LinkArtifacts artifacts;
 
+        // ---- output: everything ProgramSpirvTask needs to run without this node's
+        //      artifacts, filled at the tail of a successful RunBody() ----
+        //
+        // THIS IS NOT `artifacts` AND MUST NOT BE MERGED INTO IT. The GL thread MOVES
+        // `artifacts` out of this node at the join, and phase B runs on a worker afterwards -
+        // so phase B may read `spirvHandoff` and `in` (neither is ever touched by the join)
+        // and this node's JobState, and nothing else on it. Reading `artifacts` or
+        // `diagnostics` from phase B would race the publish.
+        struct SpirvHandoff {
+            // MANDATORY, and the reason this struct exists at all: TProgram::addShader stores
+            // a RAW TShader*, and for the one-shader-per-stage case getIntermediate() returns
+            // the TShader's own intermediate rather than a copy. These used to die when
+            // RunBody() returned, which was safe only because nothing called getIntermediate()
+            // afterwards. GlslangToSpv does exactly that, so phase B has to own them.
+            //
+            // MEMORY NOTE: this is the one thing the split makes live LONGER than it used to -
+            // a glslang arena per stage, megabytes for a shaderpack, now alive from the end of
+            // phase A until phase B runs instead of dying with the link body, so a deep
+            // phase-B backlog holds one arena per queued program. Phase B clears this vector
+            // as soon as GlslangToSpv returns, but read that call site's comment before
+            // relying on it: for the COMMON case (a shader linked into exactly one program)
+            // the compile node co-owns the same TShader and phase A pins that node, so the
+            // clear frees nothing and only the re-parsed CAS-loser shaders are actually
+            // released. If peak RSS ever becomes the binding constraint on a pack load, THIS
+            // is the field to attack - by bounding the backlog, by releasing the compile
+            // node's own reference at claim time, or by moving GlslangToSpv back into phase A.
+            Vector<SharedPtr<glslang::TShader>> shaders;
+            // GL enum per entry of `in.shaders`, in the same order (GetSpirvBinaryFromProgram
+            // walks it to pick the intermediates).
+            Vector<GLenum> shaderTypes;
+            // The reflection slice BuildGlobalUboRouting consumes: {program, uniformLocations,
+            // uniformIndexInTProgram, tProgramUniformIndexToGl, maxUniformLocation}. Carried
+            // as a LinkArtifacts with only those five fields set, so the routing pass can keep
+            // calling ProgramObject::IsValidUniformLocation / GetUniformArraySizeByTIndex
+            // unchanged. The SharedPtr copy of `program` is also what keeps the TProgram alive
+            // for phase B after the join has moved `artifacts` away.
+            ProgramObject::LinkArtifacts reflection;
+
+            // The one flag phase B tests before doing anything: false means this link never
+            // reached the tail of RunBody (it failed, or was cancelled mid-body).
+            Bool ready = false;
+        } spirvHandoff;
+
         // Posts this job once every compile in `deps` is terminal - and not one moment
         // earlier, so the body never waits on anything (invariant I4: no job body may block
         // on another job, or the pool could deadlock with all its workers waiting on each
@@ -94,8 +145,6 @@ namespace MobileGL::MG_State::GLState {
         Bool ValidateFragmentOutputLocations();
         Bool ResolveTransformFeedbackVaryings();
         void ResolveGsTriangleStripCapture(const glslang::TIntermediate* captureIntermediate);
-        void GenerateSpirv();
-        void BuildGlobalUboRouting();
 
         // Worker-side MGLOG replacement: appended to diagnostics.logLines and replayed by the
         // join, on the GL thread, where a serial implementation would have printed it.

MobileGL/MG_State/GLState/ProgramState/ProgramObject.cpp

@@ -8,7 +8,9 @@
 
 #include "ProgramObject.h"
 #include "ProgramLinkTask.h"
+#include "ProgramSpirvTask.h"
 #include <atomic>
+#include <cstring>
 #include <MG_Util/Async/ShaderCompilePool.h>
 #include <MG_Util/Converters/GLToStr/GLEnumConverter.h>
 #include <MG_Util/ShaderTranspiler/CompileEnv.h>
@@ -68,12 +70,219 @@ namespace MobileGL::MG_State::GLState {
 
     Bool ProgramObject::IsPendingLinkTerminal() const { return m_pendingLink->IsTerminal(); }
 
+    Bool ProgramObject::IsPendingSpirvTerminal() const { return m_pendingSpirv->IsTerminal(); }
+
+    void ProgramObject::JoinPendingSpirv() const {
+        MOBILEGL_ASSERT(!MG_Util::Async::ShaderCompilePool::IsPoolThread(),
+                        "ProgramObject::EnsureSpirvJoined() reached from a pool thread; a job body must never read "
+                        "GL-thread-owned objects");
+
+        // Move the node out FIRST, for the same reason JoinPendingLink does: everything below
+        // runs GL-thread-only code that reads program state, and with m_pendingSpirv still set
+        // that would re-enter this function.
+        const SharedPtr<ProgramSpirvTask> pending = Move(m_pendingSpirv);
+        m_pendingSpirv.reset();
+
+        pending->Wait();
+        if (pending->IsComplete()) {
+            m_spirv = Move(pending->artifacts);
+        }
+        // A node that settled as Cancelled published nothing, so m_spirv stays empty with
+        // spirvStatus false: linked, queryable, not drawable. Nothing to repair.
+
+        // Order matters, and it is the GL order. The shadow arrives zero-filled; the shaders'
+        // declared uniform initializers are what it should actually start from, and only then
+        // do the application's own writes - the ones it made while the layout did not exist
+        // yet - land on top. Seeding after the replay would clobber them.
+        ApplyUniformInitialValues();
+        ReplayBufferedUniformWrites();
+
+        // The THIRD version bump of this link (enqueue, phase-A publish, phase-B publish), and
+        // it is mandatory for exactly the reason the phase-A one is (see JoinPendingLink): a
+        // backend memo taken during the A->B window - when the program was already answering
+        // as linked but had no SPIR-V and no uniform shadow - must not survive the arrival of
+        // either. The memos at risk are keyed on (lifetimeId, backendStateVersion).
+        BumpLinkObservableVersions();
+
+        MG_Util::Async::ApplyDeferredDiagnostics(*pending);
+    }
+
+    Bool ProgramObject::BufferUniformWrite(const Uint location, const SizeT byteOffsetInUniform, const void* source,
+                                           const SizeT byteSize) {
+        if (source == nullptr || byteSize == 0) return true; // nothing to record, nothing to join for
+        if (m_pendingUniformBytes.size() + byteSize > kMaxBufferedUniformBytes) {
+            // Pressure valve: stop growing and let the caller take the join. Say so once per
+            // program, because the interesting fact is WHICH program did it.
+            MGLOG_D("ProgramObject %u: buffered uniform writes exceeded %zu bytes during the SPIR-V window; the "
+                    "write joins instead",
+                    m_externalIndex, kMaxBufferedUniformBytes);
+            return false;
+        }
+        const SizeT dataOffset = m_pendingUniformBytes.size();
+        m_pendingUniformBytes.resize(dataOffset + byteSize);
+        std::memcpy(m_pendingUniformBytes.data() + dataOffset, source, byteSize);
+        m_pendingUniformWrites.push_back(PendingUniformWrite{.location = location,
+                                                             .byteOffsetInUniform =
+                                                                 static_cast<Uint>(byteOffsetInUniform),
+                                                             .byteSize = static_cast<Uint>(byteSize),
+                                                             .dataOffset = static_cast<Uint>(dataOffset)});
+        return true;
+    }
+
+    // "uniform vec3 v = vec3(10, 20, 30);" - legal desktop GLSL since 1.20, and the value is
+    // what the uniform reads until glUniform* replaces it (and again after every relink).
+    // MobileGL parses with Vulkan-relaxed rules, which sweep default-block uniforms into
+    // MGL_GLOBAL_UBO; a block member cannot carry an initializer in SPIR-V, so glslang hands
+    // the folded constants over as a side-channel (TIntermediate::getUniformInitializers) and
+    // this is where they are honoured. Without it every such uniform silently read zero -
+    // which is what half of KHR-GL43.shader_storage_buffer_object was actually failing on.
+    //
+    // Writes go straight into the shadow rather than through glUniform*: this runs INSIDE the
+    // phase-B publish, so re-entering the join gate is not available, and the location space
+    // reflection assigns (one location per array element) is all that is needed.
+    void ProgramObject::ApplyUniformInitialValues() const {
+        // Through the phase-A gate, not off m_artifacts directly: phase B can be joined by a
+        // caller that has not read anything phase A publishes yet, and reading the raw field
+        // there would find the PREVIOUS link's block (or an empty one) and drop every
+        // initializer without a trace. Artifacts() is a no-op once phase A is in.
+        const auto& initializers = Artifacts().uniformInitialValues;
+        if (initializers.empty()) return;
+        if (m_spirv.globalUboScratch.empty() || m_spirv.uniformOffsets.empty()) {
+            // Phase B published no shadow (cancelled, or superseded by a relink). The program
+            // is not drawable; there is nowhere for these to land.
+            return;
+        }
+
+        Uint8* const scratch = m_spirv.globalUboScratch.data();
+        const SizeT uboSize = m_spirv.globalUboScratch.size();
+
+        for (const auto& init : initializers) {
+            // Scalars per array ELEMENT. A matrix element carries cols * rows of them, laid
+            // out column by column - which is also the order glslang folded them in.
+            const Int columns = init.matrixCols;
+            const Int rows = init.matrixRows;
+            const Int componentsPerElement = columns > 0 ? columns * rows : init.vectorSize;
+            const Int elements = init.arraySize;
+            if (componentsPerElement <= 0 || elements <= 0) continue;
+
+            const Bool isFloat = init.basicType == glslang::EbtFloat || init.basicType == glslang::EbtFloat16;
+            const Bool isInt = init.basicType == glslang::EbtInt || init.basicType == glslang::EbtUint ||
+                               init.basicType == glslang::EbtBool;
+            // Anything else (fp64, 64-bit integers) has no 32-bit shadow encoding here, and a
+            // half-written uniform is worse than an untouched one.
+            if (!isFloat && !isInt) continue;
+            const SizeT provided = isFloat ? init.floatValues.size() : init.intValues.size();
+            if (provided < static_cast<SizeT>(componentsPerElement) * static_cast<SizeT>(elements)) continue;
+
+            const Int baseLocation = GetUniformLocation(init.name);
+            if (baseLocation < 0) continue; // optimized away, or not a default-block uniform
+
+            for (Int element = 0; element < elements; ++element) {
+                const Int location = baseLocation + element;
+                if (element > 0 && !UniformLocationsAliasSameUniform(baseLocation, location)) break;
+                if (!IsValidUniformLocation(location)) break;
+                const Uint offset = GetUniformOffset(static_cast<Uint>(location));
+                if (offset == kInvalidUniformOffset) continue;
+
+                // std140 pads every column of a float matrix out to a vec4, so the columns of
+                // a mat3 are 16 bytes apart even though each carries 12. The slot's own span
+                // states the stride the rest of the pipeline agreed on rather than guessing it.
+                const SizeT slotSpan = GetUniformStorageSpanInBytes(static_cast<Uint>(location));
+                const SizeT columnStride =
+                    columns > 0 ? slotSpan / static_cast<SizeT>(columns) : slotSpan;
+                const Int componentsPerColumn = columns > 0 ? rows : componentsPerElement;
+                const Int columnCount = columns > 0 ? columns : 1;
+
+                for (Int column = 0; column < columnCount; ++column) {
+                    const SizeT byteOffset = static_cast<SizeT>(offset) + static_cast<SizeT>(column) * columnStride;
+                    const SizeT writeSize = static_cast<SizeT>(componentsPerColumn) * sizeof(Uint32);
+                    if (byteOffset + writeSize > uboSize) break;
+                    const SizeT firstComponent = static_cast<SizeT>(element) * componentsPerElement +
+                                                 static_cast<SizeT>(column) * componentsPerColumn;
+                    for (Int component = 0; component < componentsPerColumn; ++component) {
+                        const SizeT source = firstComponent + static_cast<SizeT>(component);
+                        Uint8* const destination = scratch + byteOffset + component * sizeof(Uint32);
+                        if (isFloat) {
+                            const Float value = static_cast<Float>(init.floatValues[source]);
+                            std::memcpy(destination, &value, sizeof(value));
+                        } else {
+                            const Int32 value = static_cast<Int32>(init.intValues[source]);
+                            std::memcpy(destination, &value, sizeof(value));
+                        }
+                    }
+                }
+            }
+        }
+        MarkUBOContentDirty();
+    }
+
+    void ProgramObject::ReplayBufferedUniformWrites() const {
+        if (m_pendingUniformWrites.empty()) {
+            m_pendingUniformBytes.clear();
+            return;
+        }
+
+        // Drain into locals first: MarkUBOContentDirty below is a plain counter bump, but a
+        // future reader of this function should not be able to observe a half-drained buffer.
+        Vector<PendingUniformWrite> writes;
+        Vector<Uint8> bytes;
+        writes.swap(m_pendingUniformWrites);
+        bytes.swap(m_pendingUniformBytes);
+
+        if (m_spirv.globalUboScratch.empty() || m_spirv.uniformOffsets.empty()) {
+            // Phase B produced nothing (cancelled at teardown, or a relink superseded it).
+            // The program is not drawable, so there is nowhere for these to land and nothing
+            // that could observe them.
+            MGLOG_D("ProgramObject %u: dropping %zu buffered uniform write(s); the SPIR-V job published no shadow",
+                    m_externalIndex, writes.size());
+            return;
+        }
+
+        Uint8* const scratch = m_spirv.globalUboScratch.data();
+        const SizeT uboSize = m_spirv.globalUboScratch.size();
+        for (const PendingUniformWrite& write : writes) {
+            if (write.location >= m_spirv.uniformOffsets.size()) continue;
+            const Uint offset = m_spirv.uniformOffsets[write.location];
+            if (offset == kInvalidUniformOffset ||
+                static_cast<SizeT>(offset) + write.byteOffsetInUniform + write.byteSize > uboSize) {
+                // Same verdict the live write path reaches for a uniform without backing
+                // storage: log and drop, rather than fault.
+                MGLOG_E("ProgramObject %u: buffered uniform write at location %u has no backing storage "
+                        "(offset=%u size=%u uboSize=%zu); dropping write",
+                        m_externalIndex, write.location, offset, write.byteSize, uboSize);
+                continue;
+            }
+            Uint8* const destination = scratch + offset + write.byteOffsetInUniform;
+            const Uint8* const sourceBytes = bytes.data() + write.dataOffset;
+            // The same bytes-equal dedupe the live path applies, per record and in order, so
+            // the "an identical write does not move the content version" property survives
+            // the detour byte for byte.
+            if (std::memcmp(destination, sourceBytes, write.byteSize) == 0) continue;
+            std::memcpy(destination, sourceBytes, write.byteSize);
+            MarkUBOContentDirty();
+        }
+    }
+
     void ProgramObject::CancelLink() {
+        // Phase B first: it is chained behind phase A, so cancelling A would otherwise run A's
+        // continuation and post a node this call is about to abandon anyway. Cancelling it up
+        // front makes that continuation a no-op.
+        //
+        // Cooperative and non-blocking, both of them. A node that no worker has picked up
+        // settles immediately; one that is running is flagged and settles when its body
+        // returns, writing only into itself the whole time. Either way nothing waits, and each
+        // node keeps its own inputs alive for as long as it needs them.
+        if (m_pendingSpirv) {
+            m_pendingSpirv->Cancel();
+            m_pendingSpirv.reset();
+            // Buffered writes belong to the link that is being abandoned. A relink resets
+            // every uniform to its initial value anyway (GL 4.6 core 7.6), and the other two
+            // callers are destruction and glProgramBinary's mandated failure, so there is
+            // nothing left that could want them.
+            m_pendingUniformWrites.clear();
+            m_pendingUniformBytes.clear();
+        }
         if (!m_pendingLink) return;
-        // Cooperative and non-blocking. A node that no worker has picked up settles
-        // immediately; one that is running is flagged and settles when its body returns,
-        // writing only into itself the whole time. Either way nothing waits, and the node
-        // keeps its own inputs alive for as long as it needs them.
         m_pendingLink->Cancel();
         m_pendingLink.reset();
     }
@@ -103,23 +312,30 @@ namespace MobileGL::MG_State::GLState {
         // function has ever cleared, and its callers depend on that (they write infoLog
         // immediately AFTER calling here). Link()'s prologue does not use this - it assigns a
         // whole default-constructed block, where the ordering is explicit.
+        // Phase-B output (generatedSpirv / uniformOffsets / globalUboScratch) is NOT cleared
+        // here and is not in LinkArtifacts at all: the link body calls this on its own block,
+        // where no phase-B output exists yet. The two GL-thread callers that also have to
+        // discard phase-B output say so themselves (MarkLinkFailedByProgramBinary clears
+        // m_spirv; Link()'s prologue assigns a fresh one).
         artifacts.program.reset();
-        artifacts.generatedSpirv.clear();
         artifacts.uniformLocations.clear();
         artifacts.glUniformIndexToTProgram.clear();
         artifacts.tProgramUniformIndexToGl.clear();
         artifacts.glBlockIndexToTProgram.clear();
         artifacts.tProgramBlockIndexToGl.clear();
         artifacts.linkedExplicitUniformLocations.clear();
+        artifacts.uniformInitialValues.clear();
         artifacts.uniformIndexInTProgram.clear();
+        // GL resets every uniform to its initial value at link, so nothing is "written since
+        // link" any more - and the locations these bits index no longer mean anything either.
+        artifacts.writtenUniformLocationBits.clear();
+        artifacts.writtenUniformIndexBits.clear();
+        artifacts.writtenUniformIndices.clear();
         artifacts.uniformSamplerOrImageUnitIndex.clear();
         artifacts.explicitOpaqueUniformBindings.clear();
         artifacts.uniformBlockIndexByName.clear();
         artifacts.uniformBlockBinding.clear();
         artifacts.shaderStorageBlockBinding.clear();
-        artifacts.uniformOffsets.clear();
-        artifacts.uniformSizesInBytes.clear();
-        artifacts.globalUboScratch.clear();
         artifacts.attribs.clear();
         artifacts.attribTypes.clear();
         artifacts.activeUniformCount = 0;
@@ -238,6 +454,7 @@ namespace MobileGL::MG_State::GLState {
         // is what every gated reader sees, so it has to be the complete "not linked" state -
         // including the fields ResetLinkArtifacts deliberately preserves for its own callers.
         m_artifacts = {};
+        m_spirv = {};
 
         // ---- GL-thread-owned mutations ----
         // Remove detached shaders first
@@ -292,17 +509,33 @@ namespace MobileGL::MG_State::GLState {
             task->in.shaders.push_back({shader->GetShaderStage(), shader->GetShaderSourcePtr(), node});
         }
 
+        // Phase B of the same link: SPIR-V generation, spirv-opt and the global-UBO routing
+        // tables. Created here, alongside phase A, so that from this instant the program has
+        // BOTH pending nodes and every cancel site (this prologue, ~ProgramObject,
+        // glProgramBinary's failure) drops both through the one CancelLink().
+        auto spirvTask = MakeShared<ProgramSpirvTask>();
         m_pendingLink = task;
+        m_pendingSpirv = spirvTask;
 
         // Flag off - or glMaxShaderCompilerThreadsKHR(0), see AsyncShaderCompileActive():
         // byte-identical to the synchronous implementation. RunInline() executes the same
-        // body on this thread and the join below publishes through the same code, so the two
-        // modes differ only in WHICH thread ran RunBody().
+        // bodies on this thread, in the same order, and the join below publishes through the
+        // same code, so the two modes differ only in WHICH thread ran them.
+        //
+        // Deliberately NOT expressed as SubmitAfter here: its continuation posts to the pool,
+        // and in this mode the pool is merely unused rather than stopped - the work would
+        // silently move off-thread in the one mode whose whole contract is that it does not.
         if (!MG_Util::Async::AsyncShaderCompileActive()) {
             task->RunInline();
-            EnsureLinkJoined();
+            spirvTask->RunInlineAfter(task);
+            EnsureSpirvJoined();
             return;
         }
+        // The chain edge FIRST, while phase A is still Pending, so registering it is a plain
+        // list append rather than an inline continuation on this thread. If SubmitAfter below
+        // then fails to post phase A it cancels it, and that cancel fires this edge, which
+        // cancels phase B - nothing is left stranded either way.
+        spirvTask->SubmitAfter(task);
         task->SubmitAfter(deps);
     }
 

MobileGL/MG_State/GLState/ProgramState/ProgramObject.h

@@ -18,6 +18,9 @@ namespace MobileGL::MG_State::GLState {
     // ProgramLinkTask.h includes THIS header (it outputs a LinkArtifacts), so including it
     // back would be circular. The destructor is therefore out of line.
     class ProgramLinkTask;
+    // Phase B of the same link: SPIR-V generation, spirv-opt and the global-UBO routing
+    // tables. Chained behind the ProgramLinkTask, forward-declared for the same reason.
+    class ProgramSpirvTask;
 
     class ProgramObject {
     public:
@@ -303,8 +306,133 @@ namespace MobileGL::MG_State::GLState {
         // Sentinel for a uniform location without global-UBO backing storage (should not
         // survive linking: GenerateBinary falls back to tail-allocated scratch storage).
         static constexpr Uint kInvalidUniformOffset = ~0u;
-        Uint GetUniformOffset(Uint location) const { return Artifacts().uniformOffsets[location]; }
+        // PHASE B (joins the SPIR-V job; see EnsureSpirvJoined).
+        //
+        // BOUNDS-CHECKED, and that is not defensive padding - it is the load-bearing half of
+        // the "linked but not drawable" contract. A phase B that settles CANCELLED rather than
+        // Complete (its body threw, the pool failed to enqueue it, or teardown cancelled it
+        // while phase A had already published) publishes nothing, so the shadow is a
+        // default-constructed SpirvArtifacts with an EMPTY uniformOffsets - while LINK_STATUS
+        // stays GL_TRUE, because GL gives no way to retract one, and IsValidUniformLocation()
+        // keeps answering true out of phase-A reflection. Every glUniform*/glGetUniform* call
+        // site reaches this getter BEFORE its own kInvalidUniformOffset / null-scratch guard,
+        // so an unchecked operator[] here would be a null dereference on the query surface
+        // this design promises stays answerable. Reporting kInvalidUniformOffset instead hands
+        // each of those sites exactly the value their existing guard already handles - the
+        // same value the routing pass itself uses for a uniform the optimizer deleted.
+        Uint GetUniformOffset(Uint location) const {
+            const SpirvArtifacts& spirv = Spirv();
+            return location < spirv.uniformOffsets.size() ? spirv.uniformOffsets[location]
+                                                          : kInvalidUniformOffset;
+        }
         Uint GetUniformSizesInBytes(Uint location) const { return MG_Util::GetGLTypeSize(GetUniformType(location)); }
+        // Bytes a uniform actually occupies in the global UBO, which is not its GL type size:
+        // std140 pads each column of a float matrix out to a vec4, so a mat3 spans 48 bytes
+        // even though only 36 of them carry components. Anything reading or writing a whole
+        // uniform's storage - a bounds check, a copy between two programs' shadows - wants
+        // this rather than GetUniformSizesInBytes.
+        static SizeT UniformStorageSpanInBytes(const glslang::TType* type, SizeT tightSize) {
+            if (type != nullptr && type->isMatrix() && type->getBasicType() != glslang::EbtDouble) {
+                return static_cast<SizeT>(type->getMatrixCols()) * 4 * sizeof(Float);
+            }
+            return tightSize;
+        }
+        SizeT GetUniformStorageSpanInBytes(Uint location) const {
+            return UniformStorageSpanInBytes(GetUniformTType(location), GetUniformSizesInBytes(location));
+        }
+
+        // ---- "written since link": the per-location dirty set the pipeline composite mirrors from ----
+        //
+        // A pipeline's stage programs each own their uniform storage, but the composite the draw
+        // goes through has ONE slot per name. Mirroring every active uniform of every stage
+        // therefore lets the last stage that merely DECLARES a name overwrite the value an
+        // earlier stage was actually written with - the shared-header idiom (the same
+        // `uniform mat4 u_mvp` in the VS and the FS) rendered nothing because of it. Recording
+        // which locations an application has written is what lets the mirror carry only those.
+        //
+        // WHO PAYS: only a program that could ever be a pipeline stage, decided by the latch
+        // below. glUseProgram's uniform path - thousands of calls per frame in Minecraft - pays
+        // one predictable bool branch and nothing else.
+        //
+        // GRANULARITY is per LOCATION, not per name: glUniform*v writes array elements at
+        // element locations, and a program that wrote `arr[3]` and nothing else must mirror
+        // exactly that element. The compact index list beside it is what keeps the mirror
+        // O(uniforms actually written) instead of O(active uniforms) - it is the set of GL
+        // active-uniform indices owning at least one written location, so the mirror does its
+        // two name lookups once per written uniform rather than once per uniform in the program.
+        //
+        // NOT counted as a write: the declared initializers ProgramLinkTask seeds at link
+        // (ApplyUniformInitialValues). They are a property of the SHADERS, and the composite
+        // links the very same shader objects, so it seeds itself with the identical values -
+        // there is nothing to carry. Counting them would also re-introduce the bug this set
+        // exists to fix, by letting a stage that only declares `uniform float f = 0.0;` clobber
+        // the value the application wrote for `f` in another stage.
+        Bool TracksUniformWrites() const { return m_tracksUniformWrites; }
+
+        // Generation of the write SET itself, as distinct from the values in it. The refresh
+        // gate (ProgramPipelineObject::ComputeUniformMirrorVersions) is otherwise built out of
+        // counters that only move when BYTES move - and a write can enlarge the set without
+        // moving a byte, because both write funnels drop a value-identical write before
+        // bumping anything. glProgramUniform1f(fs, f, 0.0f) on an `f` that already reads 0.0
+        // is exactly that: it makes the FRAGMENT stage the last written-to stage for `f`, so
+        // the composite must be re-mirrored to hand it the slot, and nothing else in the gate
+        // would have noticed.
+        Uint32 GetUniformWriteSetVersion() const { return m_uniformWriteSetVersion; }
+
+        // Records that `location` has been written since the last link. Cheap and idempotent;
+        // a no-op on a program that can never be a pipeline stage.
+        void MarkUniformWrittenAtLocation(Uint location) {
+            if (!m_tracksUniformWrites) return;
+            LinkArtifacts& artifacts = Artifacts();
+            if (!IsValidUniformLocation(artifacts, static_cast<Int>(location))) return;
+
+            // Sized to cover this location AND the whole location space, so a program whose
+            // highest location is written first does not reallocate on every later write, and
+            // so the subscript below needs no second guard: the vector provably contains it.
+            const SizeT locationWord = location / 64u;
+            if (locationWord >= artifacts.writtenUniformLocationBits.size()) {
+                artifacts.writtenUniformLocationBits.resize(
+                    std::max<SizeT>(locationWord + 1u, static_cast<SizeT>(artifacts.maxUniformLocation) / 64u + 1u),
+                    0u);
+            }
+            const Uint64 locationBit = Uint64{1} << (location % 64u);
+            if ((artifacts.writtenUniformLocationBits[locationWord] & locationBit) == 0) {
+                artifacts.writtenUniformLocationBits[locationWord] |= locationBit;
+                // Only on the 0 -> 1 transition: a re-write of a location already in the set
+                // changes nothing the mirror would do differently, and moving the version for
+                // it would re-walk the set on every repeated glUniform* call.
+                ++m_uniformWriteSetVersion;
+            }
+
+            // Add the owning GL active-uniform index to the compact list, once.
+            const Int tIndex = artifacts.uniformIndexInTProgram[location];
+            if (tIndex < 0 || static_cast<SizeT>(tIndex) >= artifacts.tProgramUniformIndexToGl.size()) return;
+            const Int glIndex = artifacts.tProgramUniformIndexToGl[tIndex];
+            // -1 is a uniform the relaxed parse swept out of the GL-visible index space; the
+            // mirror enumerates GL indices, so there is nothing it could look such a one up by.
+            if (glIndex < 0) return;
+            const SizeT indexWord = static_cast<SizeT>(glIndex) / 64u;
+            if (indexWord >= artifacts.writtenUniformIndexBits.size()) {
+                artifacts.writtenUniformIndexBits.resize(
+                    std::max<SizeT>(indexWord + 1u, static_cast<SizeT>(artifacts.activeUniformCount) / 64u + 1u), 0u);
+            }
+            const Uint64 indexBit = Uint64{1} << (static_cast<SizeT>(glIndex) % 64u);
+            if ((artifacts.writtenUniformIndexBits[indexWord] & indexBit) != 0) return;
+            artifacts.writtenUniformIndexBits[indexWord] |= indexBit;
+            artifacts.writtenUniformIndices.push_back(static_cast<Uint>(glIndex));
+        }
+
+        Bool IsUniformWrittenAtLocation(Uint location) const {
+            const auto& bits = Artifacts().writtenUniformLocationBits;
+            const SizeT locationWord = location / 64u;
+            return locationWord < bits.size() &&
+                   (bits[locationWord] & (Uint64{1} << (location % 64u))) != 0;
+        }
+
+        // GL active-uniform indices owning at least one written location. Empty for every
+        // program that has not been written to since its last link - and for every program
+        // that never asked to be separable, which is what makes the mirror free for them.
+        const Vector<Uint>& GetWrittenUniformIndices() const { return Artifacts().writtenUniformIndices; }
 
         Int GetAttributeLocation(const String& name) {
             const auto it = std::find(Artifacts().attribs.begin(), Artifacts().attribs.end(), name);
@@ -381,9 +509,14 @@ namespace MobileGL::MG_State::GLState {
         const String& GetActiveAttribName(Uint index) const {
             return NormalizeBuiltinPipeInputName(Artifacts().program->getPipeInput(static_cast<Int>(index)).name);
         }
-        void* MapUBO() { return Artifacts().globalUboScratch.data(); }
-        const void* GetUBOData() const { return Artifacts().globalUboScratch.data(); }
-        Uint GetUBOSize() const { return static_cast<Uint>(Artifacts().globalUboScratch.size()); }
+        // PHASE B, all three (see EnsureSpirvJoined): the shadow buffer's layout is decided
+        // by the OPTIMIZED SPIR-V, so it does not exist until the SPIR-V job has settled - and
+        // never exists at all for a program whose SPIR-V job settled cancelled. These three
+        // degrade to nullptr/nullptr/0 in that case, which is exactly the "no backing storage"
+        // shape every caller already tests for (see GetUniformOffset's note).
+        void* MapUBO() { return Spirv().globalUboScratch.data(); }
+        const void* GetUBOData() const { return Spirv().globalUboScratch.data(); }
+        Uint GetUBOSize() const { return static_cast<Uint>(Spirv().globalUboScratch.size()); }
         // Content version of the CPU-side global-UBO shadow: writers bump it so backends
         // can skip re-uploading an unchanged UBO on every draw. ~0u is reserved as the
         // backends' "never uploaded" sentinel, so skip over it on wrap.
@@ -391,6 +524,25 @@ namespace MobileGL::MG_State::GLState {
         void MarkUBOContentDirty() const {
             if (++m_uboContentVersion == ~0u) m_uboContentVersion = 0;
         }
+        // ---- glUniform* inside the phase-A -> phase-B window ----
+        //
+        // True while the program is fully linked and fully queryable but its uniform shadow's
+        // LAYOUT (which the optimized SPIR-V decides) does not exist yet. A non-opaque
+        // glUniform* write in that window is RECORDED rather than joined, and replayed into
+        // the shadow at the phase-B publish - so a pack that sets its uniforms immediately
+        // after glLinkProgram never waits for SPIR-V.
+        //
+        // Nothing can observe the difference: the only route to those bytes is glGetUniform*
+        // (and a draw), and both of those go through the phase-B gate, which replays first.
+        // The OPAQUE branch of glUniform* is deliberately not buffered - a sampler unit is
+        // phase-A state (uniformSamplerOrImageUnitIndex), so glUniform1i(samplerLoc, unit)
+        // right after a link stays a zero-join operation, which is exactly what Iris does.
+        Bool IsSpirvPending() const { return m_pendingSpirv != nullptr; }
+        // Records one write. Returns false if it declined to buffer - the caller must then
+        // perform the write directly (which joins). Declining is the pressure valve for an
+        // application that writes megabytes of uniforms into a single pending window.
+        Bool BufferUniformWrite(Uint location, SizeT byteOffsetInUniform, const void* source, SizeT byteSize);
+
         Uint32 GetBackendStateVersion() const { return m_backendStateVersion; }
         // Bumped only by (re)linking — lets backends detect that every piece of
         // link-derived reflection (locations, block order, UBO layout) is stale.
@@ -429,14 +581,35 @@ namespace MobileGL::MG_State::GLState {
         }
 
         void SetUniformSamplerOrImageUnitIndex(Uint location, Int unit) {
-            if (location >= Artifacts().uniformSamplerOrImageUnitIndex.size() ||
-                Artifacts().uniformSamplerOrImageUnitIndex[location] == unit) {
-                return;
-            }
+            if (location >= Artifacts().uniformSamplerOrImageUnitIndex.size()) return;
+            // BEFORE the equality bail-out, not after: "written" is about the application
+            // having addressed the uniform, not about the bytes changing. glUniform1i(s, 0) on
+            // a sampler that already reads 0 still has to beat another stage's untouched
+            // declaration of the same name in the composite - which is only possible if the
+            // write is recorded. (The mirror is the only reader, and it runs this same setter
+            // on the composite, where the latch is off.)
+            MarkUniformWrittenAtLocation(location);
+            if (Artifacts().uniformSamplerOrImageUnitIndex[location] == unit) return;
             Artifacts().uniformSamplerOrImageUnitIndex[location] = unit;
             ++m_backendStateVersion;
+            // IMAGE units get their own generation, and it is not redundant with the one
+            // above. A sampler unit is re-issued to the driver per draw as a plain
+            // glUniform1i, so a backend can honour a change without rebuilding anything; an
+            // image unit cannot be, because ES forbids glUniform1i on image uniforms - Espryt
+            // has to BAKE it into the ESSL it generates (RebindImageUniformsToFrontendUnits),
+            // which means the change is only honoured by regenerating the program. That
+            // regeneration is gated on link-shaped versions, so without a counter that moves
+            // here the new unit would never reach the driver.
+            if (const glslang::TType* type = GetUniformTType(location); type != nullptr && type->isImage()) {
+                ++m_imageUnitVersion;
+            }
         }
 
+        // Generation of the image-uniform unit assignment; see SetUniformSamplerOrImageUnitIndex.
+        // A backend that compiles the unit into its program source compares this to decide
+        // whether what it built is still describing the right binding.
+        Uint32 GetImageUnitVersion() const { return m_imageUnitVersion; }
+
         Int GetUniformSamplerOrImageUnitIndex(Uint location) const {
             return Artifacts().uniformSamplerOrImageUnitIndex[location];
         }
@@ -452,7 +625,32 @@ namespace MobileGL::MG_State::GLState {
         // subset of the stages of a program pipeline. Only takes effect on the next link,
         // which is why it is plain state here rather than something Link() consults.
         Bool GetSeparable() const { return m_separable; }
-        void SetSeparable(Bool separable) { m_separable = separable; }
+        void SetSeparable(Bool separable) {
+            m_separable = separable;
+            // ---- arming the uniform-write tracking latch ----
+            //
+            // The predicate wanted is "this program can ever be a pipeline stage", and
+            // GetSeparable() is NOT it in either direction. GL_PROGRAM_SEPARABLE takes effect
+            // at the NEXT link, so it can read true on a program glUseProgramStages would
+            // still reject; that direction is merely wasteful. The other direction is a
+            // correctness hole: glProgramParameteri may clear the flag AFTER a separable link,
+            // and glUseProgramStages tests the state the program was LINKED with, so such a
+            // program is still a legal stage while GetSeparable() reads false. Tracking driven
+            // by the live flag would stop recording writes on a program the composite is still
+            // mirroring from, and those uniforms would silently stop reaching the draw.
+            //
+            // "Attached to a pipeline" is not usable either, and for a more basic reason:
+            // glProgramUniform* legitimately runs before glUseProgramStages, so the marks have
+            // to already exist by the time the program becomes a stage.
+            //
+            // So: a MONOTONE latch, armed the first time GL_PROGRAM_SEPARABLE is requested
+            // true and never cleared. It over-approximates - a program that was separable once
+            // keeps paying the bookkeeping - and over-approximating only ever costs a bitset,
+            // never a wrong value. glCreateShaderProgramv arms it through this same setter.
+            // A program that never asks (every monolithic glUseProgram program, which is the
+            // hot uniform path) never arms it and pays one bool branch per glUniform*.
+            if (separable) m_tracksUniformWrites = true;
+        }
         // glProgramBinary always fails here (there is no format it could accept) and the
         // spec then requires the program's LINK_STATUS to read FALSE.
         void MarkLinkFailedByProgramBinary() {
@@ -463,15 +661,35 @@ namespace MobileGL::MG_State::GLState {
             CancelLink();
             BumpLinkObservableVersions();
             ResetLinkArtifacts(Artifacts());
+            // ResetLinkArtifacts is a LinkArtifacts-only operation (the link body calls it on
+            // its own block, where no phase-B output exists yet), so the phase-B half is
+            // cleared here. CancelLink() above already dropped the pending SPIR-V job, so
+            // this cannot be racing a publish.
+            m_spirv = {};
             Artifacts().infoLog = "No program binary format is supported.";
         }
         Bool GetValidateStatus() const { return m_validateStatus; }
-        Int GetActiveAtomicCounterCount() const { return Artifacts().program->getNumAtomicCounters(); }
-        Int GetActiveAttributesCount() const { return Artifacts().program->getNumPipeInputs(); }
+        // Artifacts().program is null until a link produces reflection, and glGetProgramiv is
+        // perfectly legal on a program that never linked (GL 4.6 sec. 7.3: the queried state is
+        // simply its initial value, zero). Dereferencing it there took the process down with a
+        // SIGSEGV inside glslang::TProgram::getNumPipeInputs - KHR-GL30.api.coverage does exactly
+        // this after a failed glGetAttribLocation, and reached it as soon as the CopyTexImage2D
+        // throw ahead of it stopped killing the run first.
+        Int GetActiveAtomicCounterCount() const {
+            const auto& program = Artifacts().program;
+            return program ? program->getNumAtomicCounters() : 0;
+        }
+        Int GetActiveAttributesCount() const {
+            const auto& program = Artifacts().program;
+            return program ? program->getNumPipeInputs() : 0;
+        }
         // GL-visible uniform blocks only: the synthesized MGL_GLOBAL_UBO the relaxed parse
         // materializes for default-block uniforms is filtered out by DoReflection.
         Int GetActiveUniformBlocksCount() const { return static_cast<Int>(Artifacts().glBlockIndexToTProgram.size()); }
-        GLuint GetComputeLocalSize(Uint dim) const { return Artifacts().program->getLocalSize(static_cast<Int>(dim)); }
+        GLuint GetComputeLocalSize(Uint dim) const {
+            const auto& program = Artifacts().program;
+            return program ? program->getLocalSize(static_cast<Int>(dim)) : 0;
+        }
         Int GetActiveAttributesMaxLength() const { return Artifacts().attribInNameMaxLength; }
         Int GetActiveUniformBlocksMaxNameLength() const { return Artifacts().uniformBlockNameMaxLength; }
         Uint GetUniformBlockIndex(const char* name) const {
@@ -534,13 +752,23 @@ namespace MobileGL::MG_State::GLState {
             return (ubo.stages & stageMask) != 0;
         }
 
-        // Set by glUniformBlockBinding
+        // Bumped by both block-binding setters below. A program pipeline's flattened composite
+        // is a different program object from the stage programs the application rebinds blocks
+        // on, so it has to be told - and this is what tells it something is worth re-reading.
+        // Separate from m_backendStateVersion because the storage-block setter deliberately
+        // does not disturb that one (see SetShaderStorageBlockBinding).
+        Uint32 GetBlockBindingVersion() const { return m_blockBindingVersion; }
+
+        // Set by glUniformBlockBinding. The vector is seeded at link with each block's DECLARED
+        // binding (layout(binding=N), else -1), so an untouched program already reports what its
+        // shaders asked for.
         void SetUniformBlockBinding(Uint index, Uint binding) {
             if (index >= Artifacts().uniformBlockBinding.size() || Artifacts().uniformBlockBinding[index] == static_cast<Int>(binding)) {
                 return;
             }
             Artifacts().uniformBlockBinding[index] = static_cast<Int>(binding);
             ++m_backendStateVersion;
+            ++m_blockBindingVersion;
         }
 
         Uint GetUniformBlockBinding(Uint index) const { return Artifacts().uniformBlockBinding[index]; }
@@ -552,6 +780,10 @@ namespace MobileGL::MG_State::GLState {
         // means "never rebound", and the shader's declared binding still stands.
         void SetShaderStorageBlockBinding(const String& blockName, Uint binding) {
             Artifacts().shaderStorageBlockBinding[blockName] = static_cast<Int>(binding);
+            // Deliberately NOT m_backendStateVersion: Espryt's entry point never forces a
+            // program build off this, and bumping that version would start doing so. The
+            // dedicated counter carries the news to the pipeline composite instead.
+            ++m_blockBindingVersion;
         }
         // -1 when the block has never been rebound. `blockName` is the interface-query
         // spelling; an arrayed block's elements ("B[0]", "B[1]") are separate GL resources
@@ -571,8 +803,15 @@ namespace MobileGL::MG_State::GLState {
             return Artifacts().shaderStorageBlockBinding;
         }
 
-        Vector<Vector<unsigned>>& GetGeneratedSpirv() { return Artifacts().generatedSpirv; }
-        const Vector<Vector<unsigned>>& GetGeneratedSpirv() const { return Artifacts().generatedSpirv; }
+        // PHASE B (see EnsureSpirvJoined). Empty for a program whose SPIR-V job was
+        // cancelled; GetSpirvStatus() below is how a backend tells that apart from a program
+        // that never linked.
+        Vector<Vector<unsigned>>& GetGeneratedSpirv() { return Spirv().generatedSpirv; }
+        const Vector<Vector<unsigned>>& GetGeneratedSpirv() const { return Spirv().generatedSpirv; }
+        // Whether phase B produced usable SPIR-V. Joins, like the four getters above: a
+        // backend asks this exactly where it used to ask GetLinkStatus(), i.e. right before
+        // it builds or draws with the program.
+        Bool GetSpirvStatus() const { return Spirv().spirvStatus; }
 
         // The linked glslang reflection itself, for the ONE consumer that needs resource
         // lists no typed getter above exposes: the GL program-interface query layer
@@ -600,6 +839,21 @@ namespace MobileGL::MG_State::GLState {
             // Offset within the gap-free record a backend that cannot express the GL
             // layout captures into; see NeedsScatteredTransformFeedbackCapture.
             Uint32 packedOffsetBytes = 0;
+
+            // GL 4.6 core 11.1.2.1 / 7.3.1.1: a member of an output interface block is
+            // captured under "<block name>.<member>". `name` keeps that GL spelling (it is
+            // what the interface queries and the ESSL backend's driver-side capture list
+            // need, since SPIRV-Cross re-emits the block under its own type name), while
+            // the three fields below carry what a SPIR-V backend needs instead: the
+            // decoration target is the block's *instance* variable and the member index
+            // inside it. blockMemberIndex < 0 means "not a block member".
+            String blockInstanceName;
+            String blockName;
+            Int blockMemberIndex = -1;
+            // Which element of an arrayed block member this capture names, -1 for "the
+            // member as a whole". SPIR-V cannot decorate a single array element, so a
+            // backend needs the element index to tell a full run from a partial one.
+            Int blockMemberElement = -1;
         };
 
         // ---- P1: everything a link PRODUCES, in one movable block ----
@@ -620,7 +874,6 @@ namespace MobileGL::MG_State::GLState {
         // without going through the gate.
         struct LinkArtifacts {
             SharedPtr<glslang::TProgram> program;
-            Vector<Vector<unsigned>> generatedSpirv;
 
             // Attributes (Vertex in)
             Vector<String> attribs;
@@ -642,7 +895,24 @@ namespace MobileGL::MG_State::GLState {
             // layout(location = N) default-block uniform qualifiers (the relaxed parse drops
             // them from reflection; the DoReflection assigner restores them from here).
             UnorderedMap<String, Int> linkedExplicitUniformLocations;
+            // Per-link snapshot of the default-block uniform INITIALIZERS the attached shaders
+            // declared ("uniform int i = 1;"). Desktop GLSL says that value is what the uniform
+            // reads until the application overwrites it, and relinking restores it - but the
+            // relaxed parse turns those uniforms into members of MGL_GLOBAL_UBO, where SPIR-V
+            // cannot carry an initializer, so the value only survives as this side-channel.
+            // Applied into the uniform shadow at the phase-B publish (ApplyUniformInitialValues).
+            Vector<glslang::TIntermediate::TUniformInitializer> uniformInitialValues;
             UnorderedMap<String, Uint> uniformLocations;
+            // ---- "written since link" (see MarkUniformWrittenAtLocation) ----
+            // In LinkArtifacts deliberately: a link is exactly the event that retracts every
+            // write (GL resets uniforms to their initial values), so living here means the set
+            // is cleared by the same three paths that clear the rest of a link's output -
+            // Link()'s whole-struct reset, ResetLinkArtifacts, and the publish's move - and no
+            // fourth reset site can be forgotten. Empty (and never allocated) for a program
+            // that never asked to be separable.
+            Vector<Uint64> writtenUniformLocationBits;
+            Vector<Uint64> writtenUniformIndexBits;
+            Vector<Uint> writtenUniformIndices;
             // Ordered by location,
             // aka. uniformIndexInTProgram[loc] == "uniform index of TProgram at location `loc`"
             Vector<Int> uniformIndexInTProgram;
@@ -664,11 +934,6 @@ namespace MobileGL::MG_State::GLState {
             // SetShaderStorageBlockBinding for why this one is by name and not by index.
             UnorderedMap<String, Int> shaderStorageBlockBinding;
 
-            // Need to be reflected after linking of SPIR-V binary
-            Vector<Uint> uniformOffsets;
-            Vector<Uint> uniformSizesInBytes;
-            Vector<Uint8> globalUboScratch;
-
             Uint activeUniformCount = 0;
             Uint maxUniformLocation = 0;
             Int uniformNameMaxLength = 0;
@@ -697,6 +962,35 @@ namespace MobileGL::MG_State::GLState {
             Uint32 xfbPackedStride = 0;
         };
 
+        // ---- everything phase B of a link produces, in one movable block ----
+        //
+        // The membership rule is the same mechanical one LinkArtifacts uses: this is exactly
+        // what ProgramSpirvTask writes, which is what makes moving it THE publish. It is
+        // deliberately NOT part of LinkArtifacts, and that separation is what routes the five
+        // readers of SPIR-V-derived data through their own join gate by compiler rather than
+        // by review - m_spirv is private and Spirv() is the only spelling that reaches it.
+        //
+        // Why these three and nothing else: `generatedSpirv` has no GL-thread reader at all
+        // (every consumer is a backend draw/prepare path), and `uniformOffsets` +
+        // `globalUboScratch` are the ONLY things glUniform*/glGetUniform* need that are
+        // derived from the OPTIMIZED SPIR-V rather than from glslang reflection - spirv-opt
+        // runs in place and can delete a uniform, or the whole global UBO, so the offsets
+        // cannot be lifted out of glslang's reflection instead.
+        struct SpirvArtifacts {
+            Vector<Vector<unsigned>> generatedSpirv;
+            // Byte offset of each uniform location inside globalUboScratch, or
+            // kInvalidUniformOffset. Sized maxUniformLocation + 1 by the routing pass.
+            Vector<Uint> uniformOffsets;
+            Vector<Uint8> globalUboScratch;
+            // False for a program whose SPIR-V was never produced (phase B cancelled at
+            // teardown or by a relink) or whose optimizer run failed. GL has no way to
+            // retract a LINK_STATUS it already reported true, so such a program stays
+            // "linked" and every reflection answer it has given stays correct - it is simply
+            // not drawable, which the backends already express through their link-status
+            // gates.
+            Bool spirvStatus = false;
+        };
+
         // ---- artifacts-only helpers, shared with ProgramLinkTask ----
         // Static and taking the block explicitly, because from stage 4 the link BODY needs
         // them while its artifacts still live on the job node, not on any ProgramObject. The
@@ -736,9 +1030,20 @@ namespace MobileGL::MG_State::GLState {
         // Blocks until a pending link has published its artifacts. Public because a few call
         // sites have to join without reading anything - see the explicit-join list (J1-J8) in
         // the P1 design. GL thread only.
+        //
+        // PHASE A ONLY. After this returns, LINK_STATUS and the whole GL query surface are
+        // final and truthful, but the SPIR-V and the uniform shadow may still be in flight.
         void JoinLink() const { EnsureLinkJoined(); }
 
-        // Drops a link that is still in flight, without waiting for it. Called at the points
+        // Both phases. The draw path uses this, and must: the backends sample lifetimeId /
+        // backendStateVersion / the UBO content version OUTSIDE the gate, so a draw that
+        // joined only phase A would sample a version, join phase B later inside the same draw
+        // (through GetGeneratedSpirv), and memoize under a version the phase-B publish had
+        // already superseded - the exact lost-invalidation hazard J1 exists to prevent.
+        void JoinLinkAndSpirv() const { EnsureSpirvJoined(); }
+
+        // Drops BOTH phases of a link that is still in flight, without waiting for either.
+        // Called at the points
         // where the pending link's result stops being the answer to "what did this program
         // link to": a re-link supersedes it, glProgramBinary must force LINK_STATUS false,
         // and a destroyed program has no observers left.
@@ -757,7 +1062,15 @@ namespace MobileGL::MG_State::GLState {
         // MUST NOT JOIN - this is what GL_COMPLETION_STATUS_KHR reads when the extension
         // surface lands. "No job at all" counts as complete: there is nothing outstanding to
         // wait for.
-        Bool IsLinkComplete() const { return m_pendingLink == nullptr || IsPendingLinkTerminal(); }
+        //
+        // BOTH phases, deliberately: an application that polls GL_COMPLETION_STATUS_KHR and
+        // then draws must not be told "done" while the SPIR-V is still being generated, or
+        // the draw it was cleared for is the thing that blocks.
+        Bool IsLinkComplete() const { return IsPhaseALinkComplete() && IsSpirvComplete(); }
+        // Phase A alone, for the callers that only care about the query surface (and for the
+        // tests that pin the two phases apart).
+        Bool IsPhaseALinkComplete() const { return m_pendingLink == nullptr || IsPendingLinkTerminal(); }
+        Bool IsSpirvComplete() const { return m_pendingSpirv == nullptr || IsPendingSpirvTerminal(); }
 
         void SetTransformFeedbackVaryings(Vector<String>&& names, GLenum bufferMode) {
             m_requestedXfbVaryings = Move(names);
@@ -824,6 +1137,44 @@ namespace MobileGL::MG_State::GLState {
         // node's state goes through this out-of-line helper.
         Bool IsPendingLinkTerminal() const;
 
+        // ---- the second join gate: phase-B (SPIR-V) output only ----
+        // Phase A FIRST, always. Two reasons: the phase-B publish replays the uniform writes
+        // that were buffered during its window, and those need the phase-A reflection to
+        // validate against; and a caller that reaches a phase-B getter without having settled
+        // phase A would otherwise leave the link half-published.
+        //
+        // Same inline/out-of-line split as the phase-A gate, for the same reason: the five
+        // getters behind this one include the per-draw uniform upload path.
+        void EnsureSpirvJoined() const {
+            if (m_pendingLink) JoinPendingLink();
+            if (m_pendingSpirv) JoinPendingSpirv();
+        }
+        void JoinPendingSpirv() const;
+        Bool IsPendingSpirvTerminal() const;
+
+        // One buffered non-opaque glUniform* write. `dataOffset` indexes m_pendingUniformBytes,
+        // which is one append-only blob rather than a per-record allocation.
+        struct PendingUniformWrite {
+            Uint location = 0;
+            Uint byteOffsetInUniform = 0;
+            Uint byteSize = 0;
+            Uint dataOffset = 0;
+        };
+        // Replays the buffer into the freshly published shadow, in write order, and drains it.
+        // Each record re-does the bounds check and the bytes-equal dedupe the live write path
+        // performs, so "an identical write does not move the content version" survives the
+        // detour exactly - and a record that really does change bytes moves the version, which
+        // is what makes a backend re-upload the UBO it cached during the window.
+        void ReplayBufferedUniformWrites() const;
+        // Seeds the freshly published uniform shadow with the declared initializers. Runs at
+        // the phase-B publish, BEFORE ReplayBufferedUniformWrites, so an application write
+        // made during the A->B window still wins - which is the GL ordering.
+        void ApplyUniformInitialValues() const;
+        // Past this, BufferUniformWrite declines and the write joins instead. Sized so an
+        // ordinary pack load never reaches it (a pending window is one program's worth of
+        // uniforms) while a pathological writer cannot grow the heap without bound.
+        static constexpr SizeT kMaxBufferedUniformBytes = 4u << 20;
+
         LinkArtifacts& Artifacts() {
             EnsureLinkJoined();
             return m_artifacts;
@@ -832,6 +1183,14 @@ namespace MobileGL::MG_State::GLState {
             EnsureLinkJoined();
             return m_artifacts;
         }
+        SpirvArtifacts& Spirv() {
+            EnsureSpirvJoined();
+            return m_spirv;
+        }
+        const SpirvArtifacts& Spirv() const {
+            EnsureSpirvJoined();
+            return m_spirv;
+        }
 
         // GL-thread-only companion to ResetLinkArtifacts (see its definition). Const because
         // the publish half of the join calls it; see the mutable counters below.
@@ -869,11 +1228,22 @@ namespace MobileGL::MG_State::GLState {
         Bool m_deleteStatus = false;
         Bool m_binaryRetrievableHint = false;
         Bool m_separable = false;
+        // Monotone "this program may ever be a pipeline stage" latch; see SetSeparable for why
+        // it is a latch and not just m_separable. Outside LinkArtifacts on purpose: a relink
+        // clears the write SET, but a program that was separable is still separable after it.
+        Bool m_tracksUniformWrites = false;
+        // Generation counters that must NOT be reset by a link, for the same reason the memo
+        // versions above are not: a reader compares them for INEQUALITY, so a reset could make
+        // a stale cache compare equal to a fresh program. See their getters.
+        Uint32 m_uniformWriteSetVersion = 0;
+        Uint32 m_imageUnitVersion = 0;
         Bool m_validateStatus = true;
         // Mutable, like m_artifacts and for the same reason: publishing a pending link is a
         // READ-side operation (the first gated getter is what pulls the result in), and the
         // publish has to bump these. Still GL-thread-only - a worker never touches them.
         mutable Uint32 m_backendStateVersion = 0;
+        // Interface-block binding generation; see GetBlockBindingVersion.
+        Uint32 m_blockBindingVersion = 0;
 
         // Backend-owned content-hash memo (see GetBackendHashMemo): valid only while
         // m_backendStateVersion matches. Several slots, not one: a backend may resolve the same
@@ -899,10 +1269,22 @@ namespace MobileGL::MG_State::GLState {
         // Mutable because publishing is a READ-side operation: a const getter has to be able
         // to settle an outstanding link before answering it.
         mutable LinkArtifacts m_artifacts;
+        // Phase-B output. Same mutability argument as m_artifacts, reached only through
+        // Spirv().
+        mutable SpirvArtifacts m_spirv;
 
         // The link job, from enqueue until the first observable read pulls its result. Null
         // means m_artifacts is already the answer - which is the state every reader outside
         // the pending window sees, and the whole reason the gate above is one branch.
         mutable SharedPtr<ProgramLinkTask> m_pendingLink;
+        // The SPIR-V job, chained behind m_pendingLink. Null means m_spirv is already the
+        // answer. A program can be in the window where m_pendingLink is already null (phase A
+        // published, the query surface is live) while this is still set.
+        mutable SharedPtr<ProgramSpirvTask> m_pendingSpirv;
+        // glUniform* writes taken while m_pendingSpirv was set, in call order, plus their
+        // bytes. Drained by the phase-B publish and cleared by every cancel site (a relink's
+        // uniforms are not the previous link's uniforms).
+        mutable Vector<PendingUniformWrite> m_pendingUniformWrites;
+        mutable Vector<Uint8> m_pendingUniformBytes;
     };
 } // namespace MobileGL::MG_State::GLState

MobileGL/MG_State/GLState/ProgramState/ProgramPipelineObject.h

@@ -40,25 +40,112 @@ namespace MobileGL {
 
                 Uint GetExternalIndex() const { return m_externalIndex; }
 
+                // glIsProgramPipeline's answer, and NOT the same question as "does this object
+                // exist" (GL 4.6 core 7.4: a GenProgramPipelines name "acquires program pipeline
+                // state only when first bound"). The object is materialized by any of the
+                // commands that take state from a reserved name - including the pure queries
+                // glGetProgramPipelineiv and glGetProgramPipelineInfoLog, which have to answer
+                // out of default state without ever making the name report as an object. So
+                // existence is map membership and this is a separate latch, exactly as
+                // TransformFeedbackObject::everBound is.
+                Bool GetEverBound() const { return m_everBound; }
+                void MarkEverBound() { m_everBound = true; }
+
+                // The stages a DRAW is built from: every stage but compute. GL 4.6 core 7.4
+                // makes the compute stage exclusive - a program object containing a compute
+                // shader may contain no other stage, and a pipeline's compute stage is
+                // dispatched on its own and never participates in a draw. So the compute stage
+                // is not merely irrelevant to the composite below, it must never enter it: a
+                // compute module handed to vkCreateGraphicsPipelines is a driver crash rather
+                // than an error return (Adreno 830 SIGSEGVs inside it).
+                static constexpr SizeT kGraphicsStageCount = static_cast<SizeT>(ShaderStage::Compute);
+                static_assert(static_cast<SizeT>(ShaderStage::Compute) + 1 ==
+                                  static_cast<SizeT>(ShaderStage::ShaderStageCount),
+                              "ShaderStage must keep Compute last so the graphics stages are a prefix");
+
                 // A draw sees one program, but a pipeline holds one program per stage. The
-                // stages are composited into a single hidden program object, rebuilt whenever
-                // the stage set - or any stage program's own link - changes. The signature is
-                // what that "changes" means: a stage program's lifetime id pins the object and
-                // its backend state version pins the link generation.
-                using DrawProgramSignature =
-                    Array<Uint64, static_cast<SizeT>(ShaderStage::ShaderStageCount) * 2>;
+                // GRAPHICS stages are composited into a single hidden program object, rebuilt
+                // whenever the stage set - or any stage program's own link - changes. The
+                // signature is what that "changes" means: a stage program's lifetime id pins the
+                // object and its LINK version pins the link generation. It covers exactly the
+                // stages the composite is built from, so attaching or relinking a compute stage
+                // never invalidates a perfectly good graphics composite - and the compute stage,
+                // having no composite of its own, can never collide with it.
+                //
+                // GetLinkVersion() and NOT GetBackendStateVersion(), which is what this used to
+                // key on. The backend state version moves on every glUniform1i to a sampler and
+                // every glUniformBlockBinding, so the "set a sampler unit, draw" loop that the
+                // SSO conformance cases run threw the composite away and REBUILT it on every
+                // single draw: a fresh ProgramObject, a full Link(true) settled synchronously
+                // (glslang + SPIR-V + spirv-opt), a full re-mirror, and a brand-new program
+                // identity that invalidated both backends' per-program registries and pipeline
+                // memos along the way. The composite's CONTENT depends on the link generations
+                // and nothing else, and m_linkVersion is bumped by exactly those
+                // (BumpLinkObservableVersions).
+                //
+                // The prerequisite that makes the narrowing legal: because the composite no
+                // longer rebuilds when per-program uniform STATE changes, every such change must
+                // reach it through the refresh below instead. Both do - sampler/image units via
+                // MirrorUniformValues, interface block bindings via MirrorBlockBindings - and
+                // the two setters that write them still bump the counters the REFRESH gate reads
+                // (see ComputeUniformMirrorVersions), which is a separate question from what
+                // this signature reads. They are the only two writers of m_backendStateVersion
+                // outside the link paths, so nothing else was ever riding on the rebuild.
+                using DrawProgramSignature = Array<Uint64, kGraphicsStageCount * 2>;
 
                 DrawProgramSignature ComputeDrawProgramSignature() const {
                     DrawProgramSignature signature{};
-                    for (SizeT stage = 0; stage < static_cast<SizeT>(ShaderStage::ShaderStageCount); ++stage) {
+                    for (SizeT stage = 0; stage < kGraphicsStageCount; ++stage) {
                         const auto& program = m_stagePrograms[stage];
                         if (!program) continue;
                         signature[stage * 2] = program->GetLifetimeId();
-                        signature[stage * 2 + 1] = program->GetBackendStateVersion();
+                        signature[stage * 2 + 1] = program->GetLinkVersion();
                     }
                     return signature;
                 }
 
+                // Per-program state is written to the STAGE programs - glUniform* addresses the
+                // pipeline's active program (GL 4.6 core 7.6.1), glProgramUniform* addresses a
+                // named one, and the two block-binding calls address a named one - while the
+                // draw reads the composite. Two different objects' state, so the composite is
+                // refreshed from its stage programs before each draw that needs it. These are
+                // the per-stage versions "needs it" is measured against. All zero after a
+                // rebuild, because a fresh composite holds only what its shaders declared and
+                // so needs a full refresh.
+                //
+                // backendStateVersion belongs HERE even though ComputeDrawProgramSignature no
+                // longer reads it, and that is the whole point of the split: a sampler-unit or
+                // uniform-block-binding write must still trip the MIRROR (it is now the only
+                // route those values have to the composite) while deliberately NOT tripping the
+                // rebuild. uboContentVersion covers ordinary uniform writes, and
+                // blockBindingVersion covers the storage-block setter, which moves neither of
+                // the other two.
+                using UniformMirrorVersions = Array<Uint64, kGraphicsStageCount * 2>;
+
+                UniformMirrorVersions ComputeUniformMirrorVersions() const {
+                    UniformMirrorVersions versions{};
+                    for (SizeT stage = 0; stage < kGraphicsStageCount; ++stage) {
+                        const auto& program = m_stagePrograms[stage];
+                        if (!program) continue;
+                        versions[stage * 2] = (static_cast<Uint64>(program->GetBackendStateVersion()) << 32) |
+                                              static_cast<Uint64>(program->GetUBOContentVersion());
+                        // Their own slot rather than folded into the pair above: the
+                        // storage-block setter moves the block-binding version and NOTHING
+                        // else, so a rebinding would otherwise be invisible to the refresh
+                        // gate - and the write-set version is the only counter that moves for
+                        // a write which ENLARGES the set without changing a byte (see
+                        // ProgramObject::GetUniformWriteSetVersion), which is what decides
+                        // which stage owns a shared name.
+                        versions[stage * 2 + 1] = (static_cast<Uint64>(program->GetBlockBindingVersion()) << 32) |
+                                                  static_cast<Uint64>(program->GetUniformWriteSetVersion());
+                    }
+                    return versions;
+                }
+                const UniformMirrorVersions& GetMirroredUniformVersions() const { return m_mirroredUniformVersions; }
+                void SetMirroredUniformVersions(const UniformMirrorVersions& versions) {
+                    m_mirroredUniformVersions = versions;
+                }
+
                 const SharedPtr<ProgramObject>& GetCachedDrawProgram(const DrawProgramSignature& signature) const {
                     static const SharedPtr<ProgramObject> nullProgram = nullptr;
                     if (!m_drawProgram || m_drawProgramSignature != signature) return nullProgram;
@@ -67,6 +154,8 @@ namespace MobileGL {
                 void SetCachedDrawProgram(const DrawProgramSignature& signature, SharedPtr<ProgramObject> program) {
                     m_drawProgramSignature = signature;
                     m_drawProgram = Move(program);
+                    // A rebuilt composite holds none of its stage programs' uniform values yet.
+                    m_mirroredUniformVersions = {};
                 }
 
             private:
@@ -74,9 +163,11 @@ namespace MobileGL {
                 SharedPtr<ProgramObject> m_activeProgram;
                 SharedPtr<ProgramObject> m_drawProgram;
                 DrawProgramSignature m_drawProgramSignature{};
+                UniformMirrorVersions m_mirroredUniformVersions{};
                 String m_infoLog;
                 const Uint m_externalIndex = 0;
                 Bool m_validateStatus = false;
+                Bool m_everBound = false;
             };
         } // namespace GLState
     } // namespace MG_State

MobileGL/MG_State/GLState/ProgramState/ProgramSpirvTask.cpp

@@ -0,0 +1,318 @@
+// MobileGL - MobileGL/MG_State/GLState/ProgramState/ProgramSpirvTask.cpp
+// Copyright (c) 2025-2026 MobileGL-Dev
+// Licensed under the GNU Lesser General Public License v3.0:
+//   https://www.gnu.org/licenses/gpl-3.0.txt
+//   https://www.gnu.org/licenses/lgpl-3.0.txt
+// SPDX-License-Identifier: LGPL-3.0-only
+// End of Source File Header
+
+#include "ProgramSpirvTask.h"
+
+#include <MG_State/GLState/ProgramState/ShaderCompileTask.h> // GlslangThreadAllocatorGuard
+#include <MG_Util/Async/ShaderCompilePool.h>
+#include <MG_Util/ShaderTranspiler/ShaderCompiler.h>
+#include <MG_Util/ShaderTranspiler/SpvcSession.h>
+#include <MG_Util/ShaderTranspiler/Types.h>
+
+#include <cstring>
+
+namespace MobileGL::MG_State::GLState {
+    void ProgramSpirvTask::DeferLog(String line) { diagnostics.logLines.push_back(Move(line)); }
+
+    void ProgramSpirvTask::SubmitAfter(const SharedPtr<ProgramLinkTask>& phaseA) {
+        MOBILEGL_ASSERT(phaseA != nullptr, "ProgramSpirvTask::SubmitAfter: the phase-A node is missing");
+        m_phaseA = phaseA;
+
+        auto self = std::static_pointer_cast<ProgramSpirvTask>(shared_from_this());
+        // ONE dependency, so no counter and no guard slot: the whole race
+        // ProgramLinkTask::SubmitAfter's +1 exists to close (a dependency settling while the
+        // remaining edges are still being registered) cannot arise with a single edge.
+        //
+        // Runs inline, right here, if phase A is already terminal.
+        phaseA->OnTerminal([self, phaseA] {
+            // "Dependency did not complete, publish nothing" - the same collapse
+            // ProgramLinkTask::CompiledArtifacts() performs for an abandoned compile. Note
+            // this reads the HANDOFF, never phaseA->artifacts: the GL thread may already be
+            // moving those out (see the class comment).
+            if (!phaseA->IsComplete() || !phaseA->spirvHandoff.ready) {
+                self->Cancel();
+                return;
+            }
+            // A cancel that landed before phase A settled (relink, glDeleteProgram, teardown).
+            // Posting would only make a worker pick up a node that immediately falls out of
+            // Run() again.
+            if (self->IsCancellationRequested()) {
+                self->Cancel();
+                return;
+            }
+            // Non-throwing by construction, and it has to be: this is a JobNode continuation,
+            // so on the pool side it runs inside an Asio handler. Post() contains its own
+            // allocation failures, and the catch below CANCELS rather than swallowing - a
+            // phase B that is never posted is a GL thread blocked forever in
+            // EnsureSpirvJoined(), which is far worse than a program reported as not drawable.
+            try {
+                MG_Util::Async::ShaderCompilePool::Get().Post(self);
+            } catch (...) {
+                self->Cancel();
+            }
+        });
+    }
+
+    void ProgramSpirvTask::RunInlineAfter(const SharedPtr<ProgramLinkTask>& phaseA) {
+        MOBILEGL_ASSERT(phaseA != nullptr, "ProgramSpirvTask::RunInlineAfter: the phase-A node is missing");
+        MOBILEGL_ASSERT(phaseA->IsTerminal(),
+                        "ProgramSpirvTask::RunInlineAfter: phase A has not settled; the inline path must run the "
+                        "two bodies in order on the same thread");
+        m_phaseA = phaseA;
+        RunInline();
+    }
+
+    // Pure CPU work only, on a pool worker (or on the GL thread in the inline mode).
+    // Everything this reads is either owned by this node or published by a terminal phase A;
+    // everything it writes is `artifacts` (and diagnostics). Same prohibitions as
+    // ProgramLinkTask::RunBody - no GL/EGL call, no pActiveBackendObject read, no
+    // pGLContext->RecordError().
+    void ProgramSpirvTask::RunBody() {
+        // glslang leaves this worker's TLS pool allocator pointing at the last arena it
+        // touched; reset it on the way out so an unrelated later job cannot allocate out of a
+        // pool that has since been freed. Declared FIRST so it is destroyed LAST - the phase-A
+        // release below drops the TShaders (and their pools) and must happen inside it.
+        const GlslangThreadAllocatorGuard glslangGuard;
+        using namespace MG_Util::ShaderTranspiler;
+
+        // Drop phase A - and with it the TShaders, the TProgram reference and phase A's whole
+        // input snapshot - the moment this body is done, rather than at some later join. For a
+        // pack load that is the difference between W glslang arenas alive and all of them.
+        struct PhaseAReleaser {
+            SharedPtr<ProgramLinkTask>& node;
+            ~PhaseAReleaser() { node.reset(); }
+        } const phaseAReleaser{m_phaseA};
+
+        if (!m_phaseA) return;
+        // Non-const: the TShaders are dropped below, the moment GlslangToSpv is finished with
+        // them. This is safe by ownership rather than by locking - phase A is terminal and
+        // therefore immutable to everyone else, the GL-thread join touches only `artifacts`
+        // and `diagnostics`, and this node is the sole reader of the handoff.
+        ProgramLinkTask::SpirvHandoff& handoff = m_phaseA->spirvHandoff;
+        const Uint externalIndex = m_phaseA->in.externalIndex;
+        if (!handoff.ready || !handoff.reflection.program) {
+            // Phase A did not reach its tail (it failed the link, or was cancelled mid-body).
+            // Publish nothing; spirvStatus stays false.
+            return;
+        }
+
+        MGLOG_D("ProgramObject %u: Starting SPIR-V generation", externalIndex);
+        GenerateSpirv(handoff, externalIndex);
+        // GlslangToSpv was the only consumer of the parsed ASTs; everything after this point
+        // works on the SPIR-V and on the TProgram's own self-contained reflection pool. Drop
+        // them here rather than at the end of the body, which is ~87% of this node's runtime
+        // earlier (spirv-opt plus routing).
+        //
+        // WHAT THIS ACTUALLY FREES, precisely - it is LESS than "the glslang arenas", and the
+        // difference matters for the peak-RSS story:
+        //   * CAS-LOSER shaders (the re-parse in ShaderCompileTask::ClaimParsedShader, i.e.
+        //     the 2nd..Nth link of a shared shader): freed here in full. The handoff is their
+        //     ONLY owner.
+        //   * CAS-WINNER shaders (the common case - one shader object linked into one
+        //     program, which is every program of an Iris pack load): NOT freed here. The
+        //     winner branch returns a COPY of ShaderCompileTask::artifacts.shader
+        //     (ShaderCompileTask.cpp:320) and the node never releases its own reference, while
+        //     phase A holds that node through in.shaders[i].compiled for its whole life - and
+        //     phase A lives until PhaseAReleaser fires at the end of this body. So the
+        //     refcount goes 2 -> 1 here and the arena dies where it would have died anyway.
+        //
+        // Making it free the winner's arena too means releasing whatever pins the TShader
+        // inside the compile node, and neither obvious route is safe as a drive-by: moving out
+        // of artifacts.shader at claim time races ShaderObject::GetCompiledShader() on the GL
+        // thread and breaks JobNode's "a terminal node is immutable" invariant, and dropping
+        // phase A's in.shaders[i].compiled reference only helps when nothing else holds the
+        // node (the adoption map is a WeakPtr index, so it would also change which nodes stay
+        // adoptable). Both belong in a change that can be reviewed against the consume-once
+        // and adoption semantics on their own terms.
+        handoff.shaders.clear();
+
+        MGLOG_D("ProgramObject %u: Building global-UBO routing tables", externalIndex);
+        BuildGlobalUboRouting(handoff, externalIndex);
+        MGLOG_D("ProgramObject %u: Binary generation finished (generatedSpirv size=%zu)", externalIndex,
+                artifacts.generatedSpirv.size());
+    }
+
+    void ProgramSpirvTask::GenerateSpirv(const ProgramLinkTask::SpirvHandoff& handoff, const Uint externalIndex) {
+        /* As we passed first stage compilation/linking,
+         * we'll assume all the operations here should
+         * pass. We may be able to employ some optimizations
+         * here without the burden of error reporting.
+         */
+        using namespace MG_Util::ShaderTranspiler;
+        MGLOG_D("ProgramObject %u: GenerateSpirv - start", externalIndex);
+
+        // The shaders were parsed once, in the link-compatible (relaxed Vulkan-rules)
+        // configuration, and the handoff's program linked those parses - so it IS the program
+        // the backends consume. Generate SPIR-V straight from its intermediates, which the
+        // handoff's TShaders keep alive.
+        ProgramBinaryAttrib binaryAttrib{
+            .shaderTypes = handoff.shaderTypes,
+            .program = *handoff.reflection.program,
+        };
+        MGLOG_D("ProgramObject %u: GenerateSpirv - requesting SPIR-V binary from program", externalIndex);
+        auto binaryResult = ShaderCompiler::GetSpirvBinaryFromProgram(binaryAttrib);
+        if (!binaryResult) {
+            DeferLog(std::format("ProgramObject {}: GenerateSpirv - GetSpirvBinaryFromProgram failed", externalIndex));
+            MOBILEGL_ASSERT(binaryResult, "GetSpirvBinaryFromProgram failed");
+            return; // spirvStatus stays false: linked, but not drawable.
+        }
+        artifacts.generatedSpirv = Move(binaryResult.value());
+        MGLOG_D("ProgramObject %u: GenerateSpirv - generated %zu SPIR-V modules", externalIndex,
+                artifacts.generatedSpirv.size());
+
+        // Linked SPIR-V generated, sanitize and optimize it
+        Bool allOptimized = true;
+        {
+            for (auto& spv : artifacts.generatedSpirv) {
+                auto success = ShaderCompiler::SanitizeAndOptimizeBinary(spv, spv);
+                if (!success) {
+                    // The one genuine phase-B failure mode: one of the seven optimizer passes
+                    // reported failure, so `spv` is whatever the run left behind. A fordebug
+                    // build trips the assert below; a release build used to hand that binary
+                    // to the backend regardless. It no longer does - the program keeps its
+                    // (truthful) LINK_STATUS and its whole query surface, and the routing
+                    // tables below still give every settable uniform storage so glUniform*
+                    // and glGetUniform* keep working, but spirvStatus stays false and the
+                    // backends refuse to build or draw with it.
+                    allOptimized = false;
+                    DeferLog(std::format("ProgramObject {}: SanitizeAndOptimizeBinary failed; the program is linked "
+                                         "and queryable but not drawable",
+                                         externalIndex));
+                }
+                MOBILEGL_ASSERT(success, "SanitizeBinary failed");
+            }
+        }
+        artifacts.spirvStatus = allOptimized;
+    }
+
+    void ProgramSpirvTask::BuildGlobalUboRouting(const ProgramLinkTask::SpirvHandoff& handoff,
+                                                 const Uint externalIndex) {
+        using namespace MG_Util::ShaderTranspiler;
+        // The phase-A reflection slice this pass keys off. Carried in the handoff rather than
+        // read off the phase-A node's artifacts, which the join has very likely already moved.
+        const ProgramObject::LinkArtifacts& reflection = handoff.reflection;
+
+        artifacts.uniformOffsets.clear();
+        artifacts.globalUboScratch.clear();
+        // kInvalidUniformOffset marks locations that end up without global-UBO backing
+        // (e.g. the optimizer eliminated every use of the uniform); the fallback pass
+        // below gives those locations tail storage so glUniform* always has a target.
+        artifacts.uniformOffsets.resize(reflection.maxUniformLocation + 1, ProgramObject::kInvalidUniformOffset);
+        for (SizeT i = 0; i < artifacts.generatedSpirv.size(); i++) {
+            auto& spv = artifacts.generatedSpirv[i];
+
+            auto shaderType = i < handoff.shaderTypes.size() ? handoff.shaderTypes[i] : GLenum{0};
+            MGLOG_D("ProgramObject %u: BuildGlobalUboRouting - parsing SPIR-V meta data for module %zu "
+                    "(shaderType=%u, wordCount=%zu)",
+                    externalIndex, i, shaderType, spv.size());
+            SpvcSession session(spv, SessionUsageBit::Reflection);
+            auto result = session.ParseMetaData();
+            if (result < 0) {
+                MGLOG_D("ProgramObject %u: BuildGlobalUboRouting - SpvcSession::ParseMetaData failed for module %zu, "
+                        "err = %d%s",
+                        externalIndex, i, result,
+                        (result == SPVC_ERROR_INVALID_SPIRV ? ". Probably no global UBO?" : ""));
+                continue;
+            } else {
+                auto& meta = session.GetMetadata();
+                auto size = meta.globalUboSize;
+                MGLOG_D("ProgramObject %u: BuildGlobalUboRouting - SPIR-V meta: uboSize=%zu plainUniformCount=%zu "
+                        "plainUniformOffsets=%zu",
+                        externalIndex, meta.globalUboSize, meta.plainUniformMemberSizesInBytes.size(),
+                        meta.plainUniformOffsetsInUBO.size());
+                if (size == 0) {
+                    continue;
+                }
+                if (artifacts.globalUboScratch.size() < size) {
+                    artifacts.globalUboScratch.resize(size);
+                }
+                for (const auto& [name, offset] : meta.plainUniformOffsetsInUBO) {
+                    // SPIRV-Reflect leaf names never carry a "[0]" suffix; frontend
+                    // reflection keys arrays as "arr[0]" (GL naming), so retry with the
+                    // suffix before declaring the uniform unbacked.
+                    auto locationIt = reflection.uniformLocations.find(name);
+                    if (locationIt == reflection.uniformLocations.end()) {
+                        locationIt = reflection.uniformLocations.find(name + "[0]");
+                    }
+                    if (locationIt == reflection.uniformLocations.end()) {
+                        MGLOG_D("ProgramObject %u: BuildGlobalUboRouting - uniform '%s' offset=%u but not found in "
+                                "uniformLocations",
+                                externalIndex, name.c_str(), offset);
+                        continue;
+                    }
+                    const Uint baseLocation = locationIt->second;
+                    if (!ProgramObject::IsValidUniformLocation(reflection, static_cast<Int>(baseLocation))) {
+                        continue;
+                    }
+
+                    const Int uniformIndex = reflection.uniformIndexInTProgram[baseLocation];
+                    const GLint arraySize = ProgramObject::GetUniformArraySizeByTIndex(reflection, uniformIndex);
+                    Uint arrayStride = 0;
+                    const auto strideIt = meta.plainUniformArrayStridesInUBO.find(name);
+                    if (strideIt != meta.plainUniformArrayStridesInUBO.end()) {
+                        arrayStride = strideIt->second;
+                    }
+
+                    // Array uniforms span one location per element (see DoReflection);
+                    // give each element its real byte offset inside the UBO.
+                    const GLint elementCount = (arraySize > 1 && arrayStride == 0) ? 1 : std::max(arraySize, 1);
+                    for (GLint element = 0; element < elementCount; ++element) {
+                        const Uint location = baseLocation + static_cast<Uint>(element);
+                        if (location > reflection.maxUniformLocation ||
+                            reflection.uniformIndexInTProgram[location] != uniformIndex) {
+                            break;
+                        }
+                        artifacts.uniformOffsets[location] = offset + static_cast<Uint>(element) * arrayStride;
+                    }
+                    MGLOG_D("ProgramObject %u: BuildGlobalUboRouting - uniform '%s' offset=%u stride=%u assigned "
+                            "to locations %u..%u",
+                            externalIndex, name.c_str(), offset, arrayStride, baseLocation,
+                            baseLocation + static_cast<Uint>(elementCount) - 1);
+                }
+                MGLOG_D("ProgramObject %u: BuildGlobalUboRouting - finished parsing module %zu metadata",
+                        externalIndex, i);
+            }
+        }
+
+        // Fallback pass: a linked program's active non-opaque uniforms must accept
+        // glUniform*/glGetUniform* even when the optimized SPIR-V no longer contains
+        // them (AggressiveDCE can remove a dead loop together with the only loads of a
+        // uniform -- or the entire global UBO, leaving the scratch unallocated). Hand
+        // such locations CPU-side storage at the (16-byte aligned) tail of the shadow
+        // buffer; backends bind at least the SPIR-V-declared UBO range, and the GPU
+        // never reads these bytes, so this only keeps the GL-visible state coherent.
+        for (Uint location = 0; location <= reflection.maxUniformLocation; ++location) {
+            if (artifacts.uniformOffsets[location] != ProgramObject::kInvalidUniformOffset) continue;
+            if (!ProgramObject::IsValidUniformLocation(reflection, static_cast<Int>(location))) continue;
+            const auto& uniform = reflection.program->getUniform(reflection.uniformIndexInTProgram[location]);
+            const glslang::TType* type = uniform.getType();
+            if (type != nullptr && type->isOpaque()) continue;
+            if (uniform.index >= 0 && uniform.index < reflection.program->getNumUniformBlocks() &&
+                std::strstr(reflection.program->getUniformBlock(uniform.index).name.c_str(),
+                            MG_Util::ShaderTranspiler::GLOBAL_UBO_NAME) == nullptr) {
+                // Member of a named uniform block: not settable through glUniform*, so it
+                // needs no global-UBO shadow storage.
+                continue;
+            }
+
+            // std140-style slot: the matrix upload paths write column vectors at
+            // 16-byte strides, so a matrix slot must cover cols * 16 bytes.
+            SizeT slotSize = MG_Util::GetGLTypeSize(uniform.glDefineType);
+            if (type != nullptr && type->isMatrix()) {
+                slotSize = static_cast<SizeT>(type->getMatrixCols()) * 16u;
+            }
+            slotSize = (slotSize + 15u) & ~static_cast<SizeT>(15u);
+            const SizeT slotOffset = (artifacts.globalUboScratch.size() + 15u) & ~static_cast<SizeT>(15u);
+            artifacts.globalUboScratch.resize(slotOffset + slotSize, 0);
+            artifacts.uniformOffsets[location] = static_cast<Uint>(slotOffset);
+            MGLOG_D("ProgramObject %u: BuildGlobalUboRouting - uniform '%s' location %u has no UBO backing in the "
+                    "generated SPIR-V (optimized out?); allocated %zu fallback bytes at scratch offset %zu",
+                    externalIndex, uniform.name.c_str(), location, slotSize, slotOffset);
+        }
+    }
+} // namespace MobileGL::MG_State::GLState

MobileGL/MG_State/GLState/ProgramState/ProgramSpirvTask.h

@@ -0,0 +1,77 @@
+// MobileGL - MobileGL/MG_State/GLState/ProgramState/ProgramSpirvTask.h
+// Copyright (c) 2025-2026 MobileGL-Dev
+// Licensed under the GNU Lesser General Public License v3.0:
+//   https://www.gnu.org/licenses/gpl-3.0.txt
+//   https://www.gnu.org/licenses/lgpl-3.0.txt
+// SPDX-License-Identifier: LGPL-3.0-only
+// End of Source File Header
+
+#pragma once
+#include <Includes.h>
+#include <MG_State/GLState/ProgramState/ProgramLinkTask.h>
+#include <MG_Util/Async/JobNode.h>
+
+namespace MobileGL::MG_State::GLState {
+    // PHASE B of one glLinkProgram: GlslangToSpv, spirv-opt, and the SPIRV-Cross pass that
+    // builds the glUniform*-to-scratch routing tables. Chained behind exactly one
+    // ProgramLinkTask and joined by exactly five ProgramObject getters (GetGeneratedSpirv,
+    // GetUniformOffset, MapUBO, GetUBOData, GetUBOSize), so ~120 other getters and the whole
+    // GL query surface stay on the phase-A gate and answer without waiting for any of this.
+    //
+    // ---- what this node may read, and what it may not ----
+    // It holds the phase-A node by SharedPtr and reads `phaseA->spirvHandoff` plus
+    // `phaseA->in`. It must NEVER read `phaseA->artifacts` or `phaseA->diagnostics`: the GL
+    // thread MOVES the artifacts out of the node at the phase-A join and DRAINS the
+    // diagnostics there, and both of those can happen while this body runs. The handoff exists
+    // precisely so this node has a copy of everything it needs that the join does not touch.
+    // (The general JobNode rule - a terminal node is immutable, so its outputs need no further
+    // synchronization - covers everything except the two members the join consumes.)
+    //
+    // ---- lifetime ----
+    // The handoff owns the Vector<SharedPtr<glslang::TShader>>, and that is mandatory rather
+    // than tidy: glslang::TProgram stores raw TShader* and, for the one-shader-per-stage case,
+    // BORROWS each stage's TIntermediate from its TShader. GlslangToSpv reads exactly those
+    // intermediates. Before the split the shaders died when ProgramLinkTask::RunBody returned,
+    // which was safe only because nothing called getIntermediate() afterwards.
+    //
+    // ---- failure ----
+    // A cancel (relink, teardown, program destruction) or an optimizer failure publishes
+    // spirvStatus = false rather than a half-built program. GL cannot retract a LINK_STATUS it
+    // already reported true, so such a program stays linked and fully queryable; it is just
+    // not drawable, which the backends express through their existing link-status gates.
+    class ProgramSpirvTask final : public MG_Util::Async::JobNode {
+    public:
+        // ---- output: valid iff IsComplete(), immutable afterwards ----
+        // Moved (never copied) into the ProgramObject by EnsureSpirvJoined().
+        ProgramObject::SpirvArtifacts artifacts;
+
+        // Posts this job when `phaseA` goes terminal - and not one moment earlier, so the body
+        // never waits on anything (invariant I4: no job body may block on another job). A
+        // single dependency needs no counter, just the one continuation; it runs inline right
+        // here if `phaseA` is already terminal, which is the same case
+        // ProgramLinkTask::SubmitAfter already reasons about.
+        //
+        // GL thread only, and only after the caller has stored a SharedPtr to this node: the
+        // continuation takes shared_from_this().
+        void SubmitAfter(const SharedPtr<ProgramLinkTask>& phaseA);
+
+        // The async-off / glMaxShaderCompilerThreadsKHR(0) path: run the body on the calling
+        // thread, right now, against an ALREADY-TERMINAL phase A. Deliberately not routed
+        // through SubmitAfter, whose continuation would Post() to a pool that is merely
+        // unused rather than stopped - that would move the work off-thread in the one mode
+        // whose contract is "byte-identical to the synchronous implementation".
+        void RunInlineAfter(const SharedPtr<ProgramLinkTask>& phaseA);
+
+    private:
+        void RunBody() override;
+
+        void GenerateSpirv(const ProgramLinkTask::SpirvHandoff& handoff, Uint externalIndex);
+        void BuildGlobalUboRouting(const ProgramLinkTask::SpirvHandoff& handoff, Uint externalIndex);
+
+        // Worker-side MGLOG replacement, replayed by the join on the GL thread. Same reason as
+        // ProgramLinkTask::DeferLog.
+        void DeferLog(String line);
+
+        SharedPtr<ProgramLinkTask> m_phaseA;
+    };
+} // namespace MobileGL::MG_State::GLState

MobileGL/MG_State/GLState/ProgramState/ProgramState.cpp

@@ -111,9 +111,13 @@ namespace MobileGL::MG_State::GLState {
         // that can grow, and a reallocation underneath this loop would be a use-after-free
         // that only shows up on the one GL call that walks the whole table. The copy costs a
         // refcount bump on a path a mode switch takes at most once.
+        // BOTH phases per program. This is the glMaxShaderCompilerThreadsKHR(0) path, whose
+        // contract is that nothing is outstanding when it returns - a program left with its
+        // SPIR-V job in flight would make the very next GL_COMPLETION_STATUS_KHR read GL_FALSE
+        // in a mode the extension says cannot have anything pending.
         for (SizeT i = 0; i < m_programObjects.size(); ++i) {
             const SharedPtr<ProgramObject> program = m_programObjects[i];
-            if (program) program->JoinLink();
+            if (program) program->JoinLinkAndSpirv();
         }
         for (SizeT i = 0; i < m_shaderObjects.size(); ++i) {
             const SharedPtr<ShaderObject> shader = m_shaderObjects[i];
@@ -122,7 +126,7 @@ namespace MobileGL::MG_State::GLState {
         // The currently-used program is reachable through m_programObjects unless
         // glDeleteProgram already freed its slot while it stayed current. Nothing else holds
         // a GL-visible name for it, but a draw would still join it, so settle it here too.
-        if (m_currentProgram) m_currentProgram->JoinLink();
+        if (m_currentProgram) m_currentProgram->JoinLinkAndSpirv();
     }
 
     void ProgramState::MarkShaderObjectForDeletion(Uint shader) {

MobileGL/MG_State/GLState/ProgramState/ShaderCompileAdoptionMap.cpp

@@ -70,9 +70,10 @@ namespace MobileGL::MG_State::GLState {
     void ShaderCompileAdoptionMap::SweepIfCrowded() {
         if (m_entries.size() < m_sweepThreshold) return;
 
-        // Collect first, erase after: FastSTL::unordered_map is open-addressed, so erasing
-        // through an iterator that the same loop is still advancing is not worth reasoning
-        // about on a path this cold.
+        // Collect first, erase after: the map is open-addressed and erases by shifting the
+        // rest of the probe cluster into the hole, so an erase moves entries other than the
+        // erased one. Copying the keys out sidesteps that entirely, and this path is cold
+        // enough that the extra vector is not worth reasoning about the alternative.
         Vector<ShaderSourceKey> dead;
         for (const auto& entry : m_entries) {
             const SharedPtr<ShaderCompileTask> node = entry.second.lock();

MobileGL/MG_State/GLState/ProgramState/ShaderObject.cpp

@@ -88,12 +88,19 @@ namespace MobileGL::MG_State::GLState {
         // another object, THIS object has not pulled its result yet. (An adopted node may
         // already be terminal - the join then only replays what is left of its diagnostics.)
         m_compileJoined = false;
+        // A new compile is a new story: whatever the optimistic getters promised about the
+        // previous node does not carry over.
+        m_optimisticAnswerLatched = false;
     }
 
     void ShaderObject::DropCompileNode() const {
         if (!m_compiled) return;
         m_compiled->ReleaseAdopter();
         m_compiled.reset();
+        // No node means IsCompileComplete() is trivially true and the truthful answers are
+        // "not compiled"; a stale latch would keep reporting a compile that no longer
+        // exists as GL_TRUE.
+        m_optimisticAnswerLatched = false;
     }
 
     void ShaderObject::InvalidateCompiledState() {

MobileGL/MG_State/GLState/ProgramState/ShaderObject.h

@@ -116,8 +116,10 @@ namespace MobileGL {
             Bool GetDeleteStatus() const { return m_deleteStatus; }
 
             // Blocks until a pending compile has published its artifacts. Public for the
-            // sites that must join without reading anything - ProgramObject::Link's
-            // prologue, which needs every attached shader settled before it runs.
+            // sites that must join without reading anything - ProgramState::
+            // JoinAllPendingWork, the glMaxShaderCompilerThreadsKHR(0) path that settles
+            // every outstanding job. glLinkProgram deliberately does NOT come through
+            // here: its prologue takes the nodes unjoined via CompiledNodeForLink().
             void JoinCompile() const { EnsureCompileJoined(); }
 
             // True while this object holds the outcome (success OR failure) of a Compile()
@@ -141,6 +143,23 @@ namespace MobileGL {
             // outstanding to wait for.
             Bool IsCompileComplete() const { return m_compiled == nullptr || m_compiled->IsTerminal(); }
 
+            // MOBILEGL_ASYNC_OPTIMISTIC_SHADER_STATUS's one-story-per-compile memory. The
+            // three optimistic getter sites in GL_Program ask THIS instead of a raw
+            // IsCompileComplete() peek, and the difference is the latch: without it, a job
+            // that settles between two adjacent queries hands the application a torn pair -
+            // an empty info log from the optimistic read, then the real GL_FALSE from the
+            // truthful one - and an application that aborts on that status never reaches
+            // the link join that quotes the real log. So the first optimistic answer
+            // latches: until the next AdoptCompileNode/DropCompileNode this object keeps
+            // answering optimistically even after the job settles, and a real failure
+            // surfaces exactly once, at the link. Returns whether the caller should answer
+            // optimistically; the caller has already checked the quirk is active.
+            Bool TakeOptimisticCompileAnswer() const {
+                if (!m_optimisticAnswerLatched && IsCompileComplete()) return false;
+                m_optimisticAnswerLatched = true;
+                return true;
+            }
+
         private:
             // ---- The one and only join gate for compile output (P1 invariant I5) ----
             // The fast path - no job, or a job whose result this object has already pulled -
@@ -231,6 +250,10 @@ namespace MobileGL {
             // Exactly-once latch for the pull above. Armed with every new job node, set by
             // the one join that consumes it.
             mutable Bool m_compileJoined = false;
+            // TakeOptimisticCompileAnswer's memory: this object has answered a compile
+            // query optimistically for the current node. Cleared wherever the node
+            // changes hands (AdoptCompileNode) or goes away (DropCompileNode).
+            mutable Bool m_optimisticAnswerLatched = false;
         };
     } // namespace MG_State::GLState
 } // namespace MobileGL

MobileGL/MG_State/GLState/RenderState/RenderState.cpp

@@ -7,6 +7,7 @@
 // End of Source File Header
 
 #include "RenderState.h"
+#include "MG_Util/Debug/Log.h"
 #include "MG_Util/Types.h"
 
 namespace MobileGL {
@@ -268,9 +269,14 @@ namespace MobileGL {
             }
 
             void RenderState::SetCapabilityIndexed(CapabilityInput cap, Uint index, Bool enabled) {
-                // Only for BlendState currently
+                // Only for BlendState currently. The GL entry points (glEnablei/glDisablei) already
+                // reject every non-GL_BLEND target with GL_INVALID_ENUM before reaching here, so this
+                // is a backstop - but it must stay a backstop: THROW_UNIMPL_EXCEPTION unwinds a C++
+                // exception through the C GL ABI and terminates the process.
                 if (cap != CapabilityInput::Blend) {
-                    THROW_UNIMPL_EXCEPTION;
+                    MGLOG_I("RenderState::SetCapabilityIndexed: indexed capability state exists only for "
+                            "GL_BLEND (cap=%d, index=%u); ignoring",
+                            static_cast<int>(cap), index);
                     return;
                 }
                 if (index >= MG_State::GLState::FramebufferObject::MAX_DRAW_BUFFERS) {
@@ -284,9 +290,13 @@ namespace MobileGL {
             }
 
             Bool RenderState::IsCapabilityEnabledIndexed(CapabilityInput cap, Uint index) const {
-                // Only for BlendState currently
+                // Only for BlendState currently - same backstop reasoning as SetCapabilityIndexed:
+                // glIsEnabledi has already answered GL_INVALID_ENUM/GL_FALSE for anything else, and a
+                // query must never be able to terminate the process.
                 if (cap != CapabilityInput::Blend) {
-                    THROW_UNIMPL_EXCEPTION;
+                    MGLOG_I("RenderState::IsCapabilityEnabledIndexed: indexed capability state exists only "
+                            "for GL_BLEND (cap=%d, index=%u); reporting disabled",
+                            static_cast<int>(cap), index);
                     return false;
                 }
                 if (index >= MG_State::GLState::FramebufferObject::MAX_DRAW_BUFFERS) {

MobileGL/MG_State/GLState/VertexArrayState/VertexArrayObject.cpp

@@ -29,6 +29,8 @@ namespace MobileGL::MG_State::GLState {
             attr.Normalized = false;
             attr.Stride = 0;
             attr.Offset = 0;
+            attr.LegacyStride = 0;
+            attr.LegacyPointer = 0;
             attr.Buffer = nullptr;
 
             BumpAttributeFormatVersion(index);
@@ -61,10 +63,19 @@ namespace MobileGL::MG_State::GLState {
     void VertexArrayObject::SetAttributeFormat(Uint index, int size, DataType type, Bool normalized, int stride,
                                                SizeT offset, Bool isInteger, Bool isBgra) {
         if (index >= MAX_VERTEX_ATTRIBS) return;
+        if (size < 1 || size > 4) {
+            return;
+        }
 
         // The classic pointer-style API takes back full ownership of the resolved fields.
         m_attributeUsesBindingModel[index] = false;
 
+        // The legacy query shadows: written here and nowhere else, so a later binding-model
+        // mutation cannot leak into VERTEX_ATTRIB_ARRAY_STRIDE / _POINTER. They are pure
+        // query state, so they carry no version bump of their own.
+        m_attributes[index].LegacyStride = stride;
+        m_attributes[index].LegacyPointer = offset;
+
         if (m_attributes[index].Size == size && m_attributes[index].Type == type &&
             m_attributes[index].Normalized == normalized && m_attributes[index].Stride == stride &&
             m_attributes[index].Offset == offset && m_attributes[index].IsInteger == isInteger &&
@@ -72,10 +83,6 @@ namespace MobileGL::MG_State::GLState {
             return;
         }
 
-        if (size < 1 || size > 4) {
-            return;
-        }
-
         auto& attr = m_attributes[index];
         attr.Size = size;
         attr.Type = type;
@@ -111,6 +118,12 @@ namespace MobileGL::MG_State::GLState {
         binding.Offset = offset;
         binding.Stride = effectiveStride;
         binding.Divisor = m_attributes[index].Divisor;
+
+        // Other attributes may already be pointed at this binding point through
+        // glVertexAttribBinding; they see the new buffer/offset/stride too (basic-state3
+        // checks exactly that after a glVertexAttribPointer). They are not adopted into the
+        // binding model here - only the ones already in it re-resolve.
+        ResolveAttributesForBinding(index, /*adopt: */ false);
     }
 
     void VertexArrayObject::BindAttributeBuffer(Uint index, const SharedPtr<BufferObject>& buffer) {
@@ -147,10 +160,24 @@ namespace MobileGL::MG_State::GLState {
 
     void VertexArrayObject::SetAttributeDivisor(Uint index, Uint divisor) {
         if (index >= MAX_VERTEX_ATTRIBS) return;
-        // glVertexAttribDivisor is VertexBindingDivisor on the attribute's own binding point
-        // (GL 4.6 core 10.3.2), so the binding-point view has to follow the resolved attribute.
-        if (index < MAX_VERTEX_ATTRIB_BINDINGS && m_attributeBindingIndex[index] == index) {
+        // GL 4.6 core 10.3.2 defines VertexAttribDivisor(i, d) as
+        //   VertexAttribBinding(i, i); VertexBindingDivisor(i, d)
+        // - the binding is RE-POINTED at i, it is not merely written through when it already
+        // happens to be i. Guarding the write on "binding == index" (which is what this did)
+        // left an attribute that glVertexAttribBinding had moved elsewhere pointing at the old
+        // binding, so the next resolve restored that binding's divisor and the new one was
+        // lost (KHR-GL4x.vertex_attrib_binding.basic-state4).
+        //
+        // What is deliberately NOT copied from VertexAttribBinding is the adoption into the
+        // binding model: an attribute configured the classic way keeps its pointer-resolved
+        // stride/offset, exactly as before. The binding point mirrors that state already
+        // (MirrorPointerIntoBinding), so nothing observable differs - and adopting it here
+        // would silently swap the raw pointer stride for the effective one under every
+        // application that calls glVertexAttribDivisor after glVertexAttribPointer.
+        if (index < MAX_VERTEX_ATTRIB_BINDINGS) {
+            m_attributeBindingIndex[index] = index;
             m_bindingPoints[index].Divisor = divisor;
+            ResolveAttributesForBinding(index, /*adopt: */ false);
         }
         if (m_attributes[index].Divisor == divisor) return;
         m_attributes[index].Divisor = divisor;
@@ -164,7 +191,6 @@ namespace MobileGL::MG_State::GLState {
 
     void VertexArrayObject::ResolveAttributeFromBinding(Uint attribIndex) {
         if (attribIndex >= MAX_VERTEX_ATTRIBS) return;
-        if (!m_attributeUsesBindingModel[attribIndex]) return;
 
         const Uint bindingIndex = m_attributeBindingIndex[attribIndex];
         if (bindingIndex >= MAX_VERTEX_ATTRIB_BINDINGS) return;
@@ -172,11 +198,24 @@ namespace MobileGL::MG_State::GLState {
 
         auto& attr = m_attributes[attribIndex];
 
+        // VERTEX_ATTRIB_ARRAY_DIVISOR is not independent per-attribute state: it IS the divisor
+        // of the binding point the attribute is attached to (GL 4.6 core 10.3.2), whichever API
+        // configured the attribute. glVertexBindingDivisor therefore has to reach a classic
+        // pointer-configured attribute as well - basic-state4 alternates the two spellings on
+        // the same attribute and expects each to win in turn.
+        if (attr.Divisor != binding.Divisor) {
+            attr.Divisor = binding.Divisor;
+            BumpAttributeFormatVersion(attribIndex);
+        }
+
+        // Everything else stays owned by whichever API configured the attribute: a classic
+        // glVertexAttrib*Pointer attribute keeps its pointer-resolved stride and offset.
+        if (!m_attributeUsesBindingModel[attribIndex]) return;
+
         const SizeT resolvedOffset = binding.Offset + m_attributeRelativeOffset[attribIndex];
-        if (attr.Stride != binding.Stride || attr.Offset != resolvedOffset || attr.Divisor != binding.Divisor) {
+        if (attr.Stride != binding.Stride || attr.Offset != resolvedOffset) {
             attr.Stride = binding.Stride;
             attr.Offset = resolvedOffset;
-            attr.Divisor = binding.Divisor;
             BumpAttributeFormatVersion(attribIndex);
         }
 
@@ -186,6 +225,14 @@ namespace MobileGL::MG_State::GLState {
         }
     }
 
+    void VertexArrayObject::ResolveAttributesForBinding(Uint bindingIndex, Bool adopt) {
+        for (Uint attribIndex = 0; attribIndex < MAX_VERTEX_ATTRIBS; ++attribIndex) {
+            if (m_attributeBindingIndex[attribIndex] != bindingIndex) continue;
+            if (adopt) m_attributeUsesBindingModel[attribIndex] = true;
+            ResolveAttributeFromBinding(attribIndex);
+        }
+    }
+
     void VertexArrayObject::SetBindingBuffer(Uint bindingIndex, const SharedPtr<BufferObject>& buffer, SizeT offset,
                                              int stride) {
         if (bindingIndex >= MAX_VERTEX_ATTRIB_BINDINGS) return;
@@ -195,15 +242,10 @@ namespace MobileGL::MG_State::GLState {
         binding.Offset = offset;
         binding.Stride = stride;
 
-        for (Uint attribIndex = 0; attribIndex < MAX_VERTEX_ATTRIBS; ++attribIndex) {
-            if (m_attributeBindingIndex[attribIndex] == bindingIndex) {
-                // Binding a vertex buffer to a binding point adopts every attribute currently
-                // mapped to that binding point into the binding model (the default mapping is
-                // attribute i -> binding i, which matches the GL 4.3 rules for state mixing).
-                m_attributeUsesBindingModel[attribIndex] = true;
-                ResolveAttributeFromBinding(attribIndex);
-            }
-        }
+        // Binding a vertex buffer to a binding point adopts every attribute currently mapped to
+        // that binding point into the binding model (the default mapping is attribute i ->
+        // binding i, which matches the GL 4.3 rules for state mixing).
+        ResolveAttributesForBinding(bindingIndex, /*adopt: */ true);
     }
 
     void VertexArrayObject::SetBindingDivisor(Uint bindingIndex, Uint divisor) {
@@ -211,11 +253,7 @@ namespace MobileGL::MG_State::GLState {
 
         m_bindingPoints[bindingIndex].Divisor = divisor;
 
-        for (Uint attribIndex = 0; attribIndex < MAX_VERTEX_ATTRIBS; ++attribIndex) {
-            if (m_attributeBindingIndex[attribIndex] == bindingIndex && m_attributeUsesBindingModel[attribIndex]) {
-                ResolveAttributeFromBinding(attribIndex);
-            }
-        }
+        ResolveAttributesForBinding(bindingIndex, /*adopt: */ false);
     }
 
     void VertexArrayObject::SetAttributeBinding(Uint attribIndex, Uint bindingIndex) {

MobileGL/MG_State/GLState/VertexArrayState/VertexArrayObject.h

@@ -32,6 +32,16 @@ namespace MobileGL {
                 Bool IsBgra = false;
                 Uint Divisor = 0;
                 SharedPtr<BufferObject> Buffer;
+
+                // GL 4.6 core table 23.3: VERTEX_ATTRIB_ARRAY_STRIDE and _POINTER are the
+                // arguments of the last glVertexAttrib*Pointer call on this attribute,
+                // reported verbatim, and NOTHING else writes them - not glVertexAttribFormat,
+                // not glBindVertexBuffer. Stride/Offset above are the *resolved* draw inputs
+                // and the binding model does overwrite those, so the two views have to be
+                // stored apart or the binding-model sequence reports a legacy state it never
+                // set (KHR-GL4x.vertex_attrib_binding.basic-state3).
+                int LegacyStride = 0;
+                SizeT LegacyPointer = 0;
             };
 
             // ARB_vertex_attrib_binding separate binding point. Attributes configured through the
@@ -40,7 +50,8 @@ namespace MobileGL {
             struct VertexBufferBindingPoint {
                 SharedPtr<BufferObject> Buffer;
                 SizeT Offset = 0;
-                int Stride = 0;
+                // GL 4.6 core table 23.4: the initial VERTEX_BINDING_STRIDE is 16, not 0.
+                int Stride = 16;
                 Uint Divisor = 0;
             };
 
@@ -185,6 +196,10 @@ namespace MobileGL {
                 void BumpAttributeBufferVersion(Uint index);
                 void BumpAttributeSwitchVersion(Uint index);
                 void ResolveAttributeFromBinding(Uint attribIndex);
+                // Re-resolve every attribute currently pointed at `bindingIndex`. `adopt` turns
+                // the ones that are not in the binding model yet into binding-model attributes
+                // first (what glBindVertexBuffer does, GL 4.3 rules for state mixing).
+                void ResolveAttributesForBinding(Uint bindingIndex, Bool adopt);
 
                 // The default mapping is attribute i -> binding point i. Keep it an iota over
                 // MAX_VERTEX_ATTRIBS rather than a literal list: a literal list silently leaves the

MobileGL/MG_Test/Backend/DirectGLES/CMakeLists.txt

@@ -0,0 +1,20 @@
+cmake_minimum_required(VERSION 3.14)
+
+add_executable(
+        EsslShaderPassTest
+        EsslShaderPassTest.cpp
+)
+
+target_include_directories(EsslShaderPassTest PRIVATE
+        ${MGL_ROOT}/include
+        ${MGL_ROOT}/MobileGL
+)
+
+target_link_libraries(
+        EsslShaderPassTest PRIVATE
+        GTest::gtest_main
+        ${LINK_LIBRARIES}
+)
+
+include(GoogleTest)
+gtest_discover_tests(EsslShaderPassTest DISCOVERY_TIMEOUT 30 PROPERTIES LABELS unit)

MobileGL/MG_Test/Backend/DirectGLES/EsslShaderPassTest.cpp

@@ -0,0 +1,369 @@
+// MobileGL - MobileGL/MG_Test/Backend/DirectGLES/EsslShaderPassTest.cpp
+// Copyright (c) 2025-2026 MobileGL-Dev
+// Licensed under the GNU Lesser General Public License v3.0:
+//   https://www.gnu.org/licenses/gpl-3.0.txt
+//   https://www.gnu.org/licenses/lgpl-3.0.txt
+// SPDX-License-Identifier: LGPL-3.0-only
+// End of Source File Header
+//
+// The post-transpile textual passes the DirectGLES ("Espryt") backend runs over the ESSL
+// SPIRV-Cross hands it (MG_Backend/DirectGLES/Utils.cpp). No GL context and no driver: the
+// passes are pure String -> String, so the shapes they have to survive can be pinned here
+// instead of only on a device.
+
+#include <gtest/gtest.h>
+
+#include <MG_Backend/DirectGLES/Utils.h>
+
+using namespace MobileGL;
+using MobileGL::MG_Backend::DirectGLES::PrgramImpl::IMAGE_WRITE_ALIAS_PREFIX;
+using MobileGL::MG_Backend::DirectGLES::PrgramImpl::RemoveLayoutBinding;
+using MobileGL::MG_Backend::DirectGLES::PrgramImpl::SplitReadWriteImageUniforms;
+
+namespace {
+    Bool Contains(const String& haystack, const String& needle) {
+        return haystack.find(needle) != String::npos;
+    }
+
+    SizeT CountOf(const String& haystack, const String& needle) {
+        SizeT count = 0;
+        for (SizeT pos = haystack.find(needle); pos != String::npos; pos = haystack.find(needle, pos + 1)) {
+            ++count;
+        }
+        return count;
+    }
+
+    String WriteAlias(const String& name) { return String(IMAGE_WRITE_ALIAS_PREFIX) + name; }
+} // namespace
+
+// The bug the pass exists for. SPIRV-Cross speculatively marks every storage image
+// NonWritable+NonReadable, then clears NonReadable at the OpImageRead and NonWritable at the
+// OpImageWrite, so an image the shader both reads and writes comes out carrying NEITHER
+// `readonly` nor `writeonly` - which ESSL rejects for any format other than r32f/r32i/r32ui
+// (GLSL ES 3.20 4.10). The device compile then fails and the draw silently binds program 0.
+TEST(SplitReadWriteImageUniformsTest, ReadWriteImageIsSplitIntoAnAliasingPair) {
+    const String source = R"(#version 320 es
+layout(binding = 2, rgba8) uniform highp image2D goku;
+layout(location = 0) out highp vec4 mg_FragColor;
+void main()
+{
+    highp vec4 loaded = imageLoad(goku, ivec2(gl_FragCoord.xy));
+    imageStore(goku, ivec2(gl_FragCoord.xy), loaded + vec4(0.25));
+    mg_FragColor = loaded;
+}
+)";
+    const String out = SplitReadWriteImageUniforms(source);
+
+    // Both halves: same binding, same format, same type - which is what makes two image
+    // variables on one image unit legal.
+    EXPECT_TRUE(Contains(out, "layout(binding = 2, rgba8) uniform readonly highp image2D goku;"));
+    EXPECT_TRUE(Contains(out, "layout(binding = 2, rgba8) uniform writeonly highp image2D " + WriteAlias("goku") + ";"));
+
+    // The load keeps the original name, the store moves to the writeonly half.
+    EXPECT_TRUE(Contains(out, "imageLoad(goku,"));
+    EXPECT_TRUE(Contains(out, "imageStore(" + WriteAlias("goku") + ","));
+    EXPECT_FALSE(Contains(out, "imageStore(goku,"));
+}
+
+// The split has to survive RemoveLayoutBinding, which runs straight after it: an ES image
+// unit cannot be assigned through the API, so the layout qualifier is the only binding
+// mechanism and both halves must still carry theirs afterwards.
+TEST(SplitReadWriteImageUniformsTest, BothHalvesKeepTheirBindingThroughRemoveLayoutBinding) {
+    const String source = R"(#version 320 es
+layout(binding = 5, rgba8) uniform highp image2D goku;
+void main()
+{
+    imageStore(goku, ivec2(0), imageLoad(goku, ivec2(0)));
+}
+)";
+    const String out = RemoveLayoutBinding(SplitReadWriteImageUniforms(source));
+    EXPECT_EQ(CountOf(out, "binding = 5"), 2u);
+}
+
+// Cheap hardening: the pass does not depend on SPIRV-Cross getting the read-only case right,
+// and a shader that only reads must not pay for a second uniform.
+TEST(SplitReadWriteImageUniformsTest, ReadOnlyImageGetsReadonlyAndIsNotSplit) {
+    const String source = R"(#version 320 es
+layout(binding = 1, rgba16f) uniform highp image2DArray trunks;
+layout(location = 0) out highp vec4 mg_FragColor;
+void main()
+{
+    mg_FragColor = imageLoad(trunks, ivec3(0));
+}
+)";
+    const String out = SplitReadWriteImageUniforms(source);
+    EXPECT_TRUE(Contains(out, "layout(binding = 1, rgba16f) uniform readonly highp image2DArray trunks;"));
+    EXPECT_FALSE(Contains(out, "writeonly"));
+    EXPECT_FALSE(Contains(out, IMAGE_WRITE_ALIAS_PREFIX));
+    EXPECT_EQ(CountOf(out, "image2DArray"), 1u);
+}
+
+TEST(SplitReadWriteImageUniformsTest, WriteOnlyImageGetsWriteonlyAndIsNotSplit) {
+    const String source = R"(#version 320 es
+layout(binding = 3, rgba8) uniform highp image2D gohan;
+void main()
+{
+    imageStore(gohan, ivec2(0), vec4(1.0));
+}
+)";
+    const String out = SplitReadWriteImageUniforms(source);
+    EXPECT_TRUE(Contains(out, "layout(binding = 3, rgba8) uniform writeonly highp image2D gohan;"));
+    EXPECT_FALSE(Contains(out, "readonly"));
+    EXPECT_FALSE(Contains(out, IMAGE_WRITE_ALIAS_PREFIX));
+}
+
+// r32f / r32i / r32ui are exactly the formats GLSL ES 3.20 4.10 exempts from the rule, so a
+// read+write image in one of them is already legal and must not be doubled.
+TEST(SplitReadWriteImageUniformsTest, ExemptFormatsAreLeftCompletelyAlone) {
+    for (const char* format : {"r32f", "r32i", "r32ui"}) {
+        const String type = String(format) == "r32f" ? "image2D" : (String(format) == "r32i" ? "iimage2D" : "uimage2D");
+        const String source = "#version 320 es\nlayout(binding = 4, " + String(format) + ") uniform highp " + type +
+                              " vegeta;\nvoid main()\n{\n    imageStore(vegeta, ivec2(0), imageLoad(vegeta, "
+                              "ivec2(0)));\n}\n";
+        EXPECT_EQ(SplitReadWriteImageUniforms(source), source) << "format " << format;
+    }
+}
+
+// A declaration SPIRV-Cross already qualified is none of this pass's business.
+TEST(SplitReadWriteImageUniformsTest, AlreadyQualifiedDeclarationsAreUntouched) {
+    const String source = R"(#version 320 es
+layout(binding = 0, rgba8) uniform readonly highp image2D reader;
+layout(binding = 1, rgba8) uniform writeonly highp image2D writer;
+void main()
+{
+    imageStore(writer, ivec2(0), imageLoad(reader, ivec2(0)));
+}
+)";
+    EXPECT_EQ(SplitReadWriteImageUniforms(source), source);
+}
+
+// The binding of an image array is the array's base; splitting must keep the array on both
+// halves (dropping the subscript would silently turn 3 units into 1).
+TEST(SplitReadWriteImageUniformsTest, ImageArraySplitsAndKeepsItsArraySize) {
+    const String source = R"(#version 320 es
+layout(binding = 6, rgba8) uniform highp image2D gohan[3];
+void main()
+{
+    imageStore(gohan[1], ivec2(0), imageLoad(gohan[2], ivec2(0)));
+}
+)";
+    const String out = SplitReadWriteImageUniforms(source);
+    EXPECT_TRUE(Contains(out, "layout(binding = 6, rgba8) uniform readonly highp image2D gohan[3];"));
+    EXPECT_TRUE(Contains(out,
+                         "layout(binding = 6, rgba8) uniform writeonly highp image2D " + WriteAlias("gohan") + "[3];"));
+    EXPECT_TRUE(Contains(out, "imageStore(" + WriteAlias("gohan") + "[1],"));
+    EXPECT_TRUE(Contains(out, "imageLoad(gohan[2],"));
+}
+
+// The rewrite is by identifier, not by substring: "goku" must not reach into "goku_hd", and
+// the two images have to be classified independently.
+TEST(SplitReadWriteImageUniformsTest, ANameThatIsAPrefixOfAnotherIsNotClobbered) {
+    const String source = R"(#version 320 es
+layout(binding = 1, rgba8) uniform highp image2D goku;
+layout(binding = 2, rgba8) uniform highp image2D goku_hd;
+void main()
+{
+    highp vec4 loaded = imageLoad(goku, ivec2(0));
+    imageStore(goku, ivec2(0), loaded);
+    imageStore(goku_hd, ivec2(0), loaded);
+}
+)";
+    const String out = SplitReadWriteImageUniforms(source);
+
+    // goku is read+write -> split; goku_hd is write-only -> qualified in place, not split.
+    EXPECT_TRUE(Contains(out, "layout(binding = 1, rgba8) uniform readonly highp image2D goku;"));
+    EXPECT_TRUE(Contains(out, "layout(binding = 1, rgba8) uniform writeonly highp image2D " + WriteAlias("goku") + ";"));
+    EXPECT_TRUE(Contains(out, "layout(binding = 2, rgba8) uniform writeonly highp image2D goku_hd;"));
+    EXPECT_TRUE(Contains(out, "imageStore(goku_hd,"));
+    EXPECT_FALSE(Contains(out, WriteAlias("goku") + "_hd"));
+    EXPECT_FALSE(Contains(out, WriteAlias("goku_hd")));
+}
+
+// Other qualifiers belong to both halves, and the memory qualifier goes where SPIRV-Cross
+// puts it (right after `uniform`) so the image-rebinding regex in Managers.cpp still matches.
+TEST(SplitReadWriteImageUniformsTest, ExistingQualifiersAreCarriedOntoBothHalves) {
+    const String source = R"(#version 320 es
+layout(binding = 2, rgba8) uniform coherent restrict highp image2D goku;
+void main()
+{
+    imageStore(goku, ivec2(0), imageLoad(goku, ivec2(0)));
+}
+)";
+    const String out = SplitReadWriteImageUniforms(source);
+    EXPECT_TRUE(Contains(out, "uniform readonly coherent restrict highp image2D goku;"));
+    EXPECT_TRUE(
+        Contains(out, "uniform writeonly coherent restrict highp image2D " + WriteAlias("goku") + ";"));
+}
+
+// imageSize reads no texels and writes none, so it decides nothing; readonly is what keeps
+// such a declaration legal.
+TEST(SplitReadWriteImageUniformsTest, ImageSizeAloneDoesNotCountAsALoadOrAStore) {
+    const String source = R"(#version 320 es
+layout(binding = 8, rgba8ui) uniform highp uimage2D sizeOnly;
+layout(location = 0) out highp vec4 mg_FragColor;
+void main()
+{
+    mg_FragColor = vec4(float(imageSize(sizeOnly).x));
+}
+)";
+    const String out = SplitReadWriteImageUniforms(source);
+    EXPECT_TRUE(Contains(out, "layout(binding = 8, rgba8ui) uniform readonly highp uimage2D sizeOnly;"));
+    EXPECT_FALSE(Contains(out, IMAGE_WRITE_ALIAS_PREFIX));
+}
+
+// The alias must not land on an identifier the shader already uses.
+TEST(SplitReadWriteImageUniformsTest, AliasNameAvoidsAnExistingIdentifier) {
+    const String source = R"(#version 320 es
+layout(binding = 6, rgba8) uniform highp image2D taken;
+highp vec4 mg_imageWrite_taken;
+void main()
+{
+    imageStore(taken, ivec2(0), imageLoad(taken, ivec2(0)) + mg_imageWrite_taken);
+}
+)";
+    const String out = SplitReadWriteImageUniforms(source);
+    EXPECT_FALSE(Contains(out, "image2D " + WriteAlias("taken") + ";"));
+    EXPECT_TRUE(Contains(out, "image2D " + WriteAlias("taken") + "X;"));
+    EXPECT_TRUE(Contains(out, "imageStore(" + WriteAlias("taken") + "X,"));
+    EXPECT_TRUE(Contains(out, "+ mg_imageWrite_taken)"));
+}
+
+// A use the pass cannot account for (here: the image handed to a user function) means it
+// cannot know every store site, so it declines rather than emitting a half-rewritten shader.
+TEST(SplitReadWriteImageUniformsTest, AnUnrecognizedUseLeavesTheDeclarationAlone) {
+    const String source = R"(#version 320 es
+layout(binding = 2, rgba8) uniform highp image2D passed;
+highp vec4 helper(highp image2D img) { return imageLoad(img, ivec2(0)); }
+void main()
+{
+    imageStore(passed, ivec2(0), helper(passed));
+}
+)";
+    EXPECT_EQ(SplitReadWriteImageUniforms(source), source);
+}
+
+TEST(SplitReadWriteImageUniformsTest, ShaderWithoutImagesIsReturnedUnchanged) {
+    const String source = R"(#version 320 es
+layout(binding = 0) uniform highp sampler2D goku;
+layout(location = 0) out highp vec4 mg_FragColor;
+void main()
+{
+    mg_FragColor = texture(goku, vec2(0.5));
+}
+)";
+    EXPECT_EQ(SplitReadWriteImageUniforms(source), source);
+}
+
+// ---------------------------------------------------------------------------------------
+// RetargetTextureBufferExtension
+//
+// Buffer textures are core in the OpenGL 3.1+ context MobileGL advertises, but in ES they
+// only became core in 3.2; below that they need EXT_texture_buffer or OES_texture_buffer.
+// SPIRV-Cross hardcodes the EXT spelling for every Dim=Buffer image it emits below ESSL 320
+// and offers no way to ask for the other one, so on a driver that advertises only the OES
+// name the `: require` is a hard compile error over a single token.
+// ---------------------------------------------------------------------------------------
+
+using Tier = MobileGL::MG_External::GLESCapabilities::TextureBufferTier;
+using MobileGL::MG_Backend::DirectGLES::PrgramImpl::RetargetTextureBufferExtension;
+
+namespace {
+    // What SPIRV-Cross actually emits for `uniform isamplerBuffer CloudFaces;` at ESSL 310 -
+    // the shape that empties Minecraft 26.3's cloud layer on a driver without the extension.
+    const String kBufferTextureShader = R"(#version 310 es
+#extension GL_EXT_texture_buffer : require
+precision highp float;
+uniform highp isamplerBuffer CloudFaces;
+layout(location = 0) out highp vec4 mg_FragColor;
+void main()
+{
+    mg_FragColor = vec4(texelFetch(CloudFaces, gl_VertexID).r);
+}
+)";
+} // namespace
+
+TEST(RetargetTextureBufferExtensionTest, OesOnlyDriverGetsTheOesDirective) {
+    const String out = RetargetTextureBufferExtension(kBufferTextureShader, Tier::ExtensionOES);
+    EXPECT_TRUE(Contains(out, "#extension GL_OES_texture_buffer : require"))
+        << "the OES driver's own spelling must reach the directive:\n" << out;
+    EXPECT_FALSE(Contains(out, "GL_EXT_texture_buffer"))
+        << "the EXT spelling this driver does not advertise must be gone:\n" << out;
+    // Only the directive changes; the declaration and the fetch are identical between the two
+    // extensions and must not be touched.
+    EXPECT_TRUE(Contains(out, "uniform highp isamplerBuffer CloudFaces;"));
+    EXPECT_TRUE(Contains(out, "texelFetch(CloudFaces, gl_VertexID)"));
+}
+
+TEST(RetargetTextureBufferExtensionTest, ExtDriverKeepsWhatSpirvCrossEmitted) {
+    EXPECT_EQ(RetargetTextureBufferExtension(kBufferTextureShader, Tier::ExtensionEXT),
+              kBufferTextureShader);
+}
+
+// ES 3.2 needs no directive at all, and SPIRV-Cross emits none at ESSL 320 - but a shader
+// that arrived with one anyway must not be rewritten to a name the pass was not asked for.
+TEST(RetargetTextureBufferExtensionTest, CoreAndUnsupportedTiersAreNoOps) {
+    EXPECT_EQ(RetargetTextureBufferExtension(kBufferTextureShader, Tier::CoreEs32),
+              kBufferTextureShader);
+    EXPECT_EQ(RetargetTextureBufferExtension(kBufferTextureShader, Tier::None),
+              kBufferTextureShader);
+}
+
+// The name is only the subject of a rewrite where it is the subject of an #extension
+// directive. A shader that merely mentions it - in a comment SPIRV-Cross carried through, or
+// in an identifier - is not an extension request and must come out byte-identical.
+TEST(RetargetTextureBufferExtensionTest, OnlyExtensionDirectivesAreRewritten) {
+    const String source = R"(#version 310 es
+// GL_EXT_texture_buffer is what this shader would need
+precision highp float;
+uniform highp float GL_EXT_texture_buffer_lookalike;
+layout(location = 0) out highp vec4 mg_FragColor;
+void main()
+{
+    mg_FragColor = vec4(GL_EXT_texture_buffer_lookalike);
+}
+)";
+    EXPECT_EQ(RetargetTextureBufferExtension(source, Tier::ExtensionOES), source);
+}
+
+// The dangerous collision, and the one the directive check alone does NOT catch:
+// GL_EXT_texture_buffer is a strict prefix of GL_EXT_texture_buffer_object, a different and
+// real extension that SPIRV-Cross emits from the same Dim=Buffer branch on its legacy-desktop
+// path. Rewriting it would turn a valid request into one for a GL_OES_texture_buffer_object
+// that does not exist. Only an identifier-boundary check saves this, so it gets its own test
+// with the lookalike on a genuine #extension line.
+TEST(RetargetTextureBufferExtensionTest, ALongerExtensionSharingThePrefixIsNotRewritten) {
+    const String source = R"(#version 310 es
+#extension GL_EXT_texture_buffer_object : require
+precision highp float;
+void main() {}
+)";
+    EXPECT_EQ(RetargetTextureBufferExtension(source, Tier::ExtensionOES), source);
+
+    // And when both appear, exactly the exact-match one moves.
+    const String mixed = R"(#version 310 es
+#extension GL_EXT_texture_buffer_object : require
+#extension GL_EXT_texture_buffer : require
+precision highp float;
+void main() {}
+)";
+    const String out = RetargetTextureBufferExtension(mixed, Tier::ExtensionOES);
+    EXPECT_TRUE(Contains(out, "#extension GL_EXT_texture_buffer_object : require")) << out;
+    EXPECT_TRUE(Contains(out, "#extension GL_OES_texture_buffer : require")) << out;
+    EXPECT_EQ(CountOf(out, "GL_OES_texture_buffer_object"), 0u) << out;
+}
+
+// Whitespace between '#' and the keyword is legal in GLSL, and a shader carrying several
+// extension directives must have exactly the one retargeted.
+TEST(RetargetTextureBufferExtensionTest, SpacedDirectiveIsRewrittenAndNeighboursAreLeftAlone) {
+    const String source = R"(#version 310 es
+#  extension GL_EXT_texture_buffer : require
+#extension GL_EXT_shader_io_blocks : require
+precision highp float;
+void main() {}
+)";
+    const String out = RetargetTextureBufferExtension(source, Tier::ExtensionOES);
+    EXPECT_TRUE(Contains(out, "#  extension GL_OES_texture_buffer : require")) << out;
+    EXPECT_TRUE(Contains(out, "#extension GL_EXT_shader_io_blocks : require"))
+        << "an unrelated extension must survive untouched:\n" << out;
+    EXPECT_EQ(CountOf(out, "GL_OES_texture_buffer"), 1u);
+}

MobileGL/MG_Test/BackendLoader/BackendLoaderTest.cpp

@@ -52,6 +52,19 @@ namespace {
         GLfloat maxTextureMaxAnisotropy = 16.0f;
         bool maxTextureMaxAnisotropyQueried = false;
 
+        // Buffer textures. GL_MAX_TEXTURE_BUFFER_SIZE is only a legal pname once they exist, so
+        // asking on a driver without them raises GL_INVALID_ENUM - the same shape as the
+        // anisotropy probe above. The three entry-point knobs are separate because the
+        // unsuffixed name is the ES 3.2 CORE spelling while an EXT/OES driver exports the
+        // suffixed one: a resolver that only looks for the core name declares every extension
+        // driver unsupported, which is exactly the bug these knobs exist to pin.
+        GLint maxTextureBufferSize = 131072;
+        bool maxTextureBufferSizeQueried = false;
+        bool textureBufferSizeQueryRaisesError = false;
+        bool hasCoreTexBufferEntryPoint = true;
+        bool hasExtTexBufferEntryPoint = false;
+        bool hasOesTexBufferEntryPoint = false;
+
         GLuint nextBufferId = 1;
         GLuint nextShaderId = 1;
         GLuint nextProgramId = 1;
@@ -121,6 +134,14 @@ namespace {
                     *data = g_fake.fragmentInterpolationOffsetBits;
                 }
                 break;
+            case GL_MAX_TEXTURE_BUFFER_SIZE:
+                g_fake.maxTextureBufferSizeQueried = true;
+                if (g_fake.textureBufferSizeQueryRaisesError) {
+                    g_fake.pendingError = GL_INVALID_ENUM;
+                } else {
+                    *data = g_fake.maxTextureBufferSize;
+                }
+                break;
             // FillInGLESCapabilities reads the context version before running the
             // baseInstance probe, which requires ES >= 3.1.
             case GL_MAJOR_VERSION:
@@ -332,6 +353,21 @@ namespace {
         funcs.glDisable = [](GLenum) {};
         funcs.glMemoryBarrier = [](GLbitfield) {};
 
+        // Buffer-texture entry points, each present only when its knob says so. A real loader
+        // resolves the suffixed names only on a driver whose support is that extension.
+        funcs.glTexBuffer = g_fake.hasCoreTexBufferEntryPoint
+                                ? static_cast<MobileGL::MG_External::GLES::glTexBuffer_PTR>(
+                                      [](GLenum, GLenum, GLuint) {})
+                                : nullptr;
+        funcs.glTexBufferEXT = g_fake.hasExtTexBufferEntryPoint
+                                   ? static_cast<MobileGL::MG_External::GLES::glTexBufferEXT_PTR>(
+                                         [](GLenum, GLenum, GLuint) {})
+                                   : nullptr;
+        funcs.glTexBufferOES = g_fake.hasOesTexBufferEntryPoint
+                                   ? static_cast<MobileGL::MG_External::GLES::glTexBufferOES_PTR>(
+                                         [](GLenum, GLenum, GLuint) {})
+                                   : nullptr;
+
         // The probe's vertex shader writes the gl_InstanceID it observed into the
         // result SSBO at binding 0. A conforming driver observes 0; a leaking one
         // observes the indirect command's baseInstance word (byte offset 12).
@@ -520,6 +556,150 @@ TEST(FragmentInterpolationCapabilities, QueriesOnlyWhenSupportedAndPreservesDriv
     EXPECT_EQ(funcs.glGetError(), GL_NO_ERROR);
 }
 
+// Buffer textures are core in the OpenGL 3.1+ context MobileGL advertises but need ES 3.2 or
+// EXT/OES_texture_buffer on the host. The tier decides three things at once: whether glTexBuffer
+// may be called at all, which #extension directive the emitted ESSL must carry, and whether
+// GL_MAX_TEXTURE_BUFFER_SIZE is a driver answer or MobileGL's own floor.
+using TextureBufferTier = MobileGL::MG_External::GLESCapabilities::TextureBufferTier;
+
+TEST(BufferTextureCapabilities, Es32ResolvesToCoreAndTakesTheDriverLimit) {
+    ResetFakeDriver();
+    g_fake.maxVertexSsboBlocks = 0;
+    g_fake.glesMinorVersion = 2;
+    const auto funcs = MakeFakeGLESFunctions();
+
+    MobileGL::MG_External::GLESCapabilities caps;
+    ASSERT_TRUE(MobileGL::MG_Util::BackendLoader::FillInGLESCapabilities(caps, funcs));
+
+    EXPECT_EQ(caps.TextureBufferSupport, TextureBufferTier::CoreEs32);
+    EXPECT_TRUE(caps.MaxTextureBufferSizeIsDriverReported);
+    EXPECT_EQ(caps.MaxTextureBufferSize, g_fake.maxTextureBufferSize);
+    EXPECT_TRUE(g_fake.maxTextureBufferSizeQueried);
+}
+
+// The regression this pins: an ES 3.1 driver whose support is GL_EXT_texture_buffer exports
+// glTexBufferEXT and NOT the unsuffixed core name. A resolver that requires the core pointer
+// declares this driver unsupported and then refuses to compile shaders it could have run.
+TEST(BufferTextureCapabilities, Es31WithExtResolvesThroughTheSuffixedEntryPoint) {
+    ResetFakeDriver();
+    g_fake.maxVertexSsboBlocks = 0;
+    g_fake.extensions.emplace_back("GL_EXT_texture_buffer");
+    g_fake.hasCoreTexBufferEntryPoint = false;
+    g_fake.hasExtTexBufferEntryPoint = true;
+    const auto funcs = MakeFakeGLESFunctions();
+
+    MobileGL::MG_External::GLESCapabilities caps;
+    ASSERT_TRUE(MobileGL::MG_Util::BackendLoader::FillInGLESCapabilities(caps, funcs));
+
+    EXPECT_EQ(caps.TextureBufferSupport, TextureBufferTier::ExtensionEXT);
+    EXPECT_TRUE(caps.MaxTextureBufferSizeIsDriverReported);
+    EXPECT_EQ(caps.MaxTextureBufferSize, g_fake.maxTextureBufferSize);
+}
+
+TEST(BufferTextureCapabilities, Es31WithOesResolvesThroughTheSuffixedEntryPoint) {
+    ResetFakeDriver();
+    g_fake.maxVertexSsboBlocks = 0;
+    g_fake.extensions.emplace_back("GL_OES_texture_buffer");
+    g_fake.hasCoreTexBufferEntryPoint = false;
+    g_fake.hasOesTexBufferEntryPoint = true;
+    const auto funcs = MakeFakeGLESFunctions();
+
+    MobileGL::MG_External::GLESCapabilities caps;
+    ASSERT_TRUE(MobileGL::MG_Util::BackendLoader::FillInGLESCapabilities(caps, funcs));
+
+    // The tier, not just a boolean: it is what selects the OES spelling of the #extension
+    // directive SPIRV-Cross hardcodes as EXT.
+    EXPECT_EQ(caps.TextureBufferSupport, TextureBufferTier::ExtensionOES);
+    EXPECT_TRUE(caps.MaxTextureBufferSizeIsDriverReported);
+}
+
+// EXT wins over OES on a driver advertising both, because SPIRV-Cross emits the EXT spelling
+// natively and that tier needs no directive rewriting at all.
+TEST(BufferTextureCapabilities, ExtIsPreferredWhenBothExtensionsArePresent) {
+    ResetFakeDriver();
+    g_fake.maxVertexSsboBlocks = 0;
+    g_fake.extensions.emplace_back("GL_OES_texture_buffer");
+    g_fake.extensions.emplace_back("GL_EXT_texture_buffer");
+    g_fake.hasExtTexBufferEntryPoint = true;
+    g_fake.hasOesTexBufferEntryPoint = true;
+    const auto funcs = MakeFakeGLESFunctions();
+
+    MobileGL::MG_External::GLESCapabilities caps;
+    ASSERT_TRUE(MobileGL::MG_Util::BackendLoader::FillInGLESCapabilities(caps, funcs));
+
+    EXPECT_EQ(caps.TextureBufferSupport, TextureBufferTier::ExtensionEXT);
+}
+
+// The motivating driver (the emulator SDK's ANGLE: ES 3.1, neither extension). The pname is
+// never asked - it would raise GL_INVALID_ENUM - and the floor MobileGL keeps advertising is
+// flagged as not being a driver answer, because an OpenGL 4.x context may not report 0.
+TEST(BufferTextureCapabilities, Es31WithNeitherExtensionIsUnsupportedAndNeverQueriesTheLimit) {
+    ResetFakeDriver();
+    g_fake.maxVertexSsboBlocks = 0;
+    const auto funcs = MakeFakeGLESFunctions();
+
+    MobileGL::MG_External::GLESCapabilities caps;
+    ASSERT_TRUE(MobileGL::MG_Util::BackendLoader::FillInGLESCapabilities(caps, funcs));
+
+    EXPECT_EQ(caps.TextureBufferSupport, TextureBufferTier::None);
+    EXPECT_FALSE(caps.MaxTextureBufferSizeIsDriverReported);
+    EXPECT_FALSE(g_fake.maxTextureBufferSizeQueried);
+    EXPECT_EQ(caps.MaxTextureBufferSize, 65536) << "the OpenGL 3.1 spec floor, not the fake's limit";
+}
+
+// An extension string with no entry point behind it is not support. This is the ES analogue of
+// the multi-draw stub hazard: eglGetProcAddress may hand back live-looking pointers, so the
+// two signals are required together.
+TEST(BufferTextureCapabilities, AnExtensionStringWithoutAnEntryPointIsNotSupport) {
+    ResetFakeDriver();
+    g_fake.maxVertexSsboBlocks = 0;
+    g_fake.extensions.emplace_back("GL_EXT_texture_buffer");
+    g_fake.hasCoreTexBufferEntryPoint = false;
+    g_fake.hasExtTexBufferEntryPoint = false;
+    const auto funcs = MakeFakeGLESFunctions();
+
+    MobileGL::MG_External::GLESCapabilities caps;
+    ASSERT_TRUE(MobileGL::MG_Util::BackendLoader::FillInGLESCapabilities(caps, funcs));
+
+    EXPECT_EQ(caps.TextureBufferSupport, TextureBufferTier::None);
+    EXPECT_FALSE(caps.MaxTextureBufferSizeIsDriverReported);
+}
+
+// A driver that claims buffer textures and then refuses the query is a driver bug. The floor
+// stands in, and the flag says the number was not the driver's - the POST row and the
+// capability log both branch on exactly that.
+TEST(BufferTextureCapabilities, ARejectedLimitQueryIsDrainedAndMarkedAsNotDriverReported) {
+    ResetFakeDriver();
+    g_fake.maxVertexSsboBlocks = 0;
+    g_fake.glesMinorVersion = 2;
+    g_fake.textureBufferSizeQueryRaisesError = true;
+    const auto funcs = MakeFakeGLESFunctions();
+
+    MobileGL::MG_External::GLESCapabilities caps;
+    ASSERT_TRUE(MobileGL::MG_Util::BackendLoader::FillInGLESCapabilities(caps, funcs));
+
+    EXPECT_EQ(caps.TextureBufferSupport, TextureBufferTier::CoreEs32);
+    EXPECT_TRUE(g_fake.maxTextureBufferSizeQueried);
+    EXPECT_FALSE(caps.MaxTextureBufferSizeIsDriverReported);
+    EXPECT_EQ(caps.MaxTextureBufferSize, 65536);
+    EXPECT_EQ(funcs.glGetError(), GL_NO_ERROR) << "the failed query must not leave an error behind";
+}
+
+// A stale error from an earlier probe must not be mistaken for this query failing.
+TEST(BufferTextureCapabilities, AStaleErrorDoesNotDiscardTheDriverLimit) {
+    ResetFakeDriver();
+    g_fake.maxVertexSsboBlocks = 0;
+    g_fake.glesMinorVersion = 2;
+    g_fake.pendingError = GL_INVALID_OPERATION;
+    const auto funcs = MakeFakeGLESFunctions();
+
+    MobileGL::MG_External::GLESCapabilities caps;
+    ASSERT_TRUE(MobileGL::MG_Util::BackendLoader::FillInGLESCapabilities(caps, funcs));
+
+    EXPECT_TRUE(caps.MaxTextureBufferSizeIsDriverReported);
+    EXPECT_EQ(caps.MaxTextureBufferSize, g_fake.maxTextureBufferSize);
+}
+
 TEST(FragmentInterpolationCapabilities, QueryErrorIsDrainedAndFallsBackToCoreMinimums) {
     ResetFakeDriver();
     g_fake.maxVertexSsboBlocks = 0;

MobileGL/MG_Test/Buffer/BufferTest.cpp

@@ -480,6 +480,58 @@ TEST_F(BufferTest, BindBufferRangeZeroUnbindsBindingPoint) {
     EXPECT_EQ(MobileGL::MG_Impl::GLImpl::GetError(), GL_NO_ERROR);
 }
 
+// GL 4.6 core tables 23.4/23.5: *_BUFFER_START and *_BUFFER_SIZE report the (offset, size) pair
+// glBindBufferRange was ASKED for. They are not clamped to the buffer's storage - a range may
+// legally name bytes the buffer does not have, and glBufferData may resize the buffer afterwards
+// without the binding's reported window moving. The size arm used to intersect the recorded range
+// with the buffer's current size, so binding a range on a still-empty buffer (glGenBuffers with no
+// glBufferData - exactly what KHR-GL43.shader_storage_buffer_object.basic-binding does) answered 0
+// while START still answered the offset, an internally inconsistent pair no driver reports.
+TEST_F(BufferTest, IndexedBufferSizeQueryReportsTheRequestedSizeNotTheBuffersStorage) {
+    GLint ssboAlignment = 0;
+    MobileGL::MG_Impl::GLImpl::GetIntegerv(GL_SHADER_STORAGE_BUFFER_OFFSET_ALIGNMENT, &ssboAlignment);
+    ASSERT_GT(ssboAlignment, 0);
+    const GLintptr offset = ssboAlignment;
+    const GLsizeiptr size = 512;
+
+    GLuint buffer = 0;
+    MobileGL::MG_Impl::GLImpl::GenBuffers(1, &buffer);
+    // Deliberately no glBufferData: the name exists, the storage does not.
+    MobileGL::MG_Impl::GLImpl::BindBufferRange(GL_SHADER_STORAGE_BUFFER, 1, buffer, offset, size);
+    ASSERT_EQ(MobileGL::MG_Impl::GLImpl::GetError(), GL_NO_ERROR);
+
+    GLint start32 = 0;
+    GLint size32 = 0;
+    GLint64 start64 = 0;
+    GLint64 size64 = 0;
+    MobileGL::MG_Impl::GLImpl::GetIntegeri_v(GL_SHADER_STORAGE_BUFFER_START, 1, &start32);
+    MobileGL::MG_Impl::GLImpl::GetIntegeri_v(GL_SHADER_STORAGE_BUFFER_SIZE, 1, &size32);
+    MobileGL::MG_Impl::GLImpl::GetInteger64i_v(GL_SHADER_STORAGE_BUFFER_START, 1, &start64);
+    MobileGL::MG_Impl::GLImpl::GetInteger64i_v(GL_SHADER_STORAGE_BUFFER_SIZE, 1, &size64);
+    EXPECT_EQ(start32, static_cast<GLint>(offset));
+    EXPECT_EQ(size32, static_cast<GLint>(size));
+    EXPECT_EQ(start64, static_cast<GLint64>(offset));
+    EXPECT_EQ(size64, static_cast<GLint64>(size));
+
+    // Giving the buffer storage afterwards does not move the window either way.
+    MobileGL::MG_Impl::GLImpl::BindBuffer(GL_SHADER_STORAGE_BUFFER, buffer);
+    MobileGL::MG_Impl::GLImpl::BufferData(GL_SHADER_STORAGE_BUFFER, offset + size, nullptr, GL_DYNAMIC_DRAW);
+    MobileGL::MG_Impl::GLImpl::GetIntegeri_v(GL_SHADER_STORAGE_BUFFER_SIZE, 1, &size32);
+    EXPECT_EQ(size32, static_cast<GLint>(size));
+
+    // glBindBufferBase binds the whole buffer and reports (0, 0), not the buffer's size.
+    MobileGL::MG_Impl::GLImpl::BindBufferBase(GL_SHADER_STORAGE_BUFFER, 1, buffer);
+    MobileGL::MG_Impl::GLImpl::GetIntegeri_v(GL_SHADER_STORAGE_BUFFER_START, 1, &start32);
+    MobileGL::MG_Impl::GLImpl::GetIntegeri_v(GL_SHADER_STORAGE_BUFFER_SIZE, 1, &size32);
+    EXPECT_EQ(start32, 0);
+    EXPECT_EQ(size32, 0);
+
+    MobileGL::MG_Impl::GLImpl::BindBufferBase(GL_SHADER_STORAGE_BUFFER, 1, 0);
+    MobileGL::MG_Impl::GLImpl::BindBuffer(GL_SHADER_STORAGE_BUFFER, 0);
+    MobileGL::MG_Impl::GLImpl::DeleteBuffers(1, &buffer);
+    EXPECT_EQ(MobileGL::MG_Impl::GLImpl::GetError(), GL_NO_ERROR);
+}
+
 TEST_F(BufferTest, GetInteger64vMaxShaderStorageBlockSize) {
     GLint64 maxSsboBlockSize = 0;
     MobileGL::MG_Impl::GLImpl::GetInteger64v(GL_MAX_SHADER_STORAGE_BLOCK_SIZE, &maxSsboBlockSize);

MobileGL/MG_Test/CMakeLists.txt

@@ -78,6 +78,9 @@ add_subdirectory(Query)
 add_subdirectory(Pipeline)
 add_subdirectory(ShaderTranspiler)
 add_subdirectory(Util)
+# The DirectGLES post-transpile ESSL passes are pure String -> String, so unlike the
+# DirectVulkan suite below this one needs no device and always builds.
+add_subdirectory(Backend/DirectGLES)
 if (ENABLE_INTEGRATION_TESTS)
     add_subdirectory(Backend/DirectVulkan)
 endif()

MobileGL/MG_Test/Program/AsyncLinkTest.cpp

@@ -117,11 +117,26 @@ void main() { fragColor = thisIdentifierWasNeverDeclared; }
     }
 
     // The non-joining view of the program, i.e. what GL_COMPLETION_STATUS_KHR will report.
+    // BOTH phases: a program whose SPIR-V job is still in flight is not finished, even though
+    // its whole GL query surface already answers.
     Bool LinkIsSettled(const GLuint program) {
         const auto& object = MG_State::pGLContext->GetProgramObject(program);
         return object == nullptr || object->IsLinkComplete();
     }
 
+    // Phase A alone: the half that decides LINK_STATUS, the info log, and every reflection
+    // query. This is what a read of LINK_STATUS is required to settle.
+    Bool PhaseALinkIsSettled(const GLuint program) {
+        const auto& object = MG_State::pGLContext->GetProgramObject(program);
+        return object == nullptr || object->IsPhaseALinkComplete();
+    }
+
+    // Phase B alone: SPIR-V + the uniform shadow's layout.
+    Bool SpirvIsSettled(const GLuint program) {
+        const auto& object = MG_State::pGLContext->GetProgramObject(program);
+        return object == nullptr || object->IsSpirvComplete();
+    }
+
     // Enqueues `count` distinct heavy compiles without reading anything back, so the pool is
     // left with a real backlog for the caller to race against.
     Vector<GLuint> SaturatePool(const int count, Vector<String>& sourceStorage) {
@@ -516,11 +531,63 @@ TEST_F(AsyncLinkTest, LinkProgramReturnsBeforeTheWorkIsDone) {
 
     for (const GLuint program : programs) {
         EXPECT_EQ(QueryLinkStatus(program), GL_TRUE) << QueryProgramInfoLog(program);
-        EXPECT_TRUE(LinkIsSettled(program)) << "reading LINK_STATUS must have joined";
+        // PHASE A only. Reading LINK_STATUS settles the half that decides it, and no more -
+        // the SPIR-V job may well still be running, which is the entire point of the split.
+        EXPECT_TRUE(PhaseALinkIsSettled(program)) << "reading LINK_STATUS must have joined phase A";
     }
     EXPECT_EQ(GetError(), GL_NO_ERROR);
 }
 
+// The other half of the previous case, and the property the two-phase split exists for:
+// LINK_STATUS is answerable without the SPIR-V, so a run of LINK_STATUS reads over a
+// backlog must leave SPIR-V jobs outstanding rather than draining them one by one.
+TEST_F(AsyncLinkTest, ReadingLinkStatusDoesNotSettleTheSpirvJob) {
+    const AsyncModeScope async(true);
+    MG_Util::Async::ShaderCompilePool::Get().SetMaxConcurrency(1);
+    constexpr int kPrograms = 24;
+
+    const GLuint vs = MakeShader(GL_VERTEX_SHADER, kVs);
+    Vector<GLuint> programs;
+    Vector<String> sources;
+    for (int i = 0; i < kPrograms; ++i) {
+        sources.push_back(MakeBulkySource(7900 + i));
+        const char* text = sources.back().c_str();
+        const GLuint fs = CreateShader(GL_FRAGMENT_SHADER);
+        ShaderSource(fs, 1, &text, nullptr);
+        CompileShader(fs);
+        const GLuint program = CreateProgram();
+        AttachShader(program, vs);
+        AttachShader(program, fs);
+        LinkProgram(program);
+        programs.push_back(program);
+    }
+
+    int spirvOutstanding = 0;
+    for (int i = 0; i < kPrograms; ++i) {
+        const GLuint program = programs[static_cast<SizeT>(i)];
+        EXPECT_EQ(QueryLinkStatus(program), GL_TRUE) << QueryProgramInfoLog(program);
+        EXPECT_TRUE(PhaseALinkIsSettled(program)) << "reading LINK_STATUS must have joined phase A";
+        // Reflection has to answer here too, out of phase A and with no further join.
+        const String uniformName = "uSeed" + std::to_string(7900 + i);
+        EXPECT_GE(GetUniformLocation(program, uniformName.c_str()), 0) << uniformName;
+        if (!SpirvIsSettled(program)) ++spirvOutstanding;
+    }
+    EXPECT_GT(spirvOutstanding, 0) << "the whole GL query surface was answered and yet every SPIR-V job had "
+                                      "already been drained - the reads are joining phase B";
+
+    // And the SPIR-V gate really is a gate: touching it settles the job.
+    for (const GLuint program : programs) {
+        const auto& object = MG_State::pGLContext->GetProgramObject(program);
+        ASSERT_NE(object, nullptr);
+        EXPECT_GT(object->GetGeneratedSpirv().size(), 0u);
+        EXPECT_TRUE(SpirvIsSettled(program));
+        EXPECT_TRUE(LinkIsSettled(program));
+    }
+    EXPECT_EQ(GetError(), GL_NO_ERROR);
+    MG_Util::Async::ShaderCompilePool::Get().SetMaxConcurrency(
+        MG_Util::Async::ShaderCompilePool::Get().GetThreadCount());
+}
+
 // With the flag off, a link is finished by the time glLinkProgram returns. This is the guard
 // that keeps the default shippable.
 TEST_F(AsyncLinkTest, LinkIsFullySynchronousWithAsyncOff) {
@@ -561,8 +628,8 @@ TEST_F(AsyncLinkTest, DrawThroughAPipelineWithAPendingStageProgramJoinsFirst) {
     GLuint pipeline = 0;
     GenProgramPipelines(1, &pipeline);
     ASSERT_NE(pipeline, 0u);
-    // Bind before UseProgramStages: glGenProgramPipelines only reserves the name, and the
-    // first bind is what turns it into an object glUseProgramStages can find.
+    // Bound first only because this test draws through the pipeline; glUseProgramStages no
+    // longer needs it (it materializes a reserved name itself, GL 4.6 core 7.4).
     BindProgramPipeline(pipeline);
     UseProgramStages(pipeline, GL_VERTEX_SHADER_BIT, vsProgram);
     ASSERT_EQ(GetError(), GL_NO_ERROR);

MobileGL/MG_Test/Program/CMakeLists.txt

@@ -49,6 +49,22 @@ add_executable(
         AsyncLinkTest.cpp
 )
 
+add_executable(
+        OptimisticStatusTest
+        OptimisticStatusTest.cpp
+)
+
+target_include_directories(OptimisticStatusTest PRIVATE
+        ${MGL_ROOT}/include
+        ${MGL_ROOT}/MobileGL
+)
+
+target_link_libraries(
+        OptimisticStatusTest PRIVATE
+        GTest::gtest_main
+        ${LINK_LIBRARIES}
+)
+
 target_include_directories(AsyncLinkTest PRIVATE
         ${MGL_ROOT}/include
         ${MGL_ROOT}/MobileGL
@@ -60,6 +76,24 @@ target_link_libraries(
         ${LINK_LIBRARIES}
 )
 
+# Its own binary, like the other async suites: its cases pin the compile pool down to one
+# worker so a phase-B job really is still queued while the GL query surface is being read.
+add_executable(
+        AsyncSpirvPhaseTest
+        AsyncSpirvPhaseTest.cpp
+)
+
+target_include_directories(AsyncSpirvPhaseTest PRIVATE
+        ${MGL_ROOT}/include
+        ${MGL_ROOT}/MobileGL
+)
+
+target_link_libraries(
+        AsyncSpirvPhaseTest PRIVATE
+        GTest::gtest_main
+        ${LINK_LIBRARIES}
+)
+
 add_executable(
         ShaderCompileAdoptionTest
         ShaderCompileAdoptionTest.cpp
@@ -130,6 +164,38 @@ target_link_libraries(
         ${LINK_LIBRARIES}
 )
 
+add_executable(
+        XfbBlockVaryingTest
+        XfbBlockVaryingTest.cpp
+)
+
+target_include_directories(XfbBlockVaryingTest PRIVATE
+        ${MGL_ROOT}/include
+        ${MGL_ROOT}/MobileGL
+)
+
+target_link_libraries(
+        XfbBlockVaryingTest PRIVATE
+        GTest::gtest_main
+        ${LINK_LIBRARIES}
+)
+
+add_executable(
+        ProgramPipelineCompositeTest
+        ProgramPipelineCompositeTest.cpp
+)
+
+target_include_directories(ProgramPipelineCompositeTest PRIVATE
+        ${MGL_ROOT}/include
+        ${MGL_ROOT}/MobileGL
+)
+
+target_link_libraries(
+        ProgramPipelineCompositeTest PRIVATE
+        GTest::gtest_main
+        ${LINK_LIBRARIES}
+)
+
 add_executable(
         ProgramInterfaceTest
         ProgramInterfaceTest.cpp
@@ -161,15 +227,23 @@ include(GoogleTest)
 gtest_discover_tests(ProgramUtilTest DISCOVERY_TIMEOUT 30 PROPERTIES LABELS unit)
 gtest_discover_tests(ProgramTest DISCOVERY_TIMEOUT 30 PROPERTIES LABELS unit)
 gtest_discover_tests(ProgramInterfaceTest DISCOVERY_TIMEOUT 30 PROPERTIES LABELS unit)
+gtest_discover_tests(ProgramPipelineCompositeTest DISCOVERY_TIMEOUT 30 PROPERTIES LABELS unit)
+gtest_discover_tests(XfbBlockVaryingTest DISCOVERY_TIMEOUT 30 PROPERTIES LABELS unit)
 # Heavier than the rest of the unit suite by design: several cases deliberately saturate the
 # compile pool so there is something in flight to race against.
 gtest_discover_tests(AsyncCompileTest DISCOVERY_TIMEOUT 60 PROPERTIES LABELS unit TIMEOUT 300)
 gtest_discover_tests(AsyncLinkTest DISCOVERY_TIMEOUT 60 PROPERTIES LABELS unit TIMEOUT 300)
+# Same reason: every case here links a batch against a one-worker pool so that phase B is
+# genuinely outstanding while phase A is being interrogated.
+gtest_discover_tests(AsyncSpirvPhaseTest DISCOVERY_TIMEOUT 60 PROPERTIES LABELS unit TIMEOUT 300)
 # Same reason: the stage-6 cases keep a backlog in flight so a release really can race a
 # worker, and the 48-object stress links every one of them.
 gtest_discover_tests(ShaderCompileAdoptionTest DISCOVERY_TIMEOUT 60 PROPERTIES LABELS unit TIMEOUT 300)
 # Same reason: the GL_COMPLETION_STATUS_KHR cases saturate a one-worker pool on purpose.
 gtest_discover_tests(ParallelShaderCompileTest DISCOVERY_TIMEOUT 60 PROPERTIES LABELS unit TIMEOUT 300)
+# Same reason: the optimistic-window cases need a saturated one-worker pool to observe an
+# in-flight compile, and the two-phase replay links 48 programs across both flag states.
+gtest_discover_tests(OptimisticStatusTest DISCOVERY_TIMEOUT 60 PROPERTIES LABELS unit TIMEOUT 300)
 gtest_discover_tests(AsyncTeardownTest DISCOVERY_TIMEOUT 60 PROPERTIES LABELS unit TIMEOUT 300)
 # Same reason again: several cases leave A links outstanding while B compiles and links.
 gtest_discover_tests(XfbFrontendOrderInvarianceTest DISCOVERY_TIMEOUT 60 PROPERTIES LABELS unit TIMEOUT 300)

MobileGL/MG_Test/Program/OptimisticStatusTest.cpp

@@ -0,0 +1,674 @@
+// MobileGL - MobileGL/MG_Test/Program/OptimisticStatusTest.cpp
+// Copyright (c) 2025-2026 MobileGL-Dev
+// Licensed under the GNU Lesser General Public License v3.0:
+//   https://www.gnu.org/licenses/gpl-3.0.txt
+//   https://www.gnu.org/licenses/lgpl-3.0.txt
+// SPDX-License-Identifier: LGPL-3.0-only
+// End of Source File Header
+
+// MOBILEGL_ASYNC_OPTIMISTIC_SHADER_STATUS: while a compile job is in flight, the two
+// per-shader queries that would join it - GL_COMPILE_STATUS and the info log - answer
+// optimistically instead, and the first such answer latches for that compile's lifetime
+// (ShaderObject::TakeOptimisticCompileAnswer). These cases pin the corners of that
+// contract: the default still joins, the optimistic window really answers without
+// joining, the latch keeps the three queries telling one story even after the job
+// settles, a real failure still fails the program link with the compile log quoted, and
+// the Iris-shaped two-phase batch produces reflection identical to the joining path.
+//
+// Determinism note: the cases that need "a compile that cannot have settled yet" do not
+// race the pool - they occupy its single concurrency slot with a gate-blocked job
+// (PoolBlocker), so the assertions are hard EXPECTs rather than skip-if-drained guesses.
+// A quirk that silently reverts to joining DEADLOCKS such a case into its 300s ctest
+// timeout instead of passing - ugly, but a failure, which is the point.
+//
+// Like AsyncCompileTest, every case drives the real GL entry points and flips the
+// MG_Config::Features fields itself rather than reading the environment, so one binary
+// asserts both flag states regardless of how the suite was launched.
+
+#include <gtest/gtest.h>
+
+#include <algorithm>
+#include <chrono>
+#include <condition_variable>
+#include <mutex>
+#include <string>
+#include <thread>
+#include <vector>
+
+#include "Config.h"
+#include "Includes.h"
+#include "Init.h"
+#include "MG_Impl/GLImpl/Getter/GL_Getter.h"
+#include "MG_Impl/GLImpl/Program/GL_Program.h"
+#include "MG_State/GLState/Core.h"
+#include "MG_Util/Async/JobNode.h"
+#include "MG_Util/Async/ShaderCompilePool.h"
+
+using namespace MobileGL;
+using namespace MobileGL::MG_Impl::GLImpl;
+
+namespace {
+    class AsyncModeScope {
+    public:
+        explicit AsyncModeScope(const Bool async) : m_saved(MG_Config::Features.AsyncShaderCompile) {
+            MG_Config::Features.AsyncShaderCompile =
+                async ? MG_Config::QuirkOverride::ForceOn : MG_Config::QuirkOverride::ForceOff;
+        }
+        ~AsyncModeScope() { MG_Config::Features.AsyncShaderCompile = m_saved; }
+        AsyncModeScope(const AsyncModeScope&) = delete;
+        AsyncModeScope& operator=(const AsyncModeScope&) = delete;
+
+    private:
+        const MG_Config::QuirkOverride m_saved;
+    };
+
+    class OptimisticStatusScope {
+    public:
+        explicit OptimisticStatusScope(const MG_Config::QuirkOverride mode)
+            : m_saved(MG_Config::Features.AsyncOptimisticShaderStatus) {
+            MG_Config::Features.AsyncOptimisticShaderStatus = mode;
+        }
+        ~OptimisticStatusScope() { MG_Config::Features.AsyncOptimisticShaderStatus = m_saved; }
+        OptimisticStatusScope(const OptimisticStatusScope&) = delete;
+        OptimisticStatusScope& operator=(const OptimisticStatusScope&) = delete;
+
+    private:
+        const MG_Config::QuirkOverride m_saved;
+    };
+
+    // glMaxShaderCompilerThreadsKHR writes PROCESS-wide state (the pool's concurrency budget
+    // and the suspension latch), so a case that touches it has to put both back or it
+    // poisons every case declared after it in this binary.
+    class CompilerThreadScope {
+    public:
+        CompilerThreadScope() = default;
+        ~CompilerThreadScope() {
+            MG_Util::Async::SetAsyncShaderCompileSuspended(false);
+            MG_Util::Async::ShaderCompilePool::Get().SetMaxConcurrency(
+                MG_Util::Async::ShaderCompilePool::Get().GetThreadCount());
+        }
+        CompilerThreadScope(const CompilerThreadScope&) = delete;
+        CompilerThreadScope& operator=(const CompilerThreadScope&) = delete;
+    };
+
+    // A job that occupies a pool slot until released, holding everything queued behind it
+    // in a provably-unsettled state. Same gate idea as JobNodeTest's TestJob+Gate; waiting
+    // on a test-owned gate inside a body does not violate the pool's no-job-waits-on-job
+    // rule - there is no other JOB involved.
+    class PoolBlocker final : public MG_Util::Async::JobNode {
+    public:
+        void Release() {
+            {
+                const std::lock_guard<std::mutex> lock(m_mutex);
+                m_open = true;
+            }
+            m_cv.notify_all();
+        }
+
+    protected:
+        void RunBody() override {
+            std::unique_lock<std::mutex> lock(m_mutex);
+            m_cv.wait(lock, [this] { return m_open; });
+        }
+
+    private:
+        std::mutex m_mutex;
+        std::condition_variable m_cv;
+        Bool m_open = false;
+    };
+
+    // Budget 1 + a blocked job in the only slot: from construction until Release(), no
+    // shader compile posted afterwards can run, let alone settle. The destructor releases
+    // and joins so no case can leak a wedged pool into the next one.
+    class BlockedPoolScope {
+    public:
+        BlockedPoolScope() : m_blocker(MakeShared<PoolBlocker>()) {
+            MaxShaderCompilerThreadsKHR(1);
+            MG_Util::Async::ShaderCompilePool::Get().Post(m_blocker);
+        }
+        ~BlockedPoolScope() { Release(); }
+
+        void Release() {
+            m_blocker->Release();
+            m_blocker->Wait();
+        }
+
+        BlockedPoolScope(const BlockedPoolScope&) = delete;
+        BlockedPoolScope& operator=(const BlockedPoolScope&) = delete;
+
+    private:
+        SharedPtr<PoolBlocker> m_blocker;
+    };
+
+    const char* kBrokenFs = R"(#version 460
+layout(location = 0) out vec4 fragColor;
+void main() { fragColor = thisIdentifierWasNeverDeclared; }
+)";
+
+    // Expensive enough that a compile is not instantaneous, and distinct per index so the
+    // source-hash memo and the stage-6 adoption map never turn a second instance into a
+    // no-op. Callers pass disjoint seed ranges for the same reason - two calls in one case
+    // must never regenerate the same text.
+    String MakeBulkySource(const int index) {
+        String source = "#version 460\nlayout(location = 0) out vec4 fragColor;\n";
+        source += "uniform float uSeed" + std::to_string(index) + ";\n";
+        source += "void main() {\n    float acc = uSeed" + std::to_string(index) + ";\n";
+        for (int i = 0; i < 320; ++i) {
+            source += "    acc = acc * 1.0001 + sin(acc + " + std::to_string(i) + ".0) * cos(acc);\n";
+        }
+        source += "    fragColor = vec4(acc, acc, acc, 1.0);\n}\n";
+        return source;
+    }
+
+    // The two stages of one Iris-shaped program. Distinct per index (so nothing is memoized
+    // across programs) but IDENTICAL between the quirk-off and quirk-on replays of the same
+    // index, which is what makes the reflection comparison meaningful.
+    String MakeIrisVs(const int index) {
+        String source = "#version 460\nlayout(location = 0) in vec3 aPos;\n";
+        source += "uniform mat4 uModel" + std::to_string(index) + ";\n";
+        source += "uniform vec4 uTint;\nout vec4 vColor;\n";
+        source += "void main() {\n    vColor = uTint;\n    gl_Position = uModel" + std::to_string(index) +
+                  " * vec4(aPos, 1.0);\n}\n";
+        return source;
+    }
+    String MakeIrisFs(const int index) {
+        String source = "#version 460\nlayout(location = 0) out vec4 fragColor;\nin vec4 vColor;\n";
+        source += "uniform float uSeed" + std::to_string(index) + ";\nuniform vec2 uOffset;\n";
+        source += "void main() {\n    float acc = uSeed" + std::to_string(index) + " + uOffset.x;\n";
+        for (int i = 0; i < 40; ++i) {
+            source += "    acc = acc * 1.0001 + sin(acc + " + std::to_string(i) + ".0);\n";
+        }
+        source += "    fragColor = vColor + vec4(acc, uOffset.y, 0.0, 1.0);\n}\n";
+        return source;
+    }
+
+    GLuint MakeShader(const GLenum type, const char* source) {
+        const GLuint shader = CreateShader(type);
+        ShaderSource(shader, 1, &source, nullptr);
+        CompileShader(shader);
+        return shader;
+    }
+
+    GLint QueryShaderCompletion(const GLuint shader) {
+        GLint status = -1;
+        GetShaderiv(shader, GL_COMPLETION_STATUS_KHR, &status);
+        return status;
+    }
+
+    GLint QueryCompileStatus(const GLuint shader) {
+        GLint status = GL_FALSE;
+        GetShaderiv(shader, GL_COMPILE_STATUS, &status);
+        return status;
+    }
+
+    GLint QueryInfoLogLength(const GLuint shader) {
+        GLint length = -1;
+        GetShaderiv(shader, GL_INFO_LOG_LENGTH, &length);
+        return length;
+    }
+
+    String QueryShaderInfoLog(const GLuint shader) {
+        std::vector<GLchar> buffer(65536);
+        GLsizei written = 0;
+        GetShaderInfoLog(shader, (GLsizei)buffer.size(), &written, buffer.data());
+        return String(buffer.data(), static_cast<size_t>(written));
+    }
+
+    GLint QueryLinkStatus(const GLuint program) {
+        GLint status = GL_FALSE;
+        GetProgramiv(program, GL_LINK_STATUS, &status);
+        return status;
+    }
+
+    GLint QueryProgramCompletion(const GLuint program) {
+        GLint status = -1;
+        GetProgramiv(program, GL_COMPLETION_STATUS_KHR, &status);
+        return status;
+    }
+
+    String QueryProgramInfoLog(const GLuint program) {
+        // Iris reads through an explicit 32768-byte buffer; mirror that cap so the
+        // log-ordering contract is asserted through the same window the application has.
+        std::vector<GLchar> buffer(32768);
+        GLsizei written = 0;
+        GetProgramInfoLog(program, (GLsizei)buffer.size(), &written, buffer.data());
+        return String(buffer.data(), static_cast<size_t>(written));
+    }
+
+    // Enqueues `count` distinct heavy compiles without reading anything back. Seed bases
+    // must be disjoint across calls within one case (see MakeBulkySource).
+    Vector<GLuint> SaturatePool(const int count, const int seedBase, Vector<String>& sourceStorage) {
+        Vector<GLuint> shaders;
+        shaders.reserve(static_cast<SizeT>(count));
+        for (int i = 0; i < count; ++i) {
+            sourceStorage.push_back(MakeBulkySource(seedBase + i));
+            const char* text = sourceStorage.back().c_str();
+            const GLuint shader = CreateShader(GL_FRAGMENT_SHADER);
+            ShaderSource(shader, 1, &text, nullptr);
+            CompileShader(shader);
+            shaders.push_back(shader);
+        }
+        return shaders;
+    }
+
+    // One program driven through Iris's exact phase-1 shape: create, source, compile, read
+    // the info log then the compile status (GlShader.createShader's order), attach, bind an
+    // attrib, link, detach, delete. NO program-level query of any kind.
+    GLuint RunIrisPhaseOne(const String& vsSource, const String& fsSource) {
+        const char* vsText = vsSource.c_str();
+        const char* fsText = fsSource.c_str();
+
+        const GLuint vs = CreateShader(GL_VERTEX_SHADER);
+        ShaderSource(vs, 1, &vsText, nullptr);
+        CompileShader(vs);
+        (void)QueryShaderInfoLog(vs);
+        (void)QueryCompileStatus(vs);
+
+        const GLuint fs = CreateShader(GL_FRAGMENT_SHADER);
+        ShaderSource(fs, 1, &fsText, nullptr);
+        CompileShader(fs);
+        (void)QueryShaderInfoLog(fs);
+        (void)QueryCompileStatus(fs);
+
+        const GLuint program = CreateProgram();
+        AttachShader(program, vs);
+        AttachShader(program, fs);
+        BindAttribLocation(program, 0, "aPos");
+        LinkProgram(program);
+        DetachShader(program, vs);
+        DetachShader(program, fs);
+        DeleteShader(vs);
+        DeleteShader(fs);
+        return program;
+    }
+
+    // Phase 2, also in Iris's order: LINK_STATUS first, then the by-name location lookups,
+    // then the GL_ACTIVE_UNIFORMS enumeration ProgramUniforms$Builder.buildUniforms does.
+    struct ProgramReflection {
+        GLint linkStatus = GL_FALSE;
+        Vector<std::pair<String, GLint>> locations;                       // queried name -> location
+        Vector<std::tuple<String, GLenum, GLint, GLint>> activeUniforms;  // name, type, size, location
+    };
+
+    ProgramReflection RunIrisPhaseTwo(const GLuint program, const Vector<String>& names) {
+        ProgramReflection out;
+        out.linkStatus = QueryLinkStatus(program);
+
+        for (const String& name : names) {
+            out.locations.emplace_back(name, GetUniformLocation(program, name.c_str()));
+        }
+
+        GLint activeCount = 0;
+        GetProgramiv(program, GL_ACTIVE_UNIFORMS, &activeCount);
+        for (GLint i = 0; i < activeCount; ++i) {
+            GLchar name[128] = {};
+            GLsizei written = 0;
+            GLint size = 0;
+            GLenum type = 0;
+            GetActiveUniform(program, (GLuint)i, (GLsizei)sizeof(name), &written, &size, &type, name);
+            const String nameStr(name, static_cast<size_t>(written));
+            out.activeUniforms.emplace_back(nameStr, type, size, GetUniformLocation(program, name));
+        }
+        // The enumeration order is an implementation detail; the SET is the contract.
+        std::sort(out.activeUniforms.begin(), out.activeUniforms.end());
+        return out;
+    }
+
+    class OptimisticStatusTest : public ::testing::Test {
+    protected:
+        void SetUp() override { MobileGL::Initialize(); }
+    };
+} // namespace
+
+// ---------------------------------------------------------------------------------------
+// The default still joins
+// ---------------------------------------------------------------------------------------
+
+// With the quirk unset (Auto = the shipped default), GL_COMPILE_STATUS on a pending compile
+// must join it: after the query, the node is terminal. This is the case that guards the
+// default against ever silently flipping. No blocker here - a blocked pool would turn the
+// (correct) joining behaviour into a deadlock; a plain backlog only makes the pre-join
+// state likely, and the assertion is valid either way.
+TEST_F(OptimisticStatusTest, OffByDefaultTheStatusStillJoins) {
+    const AsyncModeScope async(true);
+    const OptimisticStatusScope quirk(MG_Config::QuirkOverride::Auto);
+    const CompilerThreadScope threads;
+    MaxShaderCompilerThreadsKHR(1);
+
+    Vector<String> backlog;
+    const Vector<GLuint> saturation = SaturatePool(8, 70000, backlog);
+    const Vector<GLuint> probes = SaturatePool(1, 71000, backlog);
+    const GLuint probe = probes[0];
+
+    EXPECT_EQ(QueryCompileStatus(probe), GL_TRUE);
+    EXPECT_EQ(QueryShaderCompletion(probe), GL_TRUE)
+        << "GL_COMPILE_STATUS with the quirk off must have joined the job";
+
+    for (const GLuint shader : saturation) DeleteShader(shader);
+    DeleteShader(probe);
+    EXPECT_EQ(GetError(), GL_NO_ERROR);
+}
+
+// ---------------------------------------------------------------------------------------
+// The optimistic window, deterministically
+// ---------------------------------------------------------------------------------------
+
+// A compile that provably cannot have settled (the pool's only slot is gate-blocked)
+// answers GL_TRUE / length 0 / empty log, and GL_COMPLETION_STATUS_KHR still reads
+// GL_FALSE after all three - i.e. none of them joined. Hard EXPECTs, no skip: if the
+// quirk silently reverts to joining, the status read deadlocks against the blocked pool
+// and the case fails by timeout.
+TEST_F(OptimisticStatusTest, PendingCompileReportsTrueAndEmptyLogWithoutJoining) {
+    const AsyncModeScope async(true);
+    const OptimisticStatusScope quirk(MG_Config::QuirkOverride::ForceOn);
+    const CompilerThreadScope threads;
+    const BlockedPoolScope blocked;
+
+    Vector<String> storage;
+    const Vector<GLuint> probes = SaturatePool(1, 72000, storage);
+    const GLuint probe = probes[0];
+
+    EXPECT_EQ(QueryCompileStatus(probe), GL_TRUE) << "an in-flight compile must answer GL_TRUE";
+    EXPECT_EQ(QueryInfoLogLength(probe), 0) << "an in-flight compile must answer an empty log length";
+    EXPECT_TRUE(QueryShaderInfoLog(probe).empty()) << "an in-flight compile must answer an empty log";
+    EXPECT_EQ(QueryShaderCompletion(probe), GL_FALSE)
+        << "the three reads above must not have joined the blocked job";
+
+    DeleteShader(probe);
+    EXPECT_EQ(GetError(), GL_NO_ERROR);
+}
+
+// ---------------------------------------------------------------------------------------
+// The latch: one story per compile
+// ---------------------------------------------------------------------------------------
+
+// The torn-pair regression case. A broken shader's log and status are read while the job
+// is provably in flight (optimistic empty/GL_TRUE), the job then settles, and the app
+// re-reads: the latch must keep the answers optimistic - GL_TRUE, empty log - rather than
+// flip to the real GL_FALSE next to the already-consumed empty log. The real failure then
+// surfaces at the link, with the compile error inside the application's 32768-byte read
+// window (the compile log leads the quoted source in ConsumeShaders' format).
+TEST_F(OptimisticStatusTest, LatchKeepsOneStoryPerCompileAndTheLinkCarriesTheDiagnostic) {
+    const AsyncModeScope async(true);
+    const OptimisticStatusScope quirk(MG_Config::QuirkOverride::ForceOn);
+    const CompilerThreadScope threads;
+
+    const GLuint vs = CreateShader(GL_VERTEX_SHADER);
+    const char* vsText =
+        "#version 460\nlayout(location = 0) in vec3 aPos;\nvoid main() { gl_Position = vec4(aPos, 1.0); }\n";
+    ShaderSource(vs, 1, &vsText, nullptr);
+
+    GLuint fs = 0;
+    {
+        const BlockedPoolScope blocked;
+        CompileShader(vs);
+        fs = MakeShader(GL_FRAGMENT_SHADER, kBrokenFs);
+
+        // Iris's order, while nothing can settle: log (empty), then status (GL_TRUE).
+        EXPECT_TRUE(QueryShaderInfoLog(fs).empty());
+        EXPECT_EQ(QueryCompileStatus(fs), GL_TRUE);
+        EXPECT_EQ(QueryShaderCompletion(fs), GL_FALSE);
+    } // blocker released and joined; the broken compile can now settle
+
+    const auto deadline = std::chrono::steady_clock::now() + std::chrono::seconds(30);
+    while (QueryShaderCompletion(fs) == GL_FALSE) {
+        ASSERT_LT(std::chrono::steady_clock::now(), deadline) << "compile job never settled";
+        std::this_thread::sleep_for(std::chrono::milliseconds(1));
+    }
+
+    // Settled - but this shader already told the optimistic story, so it keeps telling it.
+    EXPECT_EQ(QueryCompileStatus(fs), GL_TRUE)
+        << "the latch must keep a queried-while-pending compile optimistic after it settles";
+    EXPECT_EQ(QueryInfoLogLength(fs), 0);
+    EXPECT_TRUE(QueryShaderInfoLog(fs).empty());
+
+    // The truth arrives where the design routes it: at the link.
+    const GLuint program = CreateProgram();
+    AttachShader(program, vs);
+    AttachShader(program, fs);
+    LinkProgram(program);
+    EXPECT_EQ(QueryLinkStatus(program), GL_FALSE) << "a latched-over failure must still fail the link";
+    EXPECT_NE(QueryProgramInfoLog(program).find("thisIdentifierWasNeverDeclared"), String::npos)
+        << "the compile error must lead the program info log, inside a 32768-byte window";
+
+    DeleteProgram(program);
+    DeleteShader(vs);
+    DeleteShader(fs);
+    EXPECT_EQ(GetError(), GL_NO_ERROR);
+}
+
+// A shader whose FIRST query arrives after the job settled was never answered
+// optimistically, so it owes no continuity: the truth comes straight back. (The
+// completion poll does not engage the latch - it is the extension's own non-joining
+// query and always tells the truth.)
+TEST_F(OptimisticStatusTest, OnceTerminalAnUnqueriedShaderTellsTheTruth) {
+    const AsyncModeScope async(true);
+    const OptimisticStatusScope quirk(MG_Config::QuirkOverride::ForceOn);
+
+    const GLuint fs = MakeShader(GL_FRAGMENT_SHADER, kBrokenFs);
+
+    const auto deadline = std::chrono::steady_clock::now() + std::chrono::seconds(30);
+    while (QueryShaderCompletion(fs) == GL_FALSE) {
+        ASSERT_LT(std::chrono::steady_clock::now(), deadline) << "compile job never settled";
+        std::this_thread::sleep_for(std::chrono::milliseconds(1));
+    }
+
+    EXPECT_EQ(QueryCompileStatus(fs), GL_FALSE) << "no optimistic answer was given, so no latch holds";
+    EXPECT_GT(QueryInfoLogLength(fs), 0);
+    EXPECT_NE(QueryShaderInfoLog(fs).find("thisIdentifierWasNeverDeclared"), String::npos);
+    DeleteShader(fs);
+    EXPECT_EQ(GetError(), GL_NO_ERROR);
+}
+
+// Recompiling resets the story: a latched optimistic answer must not survive a source
+// change (the latch clears when the node changes hands or goes away).
+TEST_F(OptimisticStatusTest, ANewCompileResetsTheLatch) {
+    const AsyncModeScope async(true);
+    const OptimisticStatusScope quirk(MG_Config::QuirkOverride::ForceOn);
+    const CompilerThreadScope threads;
+
+    GLuint fs = 0;
+    {
+        const BlockedPoolScope blocked;
+        fs = MakeShader(GL_FRAGMENT_SHADER, kBrokenFs);
+        EXPECT_EQ(QueryCompileStatus(fs), GL_TRUE); // latches
+    }
+
+    // New source, new compile, no query before it settles.
+    const char* goodFs = "#version 460\nlayout(location = 0) out vec4 fragColor;\n"
+                         "void main() { fragColor = vec4(1.0); }\n";
+    ShaderSource(fs, 1, &goodFs, nullptr);
+    CompileShader(fs);
+    const auto deadline = std::chrono::steady_clock::now() + std::chrono::seconds(30);
+    while (QueryShaderCompletion(fs) == GL_FALSE) {
+        ASSERT_LT(std::chrono::steady_clock::now(), deadline) << "recompile never settled";
+        std::this_thread::sleep_for(std::chrono::milliseconds(1));
+    }
+    EXPECT_EQ(QueryCompileStatus(fs), GL_TRUE);
+    EXPECT_TRUE(QueryShaderInfoLog(fs).empty());
+    DeleteShader(fs);
+    EXPECT_EQ(GetError(), GL_NO_ERROR);
+}
+
+// ---------------------------------------------------------------------------------------
+// Failure still fails, at the link, inside the application's read window
+// ---------------------------------------------------------------------------------------
+
+// A broken fragment shader whose compile status was answered optimistically still fails
+// its program link, and the compile error is readable through a 32768-byte
+// glGetProgramInfoLog - the compile log LEADS the quoted source in ConsumeShaders'
+// format, so even this >32KB shader source cannot push it out of the window.
+TEST_F(OptimisticStatusTest, AFailingCompileStillFailsItsLink) {
+    const AsyncModeScope async(true);
+    const OptimisticStatusScope quirk(MG_Config::QuirkOverride::ForceOn);
+
+    // A >32KB broken fragment shader: the undeclared identifier sits at the top, then bulk.
+    String brokenSource = "#version 460\nlayout(location = 0) out vec4 fragColor;\n";
+    brokenSource += "void main() {\n    float acc = thisIdentifierWasNeverDeclared;\n";
+    for (int i = 0; i < 900; ++i) {
+        brokenSource += "    acc = acc * 1.0001 + sin(acc + " + std::to_string(i) + ".0) * cos(acc);\n";
+    }
+    brokenSource += "    fragColor = vec4(acc);\n}\n";
+    ASSERT_GT(brokenSource.size(), 32768u);
+
+    const GLuint vs = MakeShader(GL_VERTEX_SHADER,
+                                 "#version 460\nlayout(location = 0) in vec3 aPos;\n"
+                                 "void main() { gl_Position = vec4(aPos, 1.0); }\n");
+    const char* brokenText = brokenSource.c_str();
+    const GLuint fs = CreateShader(GL_FRAGMENT_SHADER);
+    ShaderSource(fs, 1, &brokenText, nullptr);
+    CompileShader(fs);
+    (void)QueryShaderInfoLog(fs);
+    (void)QueryCompileStatus(fs); // may latch optimistic GL_TRUE; must not matter
+
+    const GLuint program = CreateProgram();
+    AttachShader(program, vs);
+    AttachShader(program, fs);
+    LinkProgram(program);
+
+    EXPECT_EQ(QueryLinkStatus(program), GL_FALSE) << "a hidden compile failure must still fail the link";
+    const String log = QueryProgramInfoLog(program);
+    EXPECT_NE(log.find("thisIdentifierWasNeverDeclared"), String::npos)
+        << "the compile error must be readable through a 32768-byte program info log window";
+
+    DeleteProgram(program);
+    DeleteShader(vs);
+    DeleteShader(fs);
+    EXPECT_EQ(GetError(), GL_NO_ERROR);
+}
+
+// ---------------------------------------------------------------------------------------
+// The Iris two-phase replay
+// ---------------------------------------------------------------------------------------
+
+// THE LOAD-BEARING CASE. 24 programs through Iris's exact phase-1 shape (compile, read log
+// then status per shader, link, detach, delete - no program query), then phase 2 (link
+// status, by-name locations including an absent name, the active-uniform enumeration).
+// Every location and every active-uniform record must equal what the identical sequence
+// produces with the quirk off.
+//
+// Two determinism guards make this a real A/B rather than a tautology:
+//   * The quirk-on arm runs FIRST, against a cold preprocess cache, and the reference arm
+//     second - so it is the path under test that pays the full pipeline, not the control.
+//   * The quirk-on arm's phase 1 runs over a BLOCKED pool, and every program is then
+//     WITNESSED still-incomplete (GL_COMPLETION_STATUS_KHR == GL_FALSE) before the pool
+//     is released: proof that no phase-1 call joined, i.e. the quirk was really engaged.
+//     A quirk that silently reverts to joining deadlocks here and fails by timeout.
+TEST_F(OptimisticStatusTest, IrisTwoPhaseReplayProducesIdenticalReflection) {
+    constexpr int kPrograms = 24;
+
+    Vector<ProgramReflection> reference;
+    Vector<ProgramReflection> optimistic;
+
+    for (const Bool quirkOn : {true, false}) {
+        const AsyncModeScope async(true);
+        const OptimisticStatusScope quirk(quirkOn ? MG_Config::QuirkOverride::ForceOn
+                                                  : MG_Config::QuirkOverride::ForceOff);
+        const CompilerThreadScope threads;
+
+        Vector<String> vsSources, fsSources;
+        for (int i = 0; i < kPrograms; ++i) {
+            vsSources.push_back(MakeIrisVs(i));
+            fsSources.push_back(MakeIrisFs(i));
+        }
+
+        Vector<GLuint> programs;
+        if (quirkOn) {
+            const BlockedPoolScope blocked;
+            for (int i = 0; i < kPrograms; ++i) {
+                programs.push_back(RunIrisPhaseOne(vsSources[(SizeT)i], fsSources[(SizeT)i]));
+            }
+            // The witness: phase 1 finished with the pool blocked, so nothing can have
+            // settled and nothing can have been joined - every link must still be pending.
+            for (int i = 0; i < kPrograms; ++i) {
+                ASSERT_EQ(QueryProgramCompletion(programs[(SizeT)i]), GL_FALSE)
+                    << "program " << i << " settled under a blocked pool - a phase-1 call must have joined";
+            }
+        } else {
+            for (int i = 0; i < kPrograms; ++i) {
+                programs.push_back(RunIrisPhaseOne(vsSources[(SizeT)i], fsSources[(SizeT)i]));
+            }
+        }
+
+        Vector<ProgramReflection>& out = quirkOn ? optimistic : reference;
+        for (int i = 0; i < kPrograms; ++i) {
+            const Vector<String> names = {"uModel" + std::to_string(i), "uTint",
+                                          "uSeed" + std::to_string(i), "uOffset", "uDoesNotExist"};
+            out.push_back(RunIrisPhaseTwo(programs[(SizeT)i], names));
+        }
+        for (const GLuint program : programs) DeleteProgram(program);
+        ASSERT_EQ(GetError(), GL_NO_ERROR);
+    }
+
+    ASSERT_EQ(reference.size(), optimistic.size());
+    for (SizeT i = 0; i < reference.size(); ++i) {
+        EXPECT_EQ(reference[i].linkStatus, GL_TRUE) << "program " << i;
+        EXPECT_EQ(optimistic[i].linkStatus, GL_TRUE) << "program " << i;
+        EXPECT_EQ(reference[i].locations, optimistic[i].locations)
+            << "program " << i << ": by-name locations diverged under the quirk";
+        EXPECT_EQ(reference[i].activeUniforms, optimistic[i].activeUniforms)
+            << "program " << i << ": active-uniform enumeration diverged under the quirk";
+        // The absent name answers -1 in both worlds.
+        EXPECT_EQ(reference[i].locations.back().second, -1) << "program " << i;
+    }
+}
+
+// ---------------------------------------------------------------------------------------
+// The concurrency observable
+// ---------------------------------------------------------------------------------------
+
+// The crisp A/B that phase 1 stopped joining. Quirk-on arm: the phase-1 shape over a
+// blocked pool completes without joining anything - every shader is then provably still
+// in flight (hard EXPECT; an inert quirk deadlocks and fails by timeout). Quirk-off arm:
+// the same shape joins at every status read, so nothing is left in flight afterwards.
+TEST_F(OptimisticStatusTest, PhaseOneIssuesNoCompileJoin) {
+    const AsyncModeScope async(true);
+    const CompilerThreadScope threads;
+
+    // Quirk on: nothing settles, nothing joins.
+    {
+        const OptimisticStatusScope quirk(MG_Config::QuirkOverride::ForceOn);
+        const BlockedPoolScope blocked;
+        Vector<String> storage;
+        Vector<GLuint> shaders;
+        for (int i = 0; i < 12; ++i) {
+            storage.push_back(MakeBulkySource(90000 + i));
+            const char* text = storage.back().c_str();
+            const GLuint fs = CreateShader(GL_FRAGMENT_SHADER);
+            ShaderSource(fs, 1, &text, nullptr);
+            CompileShader(fs);
+            (void)QueryShaderInfoLog(fs);
+            (void)QueryCompileStatus(fs);
+            shaders.push_back(fs);
+        }
+        for (const GLuint shader : shaders) {
+            EXPECT_EQ(QueryShaderCompletion(shader), GL_FALSE)
+                << "a phase-1 read joined a compile the blocked pool could not have run";
+        }
+        for (const GLuint shader : shaders) DeleteShader(shader);
+    }
+
+    // Quirk off: every status read joins its shader.
+    {
+        const OptimisticStatusScope quirk(MG_Config::QuirkOverride::ForceOff);
+        MaxShaderCompilerThreadsKHR(1);
+        Vector<String> storage;
+        Vector<GLuint> shaders;
+        for (int i = 0; i < 12; ++i) {
+            storage.push_back(MakeBulkySource(80000 + i));
+            const char* text = storage.back().c_str();
+            const GLuint fs = CreateShader(GL_FRAGMENT_SHADER);
+            ShaderSource(fs, 1, &text, nullptr);
+            CompileShader(fs);
+            (void)QueryShaderInfoLog(fs);
+            (void)QueryCompileStatus(fs);
+            shaders.push_back(fs);
+        }
+        for (const GLuint shader : shaders) {
+            EXPECT_EQ(QueryShaderCompletion(shader), GL_TRUE)
+                << "with the quirk off every per-shader status read must have joined";
+        }
+        for (const GLuint shader : shaders) DeleteShader(shader);
+    }
+    EXPECT_EQ(GetError(), GL_NO_ERROR);
+}

MobileGL/MG_Test/Program/ParallelShaderCompileTest.cpp

@@ -239,7 +239,15 @@ TEST_F(ParallelShaderCompileTest, ProgramCompletionStatusReportsFalseWithoutJoin
 
     for (const GLuint program : programs) {
         EXPECT_EQ(QueryLinkStatus(program), GL_TRUE);
-        EXPECT_EQ(QueryProgramCompletion(program), GL_TRUE) << "GL_LINK_STATUS must have joined";
+        // GL_COMPLETION_STATUS_KHR spans BOTH phases of a link, so reading GL_LINK_STATUS -
+        // which is answered out of phase A - is no longer enough to turn it GL_TRUE. That is
+        // deliberate: an application that polls completion and then draws must not be told
+        // "done" while the SPIR-V is still being generated, or the draw it was cleared for is
+        // the thing that blocks. Settling both phases is what makes the query true.
+        const auto& object = MG_State::pGLContext->GetProgramObject(program);
+        ASSERT_NE(object, nullptr);
+        object->JoinLinkAndSpirv();
+        EXPECT_EQ(QueryProgramCompletion(program), GL_TRUE) << "a full join must have settled both phases";
     }
     EXPECT_EQ(GetError(), GL_NO_ERROR);
 }
@@ -333,6 +341,54 @@ TEST_F(ParallelShaderCompileTest, ZeroCompilerThreadsJoinsEverythingAndCompilesI
     EXPECT_EQ(GetError(), GL_NO_ERROR);
 }
 
+// The same obligation, but for LINKS that are already in flight when the zero count arrives -
+// and specifically for BOTH phases of one. A link is two chained jobs now (ProgramLinkTask,
+// then ProgramSpirvTask), and GL_COMPLETION_STATUS_KHR spans both, so
+// ProgramState::JoinAllPendingWork has to settle both or this query reads GL_FALSE in the one
+// mode the extension says cannot have anything pending. The case above creates its program
+// AFTER the zero count, so it links inline and cannot see this; here the programs are linked
+// against a saturated pool BEFORE it.
+TEST_F(ParallelShaderCompileTest, ZeroCompilerThreadsJoinsPendingLinksAndTheirSpirvJobs) {
+    const AsyncModeScope async(true);
+    const CompilerThreadScope threads;
+    MaxShaderCompilerThreadsKHR(1);
+
+    // A backlog first, so the links below cannot all drain before the zero count lands.
+    Vector<String> sources;
+    (void)EnqueueBacklog(24, 5000, sources);
+
+    Vector<GLuint> programs;
+    for (int i = 0; i < 8; ++i) {
+        sources.push_back(MakeBulkySource(5100 + i));
+        const char* text = sources.back().c_str();
+        const GLuint fs = CreateShader(GL_FRAGMENT_SHADER);
+        ShaderSource(fs, 1, &text, nullptr);
+        CompileShader(fs);
+        const GLuint vs = MakeShader(GL_VERTEX_SHADER, kVs);
+        CompileShader(vs); // this file's MakeShader only sources; it does not compile
+        const GLuint program = CreateProgram();
+        AttachShader(program, vs);
+        AttachShader(program, fs);
+        LinkProgram(program);
+        programs.push_back(program);
+    }
+
+    int outstanding = 0;
+    for (const GLuint program : programs) {
+        if (QueryProgramCompletion(program) == GL_FALSE) ++outstanding;
+    }
+
+    MaxShaderCompilerThreadsKHR(0);
+
+    for (const GLuint program : programs) {
+        EXPECT_EQ(QueryProgramCompletion(program), GL_TRUE)
+            << "glMaxShaderCompilerThreadsKHR(0) must leave neither link phase in flight";
+        EXPECT_EQ(QueryLinkStatus(program), GL_TRUE);
+    }
+    EXPECT_GT(outstanding, 0) << "every link had drained before the zero count; this case proved nothing";
+    EXPECT_EQ(GetError(), GL_NO_ERROR);
+}
+
 // ...and a later NONZERO count is what lifts it. Nothing else does: not a new context, not a
 // join, not eglInitialize. That is the documented contract, so it gets an assertion.
 TEST_F(ParallelShaderCompileTest, NonzeroCompilerThreadsRestoresAsynchronousCompilation) {

MobileGL/MG_Test/Program/ProgramInterfaceTest.cpp

@@ -1100,4 +1100,119 @@ void main() { color = u + v; }
         EXPECT_EQ(viaActiveUniformBlockiv, 5);
         EXPECT_EQ(TakeError(), GL_NO_ERROR);
     }
+
+    // ---------------------------------------------------- queries on an unlinked program ----
+    // glGetProgramiv is legal on a program that has never linked - GL 4.6 sec. 7.3 says the
+    // queried state simply has its initial value - but the reflection-backed pnames read
+    // Artifacts().program, which is null until a link produces one. That dereference was a
+    // SIGSEGV inside glslang::TProgram::getNumPipeInputs, and KHR-GL30.api.coverage walks into it
+    // (it queries GL_ACTIVE_ATTRIBUTES right after a glGetAttribLocation that failed). It only
+    // became reachable once the glCopyTexImage2D throw ahead of it in the same case stopped
+    // killing the run first.
+    TEST_F(ProgramInterfaceTest, ReflectionQueriesOnAnUnlinkedProgramAnswerZero) {
+        const GLuint neverLinked = CreateProgram();
+        ASSERT_NE(neverLinked, 0u);
+        ClearErrors();
+
+        for (const GLenum pname : {GL_ACTIVE_ATTRIBUTES, GL_ACTIVE_ATTRIBUTE_MAX_LENGTH, GL_ACTIVE_UNIFORMS,
+                                   GL_ACTIVE_UNIFORM_MAX_LENGTH, GL_ACTIVE_UNIFORM_BLOCKS,
+                                   GL_ACTIVE_ATOMIC_COUNTER_BUFFERS}) {
+            GLint value = -1;
+            GetProgramiv(neverLinked, pname, &value);
+            ClearErrors();
+            EXPECT_GE(value, 0) << "pname 0x" << std::hex << pname << " left its output untouched";
+        }
+
+        // A program that was linked and FAILED is the shape api.coverage actually hits.
+        const GLuint brokenSource = MakeProgram("#version 430\nvoid main() { this is not glsl }\n", kSimpleFs);
+        LinkProgram(brokenSource);
+        ClearErrors();
+        GLint linked = GL_TRUE;
+        GetProgramiv(brokenSource, GL_LINK_STATUS, &linked);
+        ASSERT_EQ(linked, GL_FALSE) << "the shader was supposed to fail to compile";
+        ClearErrors();
+
+        GLint attributes = -1;
+        GetProgramiv(brokenSource, GL_ACTIVE_ATTRIBUTES, &attributes);
+        ClearErrors();
+        EXPECT_EQ(attributes, 0);
+
+        // GL_COMPUTE_WORK_GROUP_SIZE is GL_INVALID_OPERATION on a program that has not linked (GL
+        // 4.6 sec. 7.13), so it is allowed to leave the output alone - but it still reaches
+        // GetComputeLocalSize(), and it may not do so through a null reflection.
+        GLint localSize[3] = {-1, -1, -1};
+        GetProgramiv(brokenSource, GL_COMPUTE_WORK_GROUP_SIZE, localSize);
+        const GLenum computeError = TakeError();
+        ClearErrors();
+        EXPECT_TRUE(computeError == GL_INVALID_OPERATION || (localSize[0] == 0 && localSize[1] == 0 &&
+                                                             localSize[2] == 0))
+            << "either the query is refused, or it answers the initial value - never both untouched "
+               "and unreported";
+    }
+
+    // ------------------------------------------------------------- length on every path ----
+    // glGetProgramResourceiv's *length is the caller's only signal for how many entries params
+    // holds, and callers are entitled to leave it uninitialised: the CTS declares `GLsizei
+    // length;` next to a 1000-entry stack array and then loops `for (i = 0; i < length; ++i)`
+    // (gl4cProgramInterfaceQueryTests.cpp:2172). Leaving it untouched on an error path therefore
+    // does not "return nothing" - it hands the caller whatever was on its stack and makes it walk
+    // that far. KHR-GL43.program_interface_query.subroutines-vertex read 0x20202020 ("    ")
+    // entries and took the process down on BOTH backends. So: zero on every exit, real count on
+    // success. Poisoning with the exact CTS-observed value keeps the assertion honest.
+    TEST_F(ProgramInterfaceTest, GetProgramResourceivReportsLengthOnEveryExitPath) {
+        const GLuint p = MakeProgram(kSimpleVs, kSimpleFs);
+        BindAttribLocation(p, 0, "position");
+        BindFragDataLocation(p, 0, "color");
+        LinkProgram(p);
+        ExpectLinked(p);
+        ClearErrors();
+
+        constexpr GLsizei kPoison = 0x20202020;
+        constexpr GLsizei kBufSize = 16;
+        GLint params[kBufSize] = {};
+
+        const GLenum nameLengthProp = GL_NAME_LENGTH;
+        const GLenum compatibleSubroutinesProp = GL_COMPATIBLE_SUBROUTINES;
+        const GLenum notAProp = GL_TEXTURE_2D;
+
+        const auto lengthAfter = [&](GLuint program, GLenum iface, GLuint index, GLsizei propCount,
+                                     const GLenum* props, GLsizei bufSize, GLint* out) {
+            GLsizei length = kPoison;
+            GetProgramResourceiv(program, iface, index, propCount, props, bufSize, &length, out);
+            ClearErrors();
+            return length;
+        };
+
+        // The case that actually crashed: no subroutine reflection exists, so the query errors
+        // out - and the caller then trusts *length.
+        EXPECT_EQ(lengthAfter(p, GL_VERTEX_SUBROUTINE_UNIFORM, 0, 1, &compatibleSubroutinesProp, kBufSize, params), 0)
+            << "GL_VERTEX_SUBROUTINE_UNIFORM";
+        // Not a program name.
+        EXPECT_EQ(lengthAfter(p + 4242, GL_UNIFORM, 0, 1, &nameLengthProp, kBufSize, params), 0) << "bad program";
+        // Not an interface enum.
+        EXPECT_EQ(lengthAfter(p, GL_TEXTURE_2D, 0, 1, &nameLengthProp, kBufSize, params), 0) << "bad interface";
+        // propCount <= 0, bufSize < 0.
+        EXPECT_EQ(lengthAfter(p, GL_PROGRAM_OUTPUT, 0, 0, &nameLengthProp, kBufSize, params), 0) << "propCount 0";
+        EXPECT_EQ(lengthAfter(p, GL_PROGRAM_OUTPUT, 0, 1, &nameLengthProp, -1, params), 0) << "negative bufSize";
+        // props == nullptr.
+        EXPECT_EQ(lengthAfter(p, GL_PROGRAM_OUTPUT, 0, 1, nullptr, kBufSize, params), 0) << "null props";
+        // A prop this command does not know at all.
+        EXPECT_EQ(lengthAfter(p, GL_PROGRAM_OUTPUT, 0, 1, &notAProp, kBufSize, params), 0) << "unknown prop";
+        // A prop it knows but this interface does not carry.
+        EXPECT_EQ(lengthAfter(p, GL_PROGRAM_OUTPUT, 0, 1, &compatibleSubroutinesProp, kBufSize, params), 0)
+            << "prop/interface mismatch";
+        // Index past the end of a real interface.
+        EXPECT_EQ(lengthAfter(p, GL_PROGRAM_OUTPUT, 9999, 1, &nameLengthProp, kBufSize, params), 0) << "bad index";
+        // Nowhere to put the values.
+        EXPECT_EQ(lengthAfter(p, GL_PROGRAM_OUTPUT, 0, 1, &nameLengthProp, kBufSize, nullptr), 0) << "null params";
+
+        // ...and the success path still reports the count it actually wrote.
+        const GLuint outputIndex = GetProgramResourceIndex(p, GL_PROGRAM_OUTPUT, "color");
+        ASSERT_NE(outputIndex, GL_INVALID_INDEX);
+        GLsizei length = kPoison;
+        GetProgramResourceiv(p, GL_PROGRAM_OUTPUT, outputIndex, 1, &nameLengthProp, kBufSize, &length, params);
+        EXPECT_EQ(TakeError(), GL_NO_ERROR);
+        EXPECT_EQ(length, 1);
+        EXPECT_EQ(params[0], 6) << "GL_NAME_LENGTH counts the terminator";
+    }
 } // namespace

MobileGL/MG_Test/Program/ProgramPipelineCompositeTest.cpp

@@ -0,0 +1,505 @@
+// MobileGL - MobileGL/MG_Test/Program/ProgramPipelineCompositeTest.cpp
+// Copyright (c) 2025-2026 MobileGL-Dev
+// Licensed under the GNU Lesser General Public License v3.0:
+//   https://www.gnu.org/licenses/gpl-3.0.txt
+//   https://www.gnu.org/licenses/lgpl-3.0.txt
+// SPDX-License-Identifier: LGPL-3.0-only
+// End of Source File Header
+
+// The hidden composite program a pipeline draw goes through (MG_State/GLState/Core.cpp,
+// GetProgramForDraw), interrogated directly rather than through pixels.
+//
+// Two properties live here that the integration scenarios cannot see, because both are about
+// the composite as an OBJECT rather than about what it paints:
+//
+//   1. WHICH stage's uniform value ends up in its single slot when several stages declare the
+//      same name. The rendering cases pin the answer for the shapes an application actually
+//      writes; these pin the rule itself, including the tie.
+//   2. WHETHER it is the same object from one draw to the next. A composite rebuild is a full
+//      synchronous Link() plus a new program identity that empties both backends' per-program
+//      registries, and nothing about the resulting IMAGE would change if it happened on every
+//      draw - so an assertion on pixels can never catch that regression.
+
+#include <gtest/gtest.h>
+
+#include <cstring>
+#include <string>
+#include <vector>
+
+#include "Config.h"
+#include "Includes.h"
+#include "Init.h"
+#include "MG_Impl/GLImpl/Getter/GL_Getter.h"
+#include "MG_Impl/GLImpl/Program/GL_Program.h"
+#include "MG_Impl/GLImpl/Program/GL_ProgramPipeline.h"
+#include "MG_State/GLState/Core.h"
+
+using namespace MobileGL;
+using namespace MobileGL::MG_Impl::GLImpl;
+
+namespace {
+
+    // Both stages declare `u_shared`, which is the shared-header idiom (one header included by
+    // every stage) and the shape that used to render nothing: the fragment stage's untouched
+    // zero default overwrote the vertex stage's written value on the way into the composite.
+    const char* kSharedUniformVs = R"(#version 430 core
+out gl_PerVertex { vec4 gl_Position; };
+uniform vec4 u_shared;
+uniform vec4 u_vsOnly;
+void main() { gl_Position = u_shared + u_vsOnly; }
+)";
+
+    const char* kSharedUniformFs = R"(#version 430 core
+uniform vec4 u_shared;
+out vec4 o_color;
+void main() { o_color = u_shared; }
+)";
+
+    const char* kArrayUniformVs = R"(#version 430 core
+out gl_PerVertex { vec4 gl_Position; };
+uniform vec4 u_arr[4];
+void main() { gl_Position = u_arr[0] + u_arr[1] + u_arr[2] + u_arr[3]; }
+)";
+
+    const char* kArrayUniformFs = R"(#version 430 core
+uniform vec4 u_arr[4];
+out vec4 o_color;
+void main() { o_color = u_arr[0] + u_arr[1] + u_arr[2] + u_arr[3]; }
+)";
+
+    const char* kSamplerVs = R"(#version 430 core
+out gl_PerVertex { vec4 gl_Position; };
+void main() { gl_Position = vec4(0.0, 0.0, 0.0, 1.0); }
+)";
+
+    const char* kSamplerFs = R"(#version 430 core
+uniform sampler2D u_tex;
+out vec4 o_color;
+void main() { o_color = texture(u_tex, vec2(0.0)); }
+)";
+
+    class ProgramPipelineCompositeTest : public ::testing::Test {
+    protected:
+        void SetUp() override { MobileGL::Initialize(); }
+
+        // Built by hand rather than through glCreateShaderProgramv, for the reason AsyncLinkTest
+        // gives: that entry point detaches the shader right after linking, so a relink would
+        // leave the stage program with nothing to composite from - and one of the cases below
+        // relinks on purpose.
+        GLuint MakeSeparableProgram(const GLenum stage, const char* source) {
+            const GLuint shader = CreateShader(stage);
+            ShaderSource(shader, 1, &source, nullptr);
+            CompileShader(shader);
+            const GLuint program = CreateProgram();
+            ProgramParameteri(program, GL_PROGRAM_SEPARABLE, GL_TRUE);
+            AttachShader(program, shader);
+            LinkProgram(program);
+            GLint linked = GL_FALSE;
+            GetProgramiv(program, GL_LINK_STATUS, &linked);
+            EXPECT_EQ(linked, GL_TRUE) << "separable stage program did not link";
+            return program;
+        }
+
+        // The composite the next draw would run, settled.
+        static SharedPtr<MG_State::GLState::ProgramObject> DrawProgram() {
+            return MG_State::pGLContext->GetProgramForDraw();
+        }
+
+        // A uniform's value read out of a program's own shadow, by name. This is what the draw
+        // would upload, which is the thing under test - glGetUniform* would answer the same for
+        // the STAGE programs but has no way to name the composite at all.
+        static std::vector<float> ReadVec4(MG_State::GLState::ProgramObject& program, const String& name) {
+            const Int location = program.GetUniformLocation(name);
+            if (location < 0) return {};
+            const Uint offset = program.GetUniformOffset(static_cast<Uint>(location));
+            const auto* ubo = static_cast<const char*>(program.GetUBOData());
+            if (ubo == nullptr || offset == MG_State::GLState::ProgramObject::kInvalidUniformOffset ||
+                offset + 4 * sizeof(float) > program.GetUBOSize()) {
+                return {};
+            }
+            std::vector<float> value(4);
+            std::memcpy(value.data(), ubo + offset, 4 * sizeof(float));
+            return value;
+        }
+    };
+
+} // namespace
+
+// ---------------------------------------------------------------------------------------
+// Which stage wins the composite's single slot
+// ---------------------------------------------------------------------------------------
+
+// THE defect. Both stages declare `u_shared`; only the VERTEX program is ever written to.
+// Walking the stages in order and copying every active uniform unconditionally meant the
+// fragment stage's untouched zero default landed last and won, so the composite drew zeros - a
+// whole frame of nothing, from a program that had been set up entirely correctly.
+TEST_F(ProgramPipelineCompositeTest, AWrittenStageValueIsNotClobberedByAnotherStagesUntouchedDeclaration) {
+    const GLuint vs = MakeSeparableProgram(GL_VERTEX_SHADER, kSharedUniformVs);
+    const GLuint fs = MakeSeparableProgram(GL_FRAGMENT_SHADER, kSharedUniformFs);
+
+    GLuint pipeline = 0;
+    GenProgramPipelines(1, &pipeline);
+    BindProgramPipeline(pipeline);
+    UseProgramStages(pipeline, GL_VERTEX_SHADER_BIT, vs);
+    UseProgramStages(pipeline, GL_FRAGMENT_SHADER_BIT, fs);
+    ASSERT_EQ(GetError(), GL_NO_ERROR);
+
+    // Exactly what an application does: point glUniform* at the vertex stage and write there.
+    // The fragment program is never written to and holds nothing but GL's zero default.
+    ActiveShaderProgram(pipeline, vs);
+    const GLint location = GetUniformLocation(vs, "u_shared");
+    ASSERT_GE(location, 0);
+    const float written[4] = {0.25f, 0.5f, 0.75f, 1.0f};
+    Uniform4fv(location, 1, written);
+    ASSERT_EQ(GetError(), GL_NO_ERROR);
+
+    const auto composite = DrawProgram();
+    ASSERT_NE(composite, nullptr);
+    const std::vector<float> value = ReadVec4(*composite, "u_shared");
+    ASSERT_EQ(value.size(), 4u) << "u_shared has no backing storage in the composite";
+    EXPECT_EQ(value, (std::vector<float>{0.25f, 0.5f, 0.75f, 1.0f}))
+        << "the fragment stage's untouched declaration overwrote the vertex stage's written value";
+
+    // The uniform only one stage declares is unaffected either way; it is here so a mirror that
+    // copied nothing at all would not pass this case by accident.
+    ActiveShaderProgram(pipeline, vs);
+    const GLint vsOnly = GetUniformLocation(vs, "u_vsOnly");
+    ASSERT_GE(vsOnly, 0);
+    const float other[4] = {1.0f, 2.0f, 3.0f, 4.0f};
+    Uniform4fv(vsOnly, 1, other);
+    const auto refreshed = DrawProgram();
+    EXPECT_EQ(ReadVec4(*refreshed, "u_vsOnly"), (std::vector<float>{1.0f, 2.0f, 3.0f, 4.0f}));
+    EXPECT_EQ(GetError(), GL_NO_ERROR);
+
+    BindProgramPipeline(0);
+    DeleteProgramPipelines(1, &pipeline);
+}
+
+// The tie the fix cannot make disappear: BOTH stages were written, and the composite still has
+// one slot. The documented rule is last WRITTEN-TO graphics stage wins, in ShaderStage enum
+// order - deterministic, and reachable only by a stage holding a real application value.
+TEST_F(ProgramPipelineCompositeTest, WhenBothStagesWereWrittenTheLastGraphicsStageWins) {
+    const GLuint vs = MakeSeparableProgram(GL_VERTEX_SHADER, kSharedUniformVs);
+    const GLuint fs = MakeSeparableProgram(GL_FRAGMENT_SHADER, kSharedUniformFs);
+
+    GLuint pipeline = 0;
+    GenProgramPipelines(1, &pipeline);
+    BindProgramPipeline(pipeline);
+    UseProgramStages(pipeline, GL_VERTEX_SHADER_BIT, vs);
+    UseProgramStages(pipeline, GL_FRAGMENT_SHADER_BIT, fs);
+
+    const float fromVs[4] = {1.0f, 1.0f, 1.0f, 1.0f};
+    const float fromFs[4] = {2.0f, 2.0f, 2.0f, 2.0f};
+    // Written in the order VS then FS...
+    ProgramUniform4fv(vs, GetUniformLocation(vs, "u_shared"), 1, fromVs);
+    ProgramUniform4fv(fs, GetUniformLocation(fs, "u_shared"), 1, fromFs);
+    ASSERT_EQ(GetError(), GL_NO_ERROR);
+    EXPECT_EQ(ReadVec4(*DrawProgram(), "u_shared"), (std::vector<float>{2.0f, 2.0f, 2.0f, 2.0f}));
+
+    // ...and in the order FS then VS. The answer is the same, because the rule is stage order
+    // and not write order - which is the honest statement of what the dirty set can support.
+    ProgramUniform4fv(fs, GetUniformLocation(fs, "u_shared"), 1, fromFs);
+    ProgramUniform4fv(vs, GetUniformLocation(vs, "u_shared"), 1, fromVs);
+    ASSERT_EQ(GetError(), GL_NO_ERROR);
+    EXPECT_EQ(ReadVec4(*DrawProgram(), "u_shared"), (std::vector<float>{2.0f, 2.0f, 2.0f, 2.0f}))
+        << "the both-written tie must be decided by stage order, deterministically";
+
+    BindProgramPipeline(0);
+    DeleteProgramPipelines(1, &pipeline);
+}
+
+// The both-written tie again, through the case that has no BYTES to move: the fragment stage
+// writes the value it was already holding.
+//
+// The refresh gate is built out of counters that move when bytes move (the UBO content
+// version, the backend state version), and both write funnels drop a value-identical write
+// before bumping either. So this write enlarges the write SET - it makes the fragment stage
+// the last written-to stage for `u_shared`, which is what decides the slot - while moving
+// nothing else. Without a generation on the set itself the gate never trips and the draw keeps
+// the vertex stage's value.
+TEST_F(ProgramPipelineCompositeTest, AValueIdenticalWriteStillTakesTheSlotForItsStage) {
+    const GLuint vs = MakeSeparableProgram(GL_VERTEX_SHADER, kSharedUniformVs);
+    const GLuint fs = MakeSeparableProgram(GL_FRAGMENT_SHADER, kSharedUniformFs);
+
+    GLuint pipeline = 0;
+    GenProgramPipelines(1, &pipeline);
+    BindProgramPipeline(pipeline);
+    UseProgramStages(pipeline, GL_VERTEX_SHADER_BIT, vs);
+    UseProgramStages(pipeline, GL_FRAGMENT_SHADER_BIT, fs);
+
+    const float fromVs[4] = {5.0f, 5.0f, 5.0f, 5.0f};
+    ProgramUniform4fv(vs, GetUniformLocation(vs, "u_shared"), 1, fromVs);
+    ASSERT_EQ(ReadVec4(*DrawProgram(), "u_shared"), (std::vector<float>{5.0f, 5.0f, 5.0f, 5.0f}));
+
+    // The fragment program's u_shared already reads all-zero, so this write changes not one
+    // byte of its shadow - and must still hand it the composite's slot.
+    const float zeros[4] = {0.0f, 0.0f, 0.0f, 0.0f};
+    ProgramUniform4fv(fs, GetUniformLocation(fs, "u_shared"), 1, zeros);
+    ASSERT_EQ(GetError(), GL_NO_ERROR);
+    EXPECT_EQ(ReadVec4(*DrawProgram(), "u_shared"), (std::vector<float>{0.0f, 0.0f, 0.0f, 0.0f}))
+        << "a write that moved no bytes never reached the refresh gate";
+
+    BindProgramPipeline(0);
+    DeleteProgramPipelines(1, &pipeline);
+}
+
+// glUseProgramStages here accepts a program that was never linked as separable (GL 4.6 core 7.4
+// says it should not, and MobileGL validates only LINK_STATUS). Such a program has recorded
+// none of its writes, because nothing ever armed its tracking latch - so the mirror has to fall
+// back to carrying everything rather than carrying nothing. Mirroring nothing would have been a
+// fresh regression on a shape that worked before the dirty set existed.
+TEST_F(ProgramPipelineCompositeTest, ANonSeparableStageProgramStillMirrorsItsUniforms) {
+    const char* vsSource = R"(#version 430 core
+uniform vec4 u_vsOnly;
+void main() { gl_Position = u_vsOnly; }
+)";
+    const GLuint shader = CreateShader(GL_VERTEX_SHADER);
+    ShaderSource(shader, 1, &vsSource, nullptr);
+    CompileShader(shader);
+    const GLuint vs = CreateProgram();
+    // Deliberately NO ProgramParameteri(GL_PROGRAM_SEPARABLE): this is the shape the latch
+    // cannot see coming.
+    AttachShader(vs, shader);
+    LinkProgram(vs);
+    GLint linked = GL_FALSE;
+    GetProgramiv(vs, GL_LINK_STATUS, &linked);
+    ASSERT_EQ(linked, GL_TRUE);
+
+    const GLuint fs = MakeSeparableProgram(GL_FRAGMENT_SHADER, kSharedUniformFs);
+
+    GLuint pipeline = 0;
+    GenProgramPipelines(1, &pipeline);
+    BindProgramPipeline(pipeline);
+    UseProgramStages(pipeline, GL_VERTEX_SHADER_BIT, vs);
+    UseProgramStages(pipeline, GL_FRAGMENT_SHADER_BIT, fs);
+    ASSERT_EQ(GetError(), GL_NO_ERROR);
+
+    const float written[4] = {3.0f, 1.0f, 4.0f, 1.0f};
+    ProgramUniform4fv(vs, GetUniformLocation(vs, "u_vsOnly"), 1, written);
+    ASSERT_EQ(GetError(), GL_NO_ERROR);
+
+    const auto composite = DrawProgram();
+    ASSERT_NE(composite, nullptr);
+    EXPECT_FALSE(MG_State::pGLContext->GetProgramObject(vs)->TracksUniformWrites())
+        << "this case is only meaningful while the stage program records nothing";
+    EXPECT_EQ(ReadVec4(*composite, "u_vsOnly"), (std::vector<float>{3.0f, 1.0f, 4.0f, 1.0f}))
+        << "a stage program with no write record must fall back to mirroring everything";
+
+    BindProgramPipeline(0);
+    DeleteProgramPipelines(1, &pipeline);
+}
+
+// glProgramUniform* addresses a program by NAME and needs neither a current program nor an
+// active shader program, so it is a write path that never touches the pipeline at all. It has
+// to record the write exactly like glUniform* does.
+TEST_F(ProgramPipelineCompositeTest, ProgramUniformOnAnUnboundStageProgramReachesTheComposite) {
+    const GLuint vs = MakeSeparableProgram(GL_VERTEX_SHADER, kSharedUniformVs);
+    const GLuint fs = MakeSeparableProgram(GL_FRAGMENT_SHADER, kSharedUniformFs);
+
+    GLuint pipeline = 0;
+    GenProgramPipelines(1, &pipeline);
+    UseProgramStages(pipeline, GL_VERTEX_SHADER_BIT, vs);
+    UseProgramStages(pipeline, GL_FRAGMENT_SHADER_BIT, fs);
+
+    // Deliberately BEFORE the bind, and with no glActiveShaderProgram anywhere: the write has
+    // to survive from here to a draw that has not been set up yet.
+    const float written[4] = {9.0f, 8.0f, 7.0f, 6.0f};
+    ProgramUniform4fv(vs, GetUniformLocation(vs, "u_vsOnly"), 1, written);
+    ASSERT_EQ(GetError(), GL_NO_ERROR);
+
+    BindProgramPipeline(pipeline);
+    EXPECT_EQ(ReadVec4(*DrawProgram(), "u_vsOnly"), (std::vector<float>{9.0f, 8.0f, 7.0f, 6.0f}));
+    EXPECT_EQ(GetError(), GL_NO_ERROR);
+
+    BindProgramPipeline(0);
+    DeleteProgramPipelines(1, &pipeline);
+}
+
+// Array uniforms are written at ELEMENT locations, so the record has to be per location and not
+// per name: a stage that wrote `u_arr[2]` and nothing else must carry element 2 across and
+// leave the rest to whichever stage owns them.
+TEST_F(ProgramPipelineCompositeTest, ArrayElementWritesMirrorPerElement) {
+    const GLuint vs = MakeSeparableProgram(GL_VERTEX_SHADER, kArrayUniformVs);
+    const GLuint fs = MakeSeparableProgram(GL_FRAGMENT_SHADER, kArrayUniformFs);
+
+    GLuint pipeline = 0;
+    GenProgramPipelines(1, &pipeline);
+    BindProgramPipeline(pipeline);
+    UseProgramStages(pipeline, GL_VERTEX_SHADER_BIT, vs);
+    UseProgramStages(pipeline, GL_FRAGMENT_SHADER_BIT, fs);
+
+    // Non-prefix on purpose: elements 1 and 3 from the vertex stage, element 2 from the fragment
+    // stage, element 0 from nobody. A per-name record would have carried whole arrays and let
+    // one stage's zeros take the other's elements.
+    const float one[4] = {11.0f, 11.0f, 11.0f, 11.0f};
+    const float three[4] = {33.0f, 33.0f, 33.0f, 33.0f};
+    const float two[4] = {22.0f, 22.0f, 22.0f, 22.0f};
+    ProgramUniform4fv(vs, GetUniformLocation(vs, "u_arr[1]"), 1, one);
+    ProgramUniform4fv(vs, GetUniformLocation(vs, "u_arr[3]"), 1, three);
+    ProgramUniform4fv(fs, GetUniformLocation(fs, "u_arr[2]"), 1, two);
+    ASSERT_EQ(GetError(), GL_NO_ERROR);
+
+    const auto composite = DrawProgram();
+    ASSERT_NE(composite, nullptr);
+    EXPECT_EQ(ReadVec4(*composite, "u_arr[0]"), (std::vector<float>{0.0f, 0.0f, 0.0f, 0.0f}));
+    EXPECT_EQ(ReadVec4(*composite, "u_arr[1]"), (std::vector<float>{11.0f, 11.0f, 11.0f, 11.0f}));
+    EXPECT_EQ(ReadVec4(*composite, "u_arr[2]"), (std::vector<float>{22.0f, 22.0f, 22.0f, 22.0f}));
+    EXPECT_EQ(ReadVec4(*composite, "u_arr[3]"), (std::vector<float>{33.0f, 33.0f, 33.0f, 33.0f}));
+    EXPECT_EQ(GetError(), GL_NO_ERROR);
+
+    // A multi-element glUniform*v run marks each location it actually reaches.
+    const float tail[8] = {44.0f, 44.0f, 44.0f, 44.0f, 55.0f, 55.0f, 55.0f, 55.0f};
+    ActiveShaderProgram(pipeline, fs);
+    Uniform4fv(GetUniformLocation(fs, "u_arr[2]"), 2, tail);
+    ASSERT_EQ(GetError(), GL_NO_ERROR);
+    const auto refreshed = DrawProgram();
+    EXPECT_EQ(ReadVec4(*refreshed, "u_arr[2]"), (std::vector<float>{44.0f, 44.0f, 44.0f, 44.0f}));
+    EXPECT_EQ(ReadVec4(*refreshed, "u_arr[3]"), (std::vector<float>{55.0f, 55.0f, 55.0f, 55.0f}))
+        << "the second element of a count=2 write was never recorded";
+
+    BindProgramPipeline(0);
+    DeleteProgramPipelines(1, &pipeline);
+}
+
+// Relinking resets a program's uniforms to their initial values (GL 4.6 core 7.6), so the record
+// of what was written has to be reset with them. If it survived, the composite built after the
+// relink would be handed values the stage program no longer holds.
+TEST_F(ProgramPipelineCompositeTest, RelinkingAStageProgramClearsWhatItHadWritten) {
+    const GLuint vs = MakeSeparableProgram(GL_VERTEX_SHADER, kSharedUniformVs);
+    const GLuint fs = MakeSeparableProgram(GL_FRAGMENT_SHADER, kSharedUniformFs);
+
+    GLuint pipeline = 0;
+    GenProgramPipelines(1, &pipeline);
+    BindProgramPipeline(pipeline);
+    UseProgramStages(pipeline, GL_VERTEX_SHADER_BIT, vs);
+    UseProgramStages(pipeline, GL_FRAGMENT_SHADER_BIT, fs);
+
+    const float written[4] = {5.0f, 6.0f, 7.0f, 8.0f};
+    ProgramUniform4fv(vs, GetUniformLocation(vs, "u_vsOnly"), 1, written);
+    ASSERT_EQ(ReadVec4(*DrawProgram(), "u_vsOnly"), (std::vector<float>{5.0f, 6.0f, 7.0f, 8.0f}));
+
+    LinkProgram(vs);
+    GLint linked = GL_FALSE;
+    GetProgramiv(vs, GL_LINK_STATUS, &linked);
+    ASSERT_EQ(linked, GL_TRUE);
+
+    const auto composite = DrawProgram();
+    ASSERT_NE(composite, nullptr);
+    EXPECT_EQ(ReadVec4(*composite, "u_vsOnly"), (std::vector<float>{0.0f, 0.0f, 0.0f, 0.0f}))
+        << "a relinked stage program carried its pre-relink value into the new composite";
+    EXPECT_EQ(GetError(), GL_NO_ERROR);
+
+    // ...and writing again after the relink is recorded afresh.
+    const float rewritten[4] = {1.5f, 2.5f, 3.5f, 4.5f};
+    ProgramUniform4fv(vs, GetUniformLocation(vs, "u_vsOnly"), 1, rewritten);
+    EXPECT_EQ(ReadVec4(*DrawProgram(), "u_vsOnly"), (std::vector<float>{1.5f, 2.5f, 3.5f, 4.5f}));
+    EXPECT_EQ(GetError(), GL_NO_ERROR);
+
+    BindProgramPipeline(0);
+    DeleteProgramPipelines(1, &pipeline);
+}
+
+// ---------------------------------------------------------------------------------------
+// Composite cache stability
+// ---------------------------------------------------------------------------------------
+
+// The SSO-conformance shape, and the reason the composite cache stopped being keyed on the
+// backend state version: pick a stage program, then per draw set a sampler unit and draw.
+// glUniform1i on a sampler bumps that version, so the signature changed on every iteration and
+// every single draw threw the composite away and relinked it - glslang, SPIR-V and spirv-opt,
+// synchronously, inside the draw - handing the backends a brand-new program identity each time.
+//
+// Asserted on the composite POINTER, which is the honest observable: it is the object both
+// backends key their per-program registries and pipeline memos on, so "same pointer" is exactly
+// the property that was lost.
+TEST_F(ProgramPipelineCompositeTest, ASamplerWritePerDrawDoesNotRebuildTheComposite) {
+    const GLuint vs = MakeSeparableProgram(GL_VERTEX_SHADER, kSamplerVs);
+    const GLuint fs = MakeSeparableProgram(GL_FRAGMENT_SHADER, kSamplerFs);
+
+    GLuint pipeline = 0;
+    GenProgramPipelines(1, &pipeline);
+    BindProgramPipeline(pipeline);
+    UseProgramStages(pipeline, GL_VERTEX_SHADER_BIT, vs);
+    UseProgramStages(pipeline, GL_FRAGMENT_SHADER_BIT, fs);
+    ActiveShaderProgram(pipeline, fs);
+    ASSERT_EQ(GetError(), GL_NO_ERROR);
+
+    const GLint sampler = GetUniformLocation(fs, "u_tex");
+    ASSERT_GE(sampler, 0);
+
+    const auto first = DrawProgram();
+    ASSERT_NE(first, nullptr);
+    const Uint64 firstLifetime = first->GetLifetimeId();
+    const Int compositeSampler = first->GetUniformLocation("u_tex");
+    ASSERT_GE(compositeSampler, 0);
+
+    for (GLint unit = 0; unit < 8; ++unit) {
+        Uniform1i(sampler, unit);
+        const auto composite = DrawProgram();
+        ASSERT_NE(composite, nullptr);
+        EXPECT_EQ(composite.get(), first.get())
+            << "the composite was rebuilt by a sampler-unit write at unit " << unit;
+        EXPECT_EQ(composite->GetLifetimeId(), firstLifetime) << "the composite's identity changed at unit " << unit;
+        // The value still has to ARRIVE - the whole point is that the mirror carries it now that
+        // the rebuild no longer does.
+        EXPECT_EQ(composite->GetUniformSamplerOrImageUnitIndex(static_cast<Uint>(compositeSampler)), unit)
+            << "the sampler unit did not reach the composite at unit " << unit;
+    }
+    EXPECT_EQ(GetError(), GL_NO_ERROR);
+
+    // A relink, by contrast, MUST replace it: that is the one thing the signature still tracks.
+    LinkProgram(fs);
+    GLint linked = GL_FALSE;
+    GetProgramiv(fs, GL_LINK_STATUS, &linked);
+    ASSERT_EQ(linked, GL_TRUE);
+    const auto afterRelink = DrawProgram();
+    ASSERT_NE(afterRelink, nullptr);
+    EXPECT_NE(afterRelink.get(), first.get()) << "a relinked stage program must rebuild the composite";
+
+    BindProgramPipeline(0);
+    DeleteProgramPipelines(1, &pipeline);
+}
+
+// The monolithic path must be untouched by any of this: a plain glUseProgram program is not
+// separable, records nothing, and is its own draw program.
+TEST_F(ProgramPipelineCompositeTest, AMonolithicProgramRecordsNothingAndIsItsOwnDrawProgram) {
+    const char* vsSource = R"(#version 430 core
+uniform vec4 u_shared;
+void main() { gl_Position = u_shared; }
+)";
+    const char* fsSource = R"(#version 430 core
+uniform vec4 u_shared;
+out vec4 o_color;
+void main() { o_color = u_shared; }
+)";
+    const GLuint vsShader = CreateShader(GL_VERTEX_SHADER);
+    ShaderSource(vsShader, 1, &vsSource, nullptr);
+    CompileShader(vsShader);
+    const GLuint fsShader = CreateShader(GL_FRAGMENT_SHADER);
+    ShaderSource(fsShader, 1, &fsSource, nullptr);
+    CompileShader(fsShader);
+
+    const GLuint program = CreateProgram();
+    AttachShader(program, vsShader);
+    AttachShader(program, fsShader);
+    LinkProgram(program);
+    GLint linked = GL_FALSE;
+    GetProgramiv(program, GL_LINK_STATUS, &linked);
+    ASSERT_EQ(linked, GL_TRUE);
+
+    UseProgram(program);
+    const float written[4] = {1.0f, 2.0f, 3.0f, 4.0f};
+    Uniform4fv(GetUniformLocation(program, "u_shared"), 1, written);
+    ASSERT_EQ(GetError(), GL_NO_ERROR);
+
+    const auto drawProgram = DrawProgram();
+    ASSERT_NE(drawProgram, nullptr);
+    EXPECT_EQ(drawProgram->GetExternalIndex(), program) << "a current program IS the draw program";
+    // Nothing was recorded, because nothing ever asked this program to be separable - which is
+    // what keeps the hot uniform path free of the bookkeeping.
+    EXPECT_FALSE(drawProgram->TracksUniformWrites());
+    EXPECT_TRUE(drawProgram->GetWrittenUniformIndices().empty());
+    EXPECT_EQ(ReadVec4(*drawProgram, "u_shared"), (std::vector<float>{1.0f, 2.0f, 3.0f, 4.0f}));
+
+    UseProgram(0);
+}