[Fix] (git) Added the ignored item .gradle

Initialize the desktop replay Request from the three iterationRP environment flags before loading MobileGL. Without this, the Request defaults caused the replay core to unset CI's exported flags, leaving all repairs disabled despite the workflow configuration.

.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)"
@@ -185,7 +209,7 @@ jobs:
       - name: Load trace cases
         id: trace-cases
         run: |
-          echo "android=$(python3 tools/trace_replay/trace_cases.py --ci --format github-apk)" >> "$GITHUB_OUTPUT"
+          echo "android=$(python3 tools/trace_replay/trace_cases.py --ci --format github-apk-matrix)" >> "$GITHUB_OUTPUT"
           echo "names=$(python3 tools/trace_replay/trace_cases.py --ci --format names)" >> "$GITHUB_OUTPUT"
 
   trace-fixtures:
@@ -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')"
@@ -281,13 +337,7 @@ jobs:
     strategy:
       fail-fast: false
       max-parallel: 4
-      matrix:
+      matrix: ${{ fromJSON(needs.trace-cases.outputs.android) }}
-        backend:
-          - name: DirectGLES
-            gpu: software
-          - name: DirectVulkan
-            gpu: lavapipe
-        case: ${{ fromJSON(needs.trace-cases.outputs.android) }}
     steps:
       - name: Set Swap Space
         uses: pierotofy/set-swap-space@v1.0
@@ -370,6 +420,9 @@ jobs:
           MOBILEGL_USE_ANGLE: ${{ matrix.backend.name == 'DirectGLES' && '1' || '0' }}
           MOBILEGL_TRACE_ANGLE_VARIANT: ${{ matrix.case.name == 'minecraft-1.21.4-fabric-iris-bliss-in-world' && '90a62123d794' || 'ec889e6ea831' }}
           MOBILEGL_MAGMA_R11G11B10F_FALLBACK: ${{ matrix.backend.name == 'DirectVulkan' && '1' || '0' }}
+          MOBILEGL_FIX_ITERATIONRP_SUBGROUP_SCRATCH: ${{ matrix.backend.name == 'DirectVulkan' && matrix.case.name == 'minecraft-1.21.4-fabric-iris-iterationrp-in-world' && '1' || '0' }}
+          MOBILEGL_DERIVE_NUM_SUBGROUPS: ${{ matrix.backend.name == 'DirectVulkan' && matrix.case.name == 'minecraft-1.21.4-fabric-iris-iterationrp-in-world' && '1' || '0' }}
+          MOBILEGL_ITERATIONRP_FIX_BARRIER: ${{ matrix.backend.name == 'DirectVulkan' && matrix.case.name == 'minecraft-1.21.4-fabric-iris-iterationrp-in-world' && '1' || '0' }}
         run: |
           apk_file="android-retrace-apks/MobileGL-plugin-trace-release-${GITHUB_SHA}.apk"
           test -f "${apk_file}"
@@ -379,6 +432,9 @@ jobs:
           if [ "${{ 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 +467,24 @@ 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 "${EMULATOR_LOG}" \
@@ -450,6 +524,13 @@ jobs:
           if [ -f "${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()
@@ -531,23 +612,42 @@ 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
+                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
-            fi
 
             echo "Deleting ${artifact_name} (${artifact_id})"
             gh api --method DELETE "repos/${GITHUB_REPOSITORY}/actions/artifacts/${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 +186,105 @@ 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"
+          MOBILEGL_FIX_ITERATIONRP_SUBGROUP_SCRATCH: "1"
+          MOBILEGL_DERIVE_NUM_SUBGROUPS: "1"
+          MOBILEGL_ITERATIONRP_FIX_BARRIER: "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 +328,18 @@ 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 +347,10 @@ jobs:
       - build-linux
       - test
       - benchmark
+      - integration
+    permissions:
+      actions: write
+      contents: read
     env:
       BUILD_DIR: build-retrace
       CCACHE_BASEDIR: ${{ github.workspace }}
@@ -240,12 +375,11 @@ 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 +445,21 @@ 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'
@@ -343,7 +492,9 @@ jobs:
     needs:
       - test
       - benchmark
+      - integration
     outputs:
+      matrix: ${{ steps.trace-cases.outputs.matrix }}
       names: ${{ steps.trace-cases.outputs.names }}
     steps:
       - name: Checkout repo
@@ -351,7 +502,9 @@ jobs:
 
       - name: Load trace cases
         id: trace-cases
-        run: echo "names=$(python3 tools/trace_replay/trace_cases.py --ci --format names)" >> "$GITHUB_OUTPUT"
+        run: |
+          echo "matrix=$(python3 tools/trace_replay/trace_cases.py --ci --format github-test-matrix)" >> "$GITHUB_OUTPUT"
+          echo "names=$(python3 tools/trace_replay/trace_cases.py --ci --format names)" >> "$GITHUB_OUTPUT"
 
   trace-fixtures:
     name: trace fixture (${{ matrix.case }})
@@ -366,9 +519,41 @@ 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')"
@@ -398,11 +583,7 @@ jobs:
     strategy:
       fail-fast: false
       max-parallel: 4
-      matrix:
+      matrix: ${{ fromJSON(needs.trace-cases.outputs.matrix) }}
-        backend:
-          - DirectGLES
-          - DirectVulkan
-        case: ${{ fromJSON(needs.trace-cases.outputs.names) }}
 
     steps:
       - name: Set Swap Space
@@ -456,9 +637,17 @@ 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 [ '${{ matrix.backend }}' = 'DirectVulkan' ]; then
             export MOBILEGL_MAGMA_R11G11B10F_FALLBACK=1
           fi
+          if [ '${{ matrix.backend }}' = 'DirectVulkan' ] \
+             && [ '${{ matrix.case }}' = 'minecraft-1.21.4-fabric-iris-iterationrp-in-world' ]; then
+            export MOBILEGL_FIX_ITERATIONRP_SUBGROUP_SCRATCH=1
+            export MOBILEGL_DERIVE_NUM_SUBGROUPS=1
+            export MOBILEGL_ITERATIONRP_FIX_BARRIER=1
+          fi
           # The blended depth-write quirk auto-enables only on Qualcomm, which no CI
           # runner has, so force it on for the OIT case it exists to fix. ForceOn
           # bypasses only the vendor gate, so this exercises the real strip on
@@ -470,6 +659,14 @@ jobs:
           fi
           ctest -V --no-tests=error -R '^MobileGLTraceReplay\.${{ 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

.gitignore

@@ -27,3 +27,4 @@ MobileGL/MG*/cmake-build*
 tools/trace_replay/work/
 __pycache__/
 *.py[cod]
+/.gradle

.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"]
+[submodule "include/ska"]
-	path = 3rdparty/libfork
+	path = include/ska
-	url = https://github.com/ConorWilliams/libfork.git
+	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)
@@ -107,6 +182,7 @@ set(ENABLE_SPVREMAPPER         OFF CACHE BOOL "Enable SPVRemapper" FORCE)
 set(ENABLE_OPT                 ON  CACHE BOOL "Enable SPIRV-Tools opt usage in glslang" FORCE)
 set(BUILD_EXTERNAL             ON  CACHE BOOL "Build external deps in External/" FORCE)
 set(ENABLE_GLSLANG_INSTALL     OFF CACHE BOOL "Install glslang targets" FORCE)
+set(SPIRV_SKIP_EXECUTABLES     ON  CACHE BOOL "Skip building SPIRV-Tools executables" FORCE)
 
 set(SPIRV_CROSS_C_API          ON  CACHE BOOL "Enable C API" FORCE)
 set(SPIRV_CROSS_ENABLE_GLSL    ON  CACHE BOOL "Enable GLSL backend" FORCE)
@@ -201,18 +277,32 @@ set(SOURCE_FILES
     MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/RenameBuiltinShadowingFunctionsPass.cpp
     MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/DecomposeWorkgroupVec3Pass.cpp
     MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/DecoratePositionInvariantPass.cpp
+    MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/DemoteFloat64Pass.cpp
     MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/LowerDrawParametersPass.cpp
     MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/PackDoubleVertexInputsPass.cpp
+    MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/FlattenXfbInterfaceBlocksPass.cpp
+    MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/SplitArrayVertexInputsPass.cpp
     MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/RebaseInstanceIndexPass.cpp
+    MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/ZeroBaseVertexPass.cpp
+    MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/DeriveNumSubgroupsPass.cpp
+    MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/FixIterationRPBarrierPass.cpp
+    MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/FixIterationRPSubgroupScratchPass.cpp
+    MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/EmulateSubgroupsPass.cpp
     MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/NormalizeRectCoordinatesPass.cpp
+    MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/Lower1DArrayImagesPass.cpp
+    MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/BakeImageFormatsPass.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
 
     MobileGL/MG_Util/SelfTest/DriverPost.cpp
+    MobileGL/MG_Util/SelfTest/DriverPostIterationRPWitness.cpp
 
     MobileGL/MG_Util/Texture/PixelStoreProcessor.cpp
     MobileGL/MG_Util/Texture/TextureFormatProcessor.cpp
@@ -296,6 +386,7 @@ 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
@@ -372,14 +463,7 @@ set(MOBILEGL_INCLUDE_DIR
         # Header-only submodule: no add_subdirectory, no link target. Only
         # 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
+        ${CMAKE_SOURCE_DIR}/3rdparty/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
@@ -591,3 +675,10 @@ if (NOT ANDROID)
         add_subdirectory(tools/trace_replay)
     endif()
 endif()
+
+# The integration binary is also useful as a standalone adb-shell executable.
+# Android cannot use the desktop-only MobileGL_s target, so its CMake module
+# links libMobileGL.so and creates an AImageReader-backed window instead.
+if (ANDROID AND MOBILEGL_BUILD_INTEGRATION_TEST)
+    add_subdirectory(MobileGL/MG_IntegrationTest)
+endif()

MobileGL/Config.h

@@ -66,23 +66,61 @@ 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).
     struct FeaturesTable {
         // MOBILEGL_DISABLE_TIMERQUERY: do not advertise or use GPU timer queries.
         Bool DisableTimerQuery = false;
+        // MOBILEGL_ENABLE_SPIRV_VALIDATION: validate generated and transformed SPIR-V.
+        // Disabled by default because validation is a diagnostics-only cost.
+        Bool EnableSpirvValidation = false;
         // MOBILEGL_USE_ANGLE: load ANGLE EGL/GLES libraries.
         Bool UseAngle = false;
 #if defined(MOBILEGL_TRACE_ANGLE_VARIANTS)
         // MOBILEGL_TRACE_ANGLE_VARIANT: signed trace-APK ANGLE build short hash.
         String TraceAngleVariant;
 #endif
-        // MOBILEGL_DISABLE_SUBGROUP: force-disable Vulkan shader subgroup support.
+        // MOBILEGL_DISABLE_SUBGROUP: force-disable Vulkan shader subgroup support,
+        // including the opt-in emulated compute path below.
         Bool DisableSubgroup = false;
+        // MOBILEGL_MAGMA_EMULATE_SUBGROUP: implement GL_KHR_shader_subgroup's compute
+        // stage on a 32-lane VIRTUAL subgroup lowered to workgroup-shared memory
+        // (ShaderTranspiler::EmulateSubgroupsPass). Strictly a last resort: it only ever
+        // engages when this flag is set AND the device has no native subgroup support at
+        // all - a device with real subgroup operations always uses them natively,
+        // whatever their width (the known iterationRP defect is patched by
+        // FixIterationRPSubgroupScratch below instead). Off by default.
+        Bool MagmaEmulateSubgroup = false;
+        // MOBILEGL_FIX_ITERATIONRP_SUBGROUP_SCRATCH: patch iterationRP's own bug - the
+        // pack declares `shared vec2 prefixSumCache[32]` for a 512-invocation exposure
+        // reduction and indexes it by gl_SubgroupID, so any device with sub-16-lane
+        // subgroups (8-lane lavapipe -> 64 subgroups) writes shared memory out of
+        // bounds. The pass grows that one array to what the device's topology needs and
+        // touches nothing else; it only rewrites modules positively matching the pack's
+        // reduction fingerprint (ShaderTranspiler::FixIterationRPSubgroupScratchPass),
+        // so every other shader passes through byte-identical - as does iterationRP
+        // itself on >= 16-lane devices. Auto is ON; ForceOff replays the pack's bug
+        // verbatim.
+        QuirkOverride FixIterationRPSubgroupScratch = QuirkOverride::Auto;
+        // MOBILEGL_ITERATIONRP_FIX_BARRIER: repair Program 203's missing workgroup
+        // rendezvous between its two reductions over prefixSumCache. Off by default and
+        // fingerprint-gated by FixIterationRPBarrierPass when enabled.
+        Bool IterationRPFixBarrier = false;
+        // MOBILEGL_DERIVE_NUM_SUBGROUPS: replace compute gl_NumSubgroups loads with
+        // ceil(workgroup invocations / gl_SubgroupSize) on the NATIVE subgroup path
+        // (ShaderTranspiler::DeriveNumSubgroupsPass). Auto is ON: GL requires
+        // gl_SubgroupID < gl_NumSubgroups, Adreno's builtin reports 1 while the same
+        // dispatch emits IDs 0..7, and the derived value is the one Vulkan guarantees
+        // whenever the pipeline can request REQUIRE_FULL_SUBGROUPS (which the renderer
+        // does whenever local_size_x is a multiple of the native width). ForceOff returns
+        // to the raw driver builtin.
+        QuirkOverride DeriveNumSubgroups = QuirkOverride::Auto;
+        // MOBILEGL_ADVERTISE_FP64: add GL_ARB_gpu_shader_fp64 to the advertised extension
+        // string. `double` in a shader always WORKS - it is narrowed to 32 bits before any
+        // module reaches a backend (ShaderTranspiler::DemoteFloat64Pass) - but the extension
+        // promises 64-bit precision, and that is the one thing the narrowing cannot deliver.
+        // Off by default so an application that checks the string before using doubles keeps
+        // its float path; on for measuring what the conformance suite makes of the demoted
+        // precision. See the DemoteFloat64Pass header and the "fp64" POST row.
+        Bool AdvertiseFp64 = false;
         // MOBILEGL_MAGMA_R11G11B10F_FALLBACK: use fallback format for R11G11B10F on Vulkan.
         Bool MagmaR11G11B10FFallback = false;
         // MOBILEGL_MAGMA_FRAMESINFLIGHT: requested Magma frames in flight, defaulting to 3.
@@ -90,6 +128,13 @@ 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
@@ -103,16 +148,19 @@ namespace MobileGL::MG_Config {
         // per-draw glBufferSubData path instead of the persistent-mapped ring allocator
         // (negative control / driver-bug escape hatch).
         Bool DisableUboRing = false;
+        // MOBILEGL_ESPRYT_FORCE_DS_READBACK_EMULATION: make DirectGLES skip the native ES
+        // depth/stencil reads and always go through the shader-sampling emulation. Core GL
+        // ES has no depth or stencil readback, but some drivers accept it anyway (Mesa does,
+        // Adreno does not), which means the emulation is dead code on exactly the stack the
+        // headless suite runs on. This forces it live so the scenarios and the CTS can
+        // exercise the path, and gives the device an A/B lever over the same choice.
+        Bool EsprytForceDepthStencilReadbackEmulation = false;
         // MOBILEGL_RELAXED_SEMANTICS: relax strict core-profile rules (e.g. VAO-0 draws,
         // texture-name reuse after delete) even on contexts that explicitly requested a core
         // profile. Without it, relaxed semantics still apply to every context that did not
         // explicitly request a core profile via EGL_CONTEXT_OPENGL_PROFILE_MASK / a >=3.1
         // version request.
         Bool RelaxedSemantics = false;
-        // MOBILEGL_QUIRK_SUBGROUP_PREFIX_SCAN: overrides the shader-source quirk that
-        // rewrites the recognized workgroup prefix-scan template on Qualcomm devices with
-        // subgroups wider than 32 lanes (see ShaderSourceProcessor's quirk registry).
-        QuirkOverride SubgroupPrefixScanQuirk = QuirkOverride::Auto;
         // MOBILEGL_MAGMA_DISABLE_BLENDED_DEPTH_WRITE: overrides the DirectVulkan quirk that
         // strips depth writes from accumulation-blended pipelines (MIN/MAX or additive
         // ONE+ONE - the multi-pass depth-equality signature) on drivers without
@@ -142,6 +190,19 @@ 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

@@ -162,20 +162,29 @@ namespace MobileGL::MG_ConfigLoader {
     inline void InitFeatures() {
         auto& features = MG_Config::Features;
         features.DisableTimerQuery = QueryEnvFlag("MOBILEGL_DISABLE_TIMERQUERY");
+        features.EnableSpirvValidation = QueryEnvFlag("MOBILEGL_ENABLE_SPIRV_VALIDATION");
         features.UseAngle = QueryEnvFlag("MOBILEGL_USE_ANGLE");
 #if defined(MOBILEGL_TRACE_ANGLE_VARIANTS)
         QueryEnvVariable("MOBILEGL_TRACE_ANGLE_VARIANT", features.TraceAngleVariant, "");
 #endif
         features.DisableSubgroup = QueryEnvFlag("MOBILEGL_DISABLE_SUBGROUP");
+        features.MagmaEmulateSubgroup = QueryEnvFlag("MOBILEGL_MAGMA_EMULATE_SUBGROUP");
+        features.FixIterationRPSubgroupScratch =
+            QueryEnvQuirkOverride("MOBILEGL_FIX_ITERATIONRP_SUBGROUP_SCRATCH");
+        features.IterationRPFixBarrier = QueryEnvFlag("MOBILEGL_ITERATIONRP_FIX_BARRIER");
+        features.DeriveNumSubgroups = QueryEnvQuirkOverride("MOBILEGL_DERIVE_NUM_SUBGROUPS");
+        features.AdvertiseFp64 = QueryEnvFlag("MOBILEGL_ADVERTISE_FP64");
         features.MagmaR11G11B10FFallback = QueryEnvFlag("MOBILEGL_MAGMA_R11G11B10F_FALLBACK");
         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");
+        features.EsprytForceDepthStencilReadbackEmulation =
+            QueryEnvFlag("MOBILEGL_ESPRYT_FORCE_DS_READBACK_EMULATION");
         features.RelaxedSemantics = QueryEnvFlag("MOBILEGL_RELAXED_SEMANTICS");
-        features.SubgroupPrefixScanQuirk = QueryEnvQuirkOverride("MOBILEGL_QUIRK_SUBGROUP_PREFIX_SCAN");
         features.MagmaDisableBlendedDepthWriteQuirk =
             QueryEnvQuirkOverride("MOBILEGL_MAGMA_DISABLE_BLENDED_DEPTH_WRITE");
         features.DisableRobustBufferAccess = QueryEnvFlag("MOBILEGL_DISABLE_ROBUST_BUFFER_ACCESS");
@@ -183,6 +192,8 @@ 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__
+// No __ANDROID_API__ pin here on purpose. The effective API level is owned by
-#undef __ANDROID_API__
+// the build system (gradle minSdk 26 -> -DANDROID_PLATFORM=android-26, enforced
-#define __ANDROID_API__ 26 // force Android API level to 26 for compatibility
+// by the configure-time guard in CMakeLists.txt), not by a macro.
-#endif
+//
+// 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 +47,23 @@
 #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.
+//
+// Severity order, ascending: DEBUG < INFO < WARN < ERROR < FATAL. MOBILEGL_LOG_ACTIVE_LEVEL
+// names the lowest severity compiled in, so the production default INFO keeps I/W/E/F and
+// drops only D. Any edit here must be mirrored in Log.h.
+#ifndef MOBILEGL_LOG_LEVEL_DEBUG
+#define MOBILEGL_LOG_LEVEL_DEBUG 0
+#define MOBILEGL_LOG_LEVEL_INFO 1
+#define MOBILEGL_LOG_LEVEL_WARN 2
+#define MOBILEGL_LOG_LEVEL_ERROR 3
+#define MOBILEGL_LOG_LEVEL_FATAL 4
+#endif
+
 #ifndef MOBILEGL_LOG_ACTIVE_LEVEL
 #define MOBILEGL_LOG_ACTIVE_LEVEL MOBILEGL_LOG_LEVEL_INFO
 #endif
@@ -68,6 +95,12 @@
 #endif
 
 // =============================== Utils ================================ //
+// Asserts are live in exactly the builds where MGLOG_D is live, i.e. DEBUG builds only;
+// an INFO build (the production default) compiles them out. DEBUG is the lowest severity
+// in the ordering above, so "ACTIVE <= DEBUG" is true only for ACTIVE == DEBUG - the same
+// gate MGLOG_D uses in Log.h. That equivalence is what makes this gate survive the
+// 2026-08-13 renumbering unchanged; the contract is and stays
+// "INFO builds: asserts OFF; DEBUG builds: asserts ON".
 #if MOBILEGL_LOG_ACTIVE_LEVEL <= MOBILEGL_LOG_LEVEL_DEBUG
     #define MOBILEGL_ASSERT(condition, ...)                                                                                \
         do {                                                                                                               \

MobileGL/Includes.h

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

MobileGL/Init.cpp

@@ -15,6 +15,7 @@
 #include <MG_Impl/GLImpl/Texture/ProxyTexture.h>
 #include <MG_Impl/GLImpl/Framebuffer/GL_Framebuffer.h>
 #include <MG_Impl/GLImpl/Sync/GL_Sync.h>
+#include <MG_Impl/GLImpl/Query/GL_Query.h>
 #include <MG_Util/Async/ShaderCompilePool.h>
 #include <MG_Util/ShaderTranspiler/ShaderCompiler.h>
 
@@ -51,6 +52,11 @@ namespace MobileGL {
             // before a re-initialized library could pair them with the wrong
             // backend's DeleteSync).
             MG_Impl::GLImpl::DestroyAllSyncObjects();
+            // Queries die with their contexts for the same reason, and their registry
+            // is the same shape of process-global map: drain it here too, while the
+            // function table can still pair each backend handle with the backend that
+            // minted it.
+            MG_Impl::GLImpl::DestroyAllQueryObjects();
             MG_Backend::pActiveBackendObject.reset();
             MG_State::pGLContext.reset();
             MG_State::pEGLContext.reset();

MobileGL/MG_Backend/DirectGLES/BackendObject_DirectGLES.cpp

@@ -712,9 +712,9 @@ namespace MobileGL::MG_Backend::DirectGLES {
                     {
                         .TargetGLVersion = {4, 0, 0},   // GL target version
                         .TargetGLSLVersion = {4, 6, 0}, // Target Shading Language Version
-                        // Baseline advertisement (no timer queries / anisotropy yet); reconciled
+                        // Baseline advertisement (no runtime capabilities yet); reconciled once
-                        // once the ES capabilities exist, see UpdateAdvertisedCapabilityExtensions.
+                        // the ES capabilities exist, see UpdateAdvertisedCapabilityExtensions.
-                        .Extensions = BuildAdvertisedExtensions(false, false),
+                        .Extensions = BuildAdvertisedExtensions(false, false, false, false),
                         .IsCompatibilityProfile = false // Is Compatibility Profile
                     },
                 .StaticBackendCapability = {.AllowVSOnlyPrograms = false} // Backend Capability
@@ -734,9 +734,11 @@ namespace MobileGL::MG_Backend::DirectGLES {
         // thread can only observe the extension string after the
         // advertisement for its context has settled; rebuilding the whole
         // list keeps the re-run after a context recreation idempotent.
-        void UpdateAdvertisedCapabilityExtensions(Bool anisotropicFilteringSupported) {
+        void UpdateAdvertisedCapabilityExtensions(const MG_External::GLESCapabilities& capabilities) {
-            MutableRendererInfo().RendererGLInfo.Extensions =
+            MutableRendererInfo().RendererGLInfo.Extensions = BuildAdvertisedExtensions(
-                BuildAdvertisedExtensions(AreTimerQueriesSupported(), anisotropicFilteringSupported);
+                AreTimerQueriesSupported(), capabilities.SupportsTextureFilterAnisotropy,
+                capabilities.SupportsDrawIndirect,
+                capabilities.SupportsDrawIndirect && capabilities.SupportsBaseInstance);
         }
     } // namespace
 
@@ -779,11 +781,11 @@ namespace MobileGL::MG_Backend::DirectGLES {
             return false;
         }
         DirectGLES::SetGLESCapabilities(m_GLESCapabilities);
-        // Now that g_GLESCapabilities knows about GL_EXT_disjoint_timer_query and
+        // Now that g_GLESCapabilities knows the host extensions, entry points, and ES version,
-        // GL_EXT_texture_filter_anisotropic, reconcile the advertisement (see the comment on
+        // reconcile every runtime-gated advertisement (see the comment on
-        // UpdateAdvertisedCapabilityExtensions for why it cannot happen when the extension
+        // UpdateAdvertisedCapabilityExtensions for why this cannot happen when the list is first
-        // list is first built).
+        // built).
-        UpdateAdvertisedCapabilityExtensions(m_GLESCapabilities.SupportsTextureFilterAnisotropy);
+        UpdateAdvertisedCapabilityExtensions(m_GLESCapabilities);
         UpdateDynamicBackendParameters();
         PopulateFormatCapabilities(m_GLESFunctions, m_GLESCapabilities, MutableFormatCapabilities());
         PrintFormatCapabilities(GetFormatCapabilities());
@@ -924,11 +926,13 @@ namespace MobileGL::MG_Backend::DirectGLES {
         return MutableRendererInfo();
     }
 
-    Vector<GLExtension> BuildAdvertisedExtensions(Bool timerQueriesSupported, Bool anisotropicFilteringSupported) {
+    Vector<GLExtension> BuildAdvertisedExtensions(Bool timerQueriesSupported, Bool anisotropicFilteringSupported,
+                                                  Bool drawIndirectSupported,
+                                                  Bool nonZeroIndirectBaseInstanceSupported) {
         Vector<GLExtension> extensions = {
             V_OpenGL30, V_OpenGL31, V_OpenGL32, V_OpenGL33, V_OpenGL40, E_GL_ARB_draw_buffers_blend,
             E_GL_ARB_compute_shader, E_GL_ARB_shader_storage_buffer_object, E_GL_ARB_shader_image_load_store,
-            E_GL_ARB_program_interface_query, E_GL_ARB_framebuffer_object, E_GL_EXT_framebuffer_object,
+            E_GL_ARB_clear_buffer_object, E_GL_ARB_program_interface_query, E_GL_ARB_framebuffer_object, E_GL_EXT_framebuffer_object,
             E_GL_ARB_depth_texture, E_GL_ARB_buffer_storage, E_GL_ARB_texture_storage,
             E_GL_ARB_texture_storage_multisample, E_GL_ARB_clear_texture, E_GL_ARB_direct_state_access,
             E_GL_ARB_multi_draw_indirect, E_GL_ARB_indirect_parameters, E_GL_ARB_shader_draw_parameters,
@@ -940,10 +944,34 @@ namespace MobileGL::MG_Backend::DirectGLES {
             // picks a whole different shader for draw_buffers without
             // explicit_attrib_location. DirectVulkan advertises both.
             E_GL_ARB_explicit_attrib_location, E_GL_ARB_texture_multisample, E_GL_ARB_shader_image_size,
+            // Core since GL 3.1 and implemented for every version advertised here. The string
+            // matters because applications gate the ENTRY POINTS on it rather than on the
+            // version: a caller that finds the extension missing never resolves
+            // glGetUniformBlockIndex / glUniformBlockBinding, and one that then uses uniform
+            // blocks anyway calls through a null pointer.
+            E_GL_ARB_uniform_buffer_object,
+            // Sampling the stencil aspect through DEPTH_STENCIL_TEXTURE_MODE. Core from 4.3,
+            // so on a 4.0 context the string is the only way to reach it. The host ES driver
+            // has had the same texture parameter since ES 3.1, which every device MobileGL
+            // runs on provides.
+            E_GL_ARB_stencil_texturing,
             // Advertised with GL_NUM_PROGRAM_BINARY_FORMATS = 0, which the
             // extension explicitly permits. It is also the only thing that
             // exposes glProgramParameteri before GL 4.1.
             E_GL_ARB_get_program_binary};
+        // Minecraft 26.3 checks this prerequisite before it even considers
+        // GL_ARB_multi_draw_indirect. ES 3.1 supplies both single-draw entry points; the loader
+        // folds the version and pointer checks into SupportsDrawIndirect.
+        if (drawIndirectSupported) {
+            extensions.push_back(E_GL_ARB_draw_indirect);
+        }
+        // ARB_base_instance also defines the last word of an indirect command. Direct calls are
+        // emulated on every Espryt device, but without host GL_EXT_base_instance a native indirect
+        // draw cannot shift divisor attributes by a GPU-authored non-zero value, so do not promise
+        // that incomplete case.
+        if (drawIndirectSupported && nonZeroIndirectBaseInstanceSupported) {
+            extensions.push_back(E_GL_ARB_base_instance);
+        }
         // GL_KHR_parallel_shader_compile is MobileGL's own capability, not the host ES
         // driver's: the compiler threads are MobileGL's, and glCompileShader/glLinkProgram
         // are serviced entirely inside the frontend. Whether the device driver advertises
@@ -960,6 +988,15 @@ namespace MobileGL::MG_Backend::DirectGLES {
         if (MG_Util::Async::AsyncShaderCompileEnabled()) {
             extensions.push_back(E_GL_KHR_parallel_shader_compile);
         }
+        // GL_ARB_gpu_shader_fp64 is opt-in (MOBILEGL_ADVERTISE_FP64). Every `double` in a
+        // shader compiles and runs already - it is narrowed to 32 bits before the module
+        // reaches this backend - so an application that simply uses doubles needs nothing
+        // advertised. What the extension additionally promises is 64-bit PRECISION, which no
+        // mobile GPU has and the narrowing cannot fake, so advertising it by default would
+        // make an application that checks the string take a path MobileGL cannot honour.
+        if (MG_Config::Features.AdvertiseFp64) {
+            extensions.push_back(E_GL_ARB_gpu_shader_fp64);
+        }
         // Only advertised when the device driver actually has usable timer queries
         // (GL_EXT_disjoint_timer_query plus its entry points) and the
         // MOBILEGL_DISABLE_TIMERQUERY escape hatch is off.
@@ -1002,6 +1039,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
             funcsTable.GL.MultiDrawElementsIndirect = MultiDrawElementsIndirect;
             funcsTable.GL.MultiDrawElementsIndirectCount = MultiDrawElementsIndirectCount;
             funcsTable.GL.MultiDrawArraysIndirect = MultiDrawArraysIndirect;
+            funcsTable.GL.MultiDrawArraysIndirectCount = MultiDrawArraysIndirectCount;
             funcsTable.GL.DrawRangeElementsBaseVertex = DrawRangeElementsBaseVertex;
             funcsTable.GL.DrawRangeElements = DrawRangeElements;
             funcsTable.GL.DrawElementsInstancedBaseVertexBaseInstance = DrawElementsInstancedBaseVertexBaseInstance;
@@ -1151,6 +1189,16 @@ 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/BackendObject_DirectGLES.h

@@ -67,9 +67,12 @@ namespace MobileGL::MG_Backend::DirectGLES {
     const RendererInfo& GetRendererIdentity();
 
     // The full OpenGL extension list Espryt advertises (glGetString(GL_EXTENSIONS))
-    // for a device whose timer queries / anisotropic filtering are (or are not) usable.
+    // for a device whose timer queries / anisotropic filtering / native indirect draws /
+    // non-zero indirect baseInstance semantics are (or are not) usable.
     // The MOBILEGL_DISABLE_TIMERQUERY escape hatch is applied inside.
-    Vector<GLExtension> BuildAdvertisedExtensions(Bool timerQueriesSupported, Bool anisotropicFilteringSupported);
+    Vector<GLExtension> BuildAdvertisedExtensions(Bool timerQueriesSupported, Bool anisotropicFilteringSupported,
+                                                  Bool drawIndirectSupported,
+                                                  Bool nonZeroIndirectBaseInstanceSupported);
 
     // Format: <OpenGL ES Renderer>, OpenGL ES <Major>.<Minor> — the exact string an
     // initialized backend returns from GetBackendAPIVersionString (and that ends up

MobileGL/MG_Backend/DirectGLES/DirectGLES.h

@@ -40,6 +40,8 @@ namespace MobileGL::MG_Backend::DirectGLES {
     void MultiDrawElementsIndirectCount(GLenum mode, GLenum type, const void* indirect, GLintptr drawcount,
                                         GLsizei maxdrawcount, GLsizei stride);
     void MultiDrawArraysIndirect(GLenum mode, const void* indirect, GLsizei drawcount, GLsizei stride);
+    void MultiDrawArraysIndirectCount(GLenum mode, const void* indirect, GLintptr drawcount, GLsizei maxdrawcount,
+                                      GLsizei stride);
     void DrawRangeElementsBaseVertex(GLenum mode, GLuint start, GLuint end, GLsizei count, GLenum type,
                                      const void* indices, GLint basevertex);
     void DrawRangeElements(GLenum mode, GLuint start, GLuint end, GLsizei count, GLenum type, const void* indices);
@@ -117,6 +119,24 @@ 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.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.
@@ -74,14 +82,26 @@ namespace MobileGL::MG_Backend::DirectGLES {
     // GLES core supports only GL_PRIMITIVE_RESTART_FIXED_INDEX. Throws when the app enabled
     // the arbitrary GL_PRIMITIVE_RESTART with a non-fixed index for this index type.
     void CheckPrimitiveRestartSupported(GLenum indexType);
-    // Feed the current program's gl_BaseInstance / gl_DrawID emulation uniforms. Both are
+    // Feed the current program's gl_BaseInstance / gl_DrawID / gl_BaseVertex emulation
-    // no-ops when the program does not read the corresponding builtin.
+    // uniforms. All are no-ops when the program does not read the corresponding builtin.
     void SetCurrentBaseInstance(Uint32 baseInstance);
     void SetCurrentDrawID(Uint32 drawId);
+    // GL's gl_BaseVertex is the base-vertex parameter of an indexed draw and zero for every
+    // command that has none - including all the DrawArrays forms - so every draw path that
+    // does not carry one must leave this at zero rather than inherit the last draw's value.
+    void SetCurrentBaseVertex(Int32 baseVertex);
     // True when the current program actually reads gl_DrawID, i.e. when a batched
     // (single driver call) multi-draw tier would have to feed it one value for the whole
     // batch and would therefore be wrong.
     Bool CurrentProgramReadsDrawID();
+    // Same question for gl_BaseVertex: a batched multi-draw tier cannot give each sub-draw
+    // its own base vertex through a uniform either.
+    Bool CurrentProgramReadsBaseVertex();
+    // Both of the above, conservatively, for a caller that must decide BEFORE PrepareForDraw
+    // has synced the program - where "does not read it" is indistinguishable from "cannot be
+    // asked yet". Answers true whenever the backend twin is missing or predates the current
+    // link.
+    Bool CurrentProgramMayNeedPerSubDrawBuiltins(Bool batchCarriesBaseVertices);
 
     template <typename StateObject, typename BackendObject>
     class StateBackendObjectRegistry {
@@ -121,6 +141,11 @@ 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()) {
@@ -261,6 +286,14 @@ namespace MobileGL::MG_Backend::DirectGLES {
             // context loss.
             Bool persistentMapped = false;
             void* persistentPtr = nullptr;
+            // The GL store behind `id` was created with glBufferStorageEXT and is
+            // therefore IMMUTABLE - glBufferData cannot respecify it and it must never be
+            // recycled through the size-keyed buffer pool. Tracked separately from
+            // persistentMapped because the two come apart: a glMapBufferRange that fails
+            // after its glBufferStorageEXT succeeded leaves immutable storage behind with
+            // no map, and a respecification then has to retire the id rather than hand it
+            // to glBufferData, which the driver would silently refuse.
+            Bool immutableStorage = false;
         };
 
         // Registered as the frontend's BufferBackendOps at backend init and on
@@ -313,6 +346,14 @@ namespace MobileGL::MG_Backend::DirectGLES {
         // client-attribute staging buffers): scrub every buffer-binding shadow that
         // could false-skip when the name is recycled.
         void NoteBufferIdDeleted(Uint id);
+        // Bumped whenever a live GLESBufferResource's driver id is retired and re-minted
+        // while its frontend buffer stays alive (persistent-map adoption, immutable-store
+        // retire). The VAO twins' baked glVertexAttribPointer / element-array bindings
+        // key on FRONTEND versions, which a backend-side re-mint does not move - without
+        // this generation the driver VAO would keep fetching through the deleted id (or
+        // its retained store) forever. Compared and stamped by
+        // BackendVertexArrayObject::SyncToBackend.
+        extern Uint64 g_bufferBackendIdGeneration;
         // Redundant-bind cache for INDEXED buffer bindings (glBindBufferBase/Range on
         // GL_UNIFORM_BUFFER / GL_SHADER_STORAGE_BUFFER): skips the GL call when the
         // (id, range) already at that index matches, like the array-buffer/texture/
@@ -432,6 +473,11 @@ namespace MobileGL::MG_Backend::DirectGLES {
             Array<Uint, MG_State::GLState::VertexArrayObject::MAX_VERTEX_ATTRIBS> m_clientAttributeBufferIds;
             Bool m_isInitialized = false;
             Uint16 m_syncedIndexBufferVersion = 0;
+            // Identity of the buffer the version above was stamped against. Raw and never
+            // dereferenced: the slot version is a wrapping Uint16 (see the ResolvedDrawBuffers
+            // IBO memo and the packed_pixels postmortem at BindCurrentFBO), so the version
+            // alone would read a wrapped-back count with a different buffer bound as clean.
+            const MG_State::GLState::BufferObject* m_syncedIndexBufferObject = nullptr;
             // Aggregate gate over the per-attribute walk below: the frontend bumps its config
             // version on every per-attribute version bump (the three Bump*Version functions are
             // its only writers), so an unchanged config version proves every per-attribute
@@ -441,6 +487,17 @@ namespace MobileGL::MG_Backend::DirectGLES {
             Uint32 m_syncedConfigVersion = 0;
             Array<MG_State::GLState::VertexAttributeVersion, MG_State::GLState::VertexArrayObject::MAX_VERTEX_ATTRIBS>
                 m_syncedAttributeVersions;
+            // Byte shift currently baked into the instanced arrays' offsets by the baseInstance
+            // emulation (see SetPendingFetchBaseInstance). It is draw state, not VAO state, so it
+            // is deliberately NOT covered by the config version: the frontend never bumps for it.
+            // Kept here because it describes what was last EMITTED, which is what the next sync
+            // has to correct.
+            Uint32 m_syncedFetchBaseInstance = 0;
+            // BufferImpl::g_bufferBackendIdGeneration as of this twin's last emit. A
+            // mismatch means some live buffer's driver id was re-minted since; the ids
+            // baked into the driver VAO's attribute/element bindings may be dead even
+            // though every frontend version matches, so the next sync re-emits them all.
+            Uint64 m_syncedBufferIdGeneration = 0;
         };
 
         extern StateBackendObjectRegistry<MG_State::GLState::VertexArrayObject, BackendVertexArrayObject>
@@ -454,6 +511,23 @@ namespace MobileGL::MG_Backend::DirectGLES {
         void InvalidateVAOBindingCache();
         // ES resets the binding to 0 when the currently bound VAO is deleted.
         void NoteVAOIdDeleted(Uint id);
+
+        // baseInstance emulation for drivers without GL_EXT_base_instance. GL fetches an
+        // instanced array at element "floor(instance / divisor) + baseInstance", and ES has no
+        // way to say the "+ baseInstance" part - so it is folded into the attribute's own byte
+        // offset (baseInstance * stride) for every divisor'd array, which is exactly equivalent.
+        // Must be set BEFORE PrepareForDraw so the VAO sync sees it, and cleared after the draw
+        // so the next one refetches from element 0; ScopedFetchBaseInstance does both.
+        void SetPendingFetchBaseInstance(Uint32 baseInstance);
+        Uint32 GetPendingFetchBaseInstance();
+
+        class ScopedFetchBaseInstance {
+        public:
+            explicit ScopedFetchBaseInstance(Uint32 baseInstance) { SetPendingFetchBaseInstance(baseInstance); }
+            ~ScopedFetchBaseInstance() { SetPendingFetchBaseInstance(0); }
+            ScopedFetchBaseInstance(const ScopedFetchBaseInstance&) = delete;
+            ScopedFetchBaseInstance& operator=(const ScopedFetchBaseInstance&) = delete;
+        };
     } // namespace VertexArrayImpl
 
     namespace TextureImpl {
@@ -613,6 +687,11 @@ 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
@@ -638,8 +717,18 @@ namespace MobileGL::MG_Backend::DirectGLES {
             FloatVec4 m_cacheBorderColor = {0.0f, 0.0f, 0.0f, 0.0f};
             Vec4<TextureSwizzleParam> m_cacheSwizzleParams = {TextureSwizzleParam::Red, TextureSwizzleParam::Green,
                                                               TextureSwizzleParam::Blue, TextureSwizzleParam::Alpha};
+            // GL_DEPTH_STENCIL_TEXTURE_MODE. GL_DEPTH_COMPONENT is the GL and ES default, so a
+            // texture that never asks for the stencil aspect never emits the call. The
+            // depth/stencil readback and replicate-blit emulations also write this parameter
+            // raw, but only ever on their own scratch textures (never on an application
+            // texture), so they cannot desynchronise this cache.
+            GLenum m_cacheDepthStencilTextureMode = GL_DEPTH_COMPONENT;
             Uint16 m_syncedSamplerVersion = 0;
             Uint16 m_syncedTextureParamsVersion = 0;
+            // Set when the driver texture underneath was regenerated and has therefore lost every
+            // parameter already pushed onto it: the params-version early-out has to be overridden
+            // once, or an unchanged version would skip the re-push forever.
+            Bool m_forceTextureParamsResync = false;
         };
 
         void ActivateTextureUnit(Uint unit);
@@ -657,15 +746,20 @@ 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 +774,7 @@ 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
@@ -722,6 +817,11 @@ namespace MobileGL::MG_Backend::DirectGLES {
 
             using FramebufferObject = MG_State::GLState::FramebufferObject;
             FramebufferObject::FramebufferAttachmentVersionArray m_syncedFrontendAttachmentVersions = {0};
+            // g_attachmentBackendIdGeneration as of this twin's last attachment walk. A
+            // mismatch means some backend texture id was re-minted since, and any of this
+            // twin's attachment points may still hold the dead id even though the frontend
+            // attachment versions match - so the walk re-attaches everything first.
+            Uint64 m_syncedBackendIdGeneration = 0;
         };
 
         extern StateBackendObjectRegistry<MG_State::GLState::FramebufferObject, BackendFramebufferObject>
@@ -811,6 +911,19 @@ namespace MobileGL::MG_Backend::DirectGLES {
         extern Array<MG_State::GLState::FramebufferObject*, SizeT(FramebufferTarget::FramebufferTargetCount)>
             g_fboSyncedObjects;
 
+        // Bumped whenever a live backend texture's driver id is re-minted while its
+        // frontend texture may still be attached to application FBOs
+        // (BackendTextureObject::RecreateBackendTexture - e.g. a respecify of a texture
+        // whose backend storage went immutable). The FBO twins' attachment memos key on
+        // FRONTEND attachment versions, which a backend-side re-mint does not move, so
+        // the driver FBO would keep the deleted texture name attached forever. The
+        // SyncCurrentFBO gate compares this generation (below) to re-enter the sync,
+        // and each twin re-arms its per-attachment memo on a mismatch (SyncToBackend).
+        extern Uint64 g_attachmentBackendIdGeneration;
+        // What g_attachmentBackendIdGeneration was when SyncCurrentFBO last stamped each
+        // target; part of the synced tuple above.
+        extern Array<Uint64, SizeT(FramebufferTarget::FramebufferTargetCount)> g_fboSyncedBackendIdGenerations;
+
         // Driver-level READ/DRAW framebuffer-binding shadow. Every backend
         // glBindFramebuffer routes through BindFramebufferId so scoped helpers can
         // save/restore the current binding without a glGetIntegerv round-trip (that
@@ -821,6 +934,10 @@ 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
@@ -997,9 +1114,13 @@ namespace MobileGL::MG_Backend::DirectGLES {
             void SetBaseInstance(Uint32 baseInstance) const;
             void SetBaseInstanceWordIndex(Int32 wordIndex) const;
             void SetDrawID(Uint32 drawId) const;
+            void SetBaseVertex(Int32 baseVertex) const;
             // True when the transpiled program kept a gl_DrawID uniform, i.e. SetDrawID
             // actually reaches a shader read rather than being discarded.
             Bool ReadsDrawID() const { return m_drawIdUniformLocation >= 0; }
+            // Same for gl_BaseVertex: only a program that reads it pays for the per-draw
+            // uniform write, and only such a program needs the reset after one.
+            Bool ReadsBaseVertex() const { return m_baseVertexUniformLocation >= 0; }
             Int GetIndirectParamsBinding() const { return m_indirectParamsBinding; }
             Uint GetBackendProgramId() const { return m_backendProgramId; }
             // False when the last SyncToBackend could not produce a usable program (a
@@ -1010,6 +1131,11 @@ 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; }
@@ -1025,14 +1151,48 @@ 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; }
+            // Whether the (unit, bound format) pairs this program's FORMAT-LESS image uniforms
+            // resolve to are still the ones its ESSL was generated against.
+            //
+            // A fourth condition of the same family as the three above, and the only one that
+            // reads live state rather than a program-side counter, because that is where the
+            // dependency actually is. GLSL ES requires a format layout qualifier on every image
+            // where desktop GLSL lets a writeonly declaration omit one, and the only correct
+            // qualifier is whatever glBindImageTexture named - so a declaration with no format
+            // is compiled against the BINDING, and a rebind to a different format makes the
+            // built program wrong. Keyed on the units the program's own images address (cached
+            // at sync, since a unit can only move by glUniform1i, which bumps the image-unit
+            // version above and forces a re-sync anyway), so the cost on a program with no
+            // format-less image - which is all but a handful - is one empty-vector test.
+            //
+            // Deliberately NOT reached from glBindImageTexture: that entry point must never
+            // trigger a build (same constraint as glShaderStorageBlockBinding). It moves the
+            // state and this comparison notices at the next Prepare, which is also what makes
+            // an image first bound AFTER link work.
+            Bool ImageUnitFormatsStillMatch() const;
+            // The value ImageUnitFormatsStillMatch() compares against, recomputed from live
+            // image-unit state. 0 when the program has no format-less image uniform.
+            Uint64 ComputeImageUnitFormatSignature() const;
 
         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;
+            Int m_baseVertexUniformLocation = -1;
             Int m_baseInstanceWordIndexUniformLocation = -1;
             Int m_indirectParamsBinding = -1;
             Uint32 m_snormFallbackClampOutputMask = 0;
@@ -1040,6 +1200,8 @@ 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 +1212,13 @@ namespace MobileGL::MG_Backend::DirectGLES {
             Uint32 m_lastUploadedGlobalUboVersion = ~0u;
             BufferImpl::UboRingAllocation m_globalUboRingAllocation;
             Uint32 m_syncedLinkVersion = ~0u;
+            Uint32 m_syncedImageUnitVersion = ~0u;
+            // Image units addressed by the program's FORMAT-LESS image uniforms, and the digest
+            // of the (unit, format) pairs the generated ESSL baked. Empty/0 for every program
+            // that declares a format on all of its images, which is the overwhelming majority -
+            // and what keeps the per-draw comparison free for them.
+            Vector<Int> m_formatlessImageUnits;
+            Uint64 m_imageUnitFormatSignature = 0;
             SamplerPassMemo m_samplerPassMemo;
         };
 
@@ -1073,26 +1242,79 @@ 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
+        // program that was just built - best effort, on the same "only where the driver has
-        // shader's DECLARED binding survives in the SPIR-V), so without this replay any
+        // the entry point" terms as ApplyShaderStorageBlockBinding above. Mirrors
-        // rebuild would silently revert rebound blocks. Mirrors DirectVulkan's
+        // DirectVulkan's reseed-on-rebuild in BuildProgramResourceCache.
-        // 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);
+
+        // Everything the image-format bake needs from one walk of a program's uniform
+        // reflection. GLSL ES requires a format layout qualifier on every image uniform;
+        // desktop GLSL lets a writeonly (or readonly) declaration omit one, and the only
+        // format that is CORRECT to substitute is whatever glBindImageTexture named for the
+        // unit that uniform addresses - so the transpile bakes it in and the build is keyed
+        // on it.
+        struct ImageFormatBakeInputs {
+            // Uniform name (SPIR-V spelling, i.e. an array named once, unsubscripted) to the GL
+            // internal format to bake. Holds only uniforms that DECLARED no format; a declared
+            // one is authoritative and is never overridden.
+            UnorderedMap<String, Uint> glFormatByUniformName;
+            // The same uniforms whose format SPIRV-Cross REFUSES to print for ESSL (it throws on
+            // its desktop-only set, which loses the stage), paired with the ESSL spelling to
+            // write into the emitted declaration instead. Disjoint from the map above by
+            // construction: a format is baked into the module or completed in the text, never
+            // both. r8ui - the stencil half of the packed_depth_stencil case - lands here.
+            UnorderedMap<String, String> esslFormatQualifierByUniformName;
+            // Units those uniforms address, kept so the draw path can re-read their formats
+            // without walking the reflection again.
+            Vector<Int> units;
+            // Digest of the (unit, format) pairs above. 0 when the program has no format-less
+            // image uniform, which is all but a handful.
+            Uint64 signature = 0;
+            // Array uniforms whose elements resolved to units holding DIFFERENT formats: one
+            // declaration carries one qualifier, so there is nothing correct to bake and they
+            // are dropped from the map above. Kept for diagnostics.
+            Vector<String> conflictedNames;
+            // Some format in play - declared or baked - is outside the GLSL ES core image
+            // format set, so the emitted ESSL needs the GL_NV_image_formats directive.
+            Bool needsExtendedImageFormats = false;
+        };
+        ImageFormatBakeInputs CollectImageFormatBakeInputs(
+            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 +1332,18 @@ 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/MultiDraw.cpp

@@ -252,7 +252,7 @@ namespace MobileGL::MG_Backend::DirectGLES::MultiDrawImpl {
             g_resolvedTier =
                 ResolveTier(g_GLESCapabilities, g_GLESFuncs, MG_Config::Features.EsprytMultiDrawMode,
                             &g_tierResolution);
-            MGLOG_I("DirectGLES multi-draw: %s", g_tierResolution.c_str());
+            MGLOG_D("DirectGLES multi-draw: %s", g_tierResolution.c_str());
         }
 
         // Which tiers have already announced themselves, one bit per GLESMultiDrawMode.
@@ -267,24 +267,29 @@ namespace MobileGL::MG_Backend::DirectGLES::MultiDrawImpl {
             const Uint32 bit = 1u << static_cast<Uint32>(tier);
             if (g_announcedTiers & bit) return;
             g_announcedTiers |= bit;
-            MGLOG_I("DirectGLES multi-draw: first batch executed via tier \"%s\"", TierName(tier));
+            MGLOG_D("DirectGLES multi-draw: first batch executed via tier \"%s\"", TierName(tier));
         }
 
         // The tier this particular batch can actually take. A tier is demoted here when
         // the batch's own shape - not the driver - rules it out; the compute tier keeps
         // its remaining feasibility checks inside its implementation, where the data it
         // has to walk is already in hand.
-        GLESMultiDrawMode ResolveTierForBatch(Bool programReadsDrawID, Bool hasIndexBuffer) {
+        GLESMultiDrawMode ResolveTierForBatch(Bool programReadsDrawID, Bool perSubDrawBaseVertex,
+                                              Bool hasIndexBuffer) {
             ResolveTierOnce();
             GLESMultiDrawMode tier = g_resolvedTier;
 
             // Batched tiers issue one driver entry for the whole batch, so the emulated
             // gl_DrawID uniform can only hold one value across every sub-draw. A program
             // that reads gl_DrawID gets an unrolled tier, which feeds each sub-draw its
-            // own index (the spec's value); nothing else observes the difference.
+            // own index (the spec's value); nothing else observes the difference. The
+            // emulated gl_BaseVertex is one uniform for the same reason, so a batch whose
+            // sub-draws carry their own base vertices unrolls too - even the Ext tier,
+            // which hands the driver the whole basevertex array, can only leave ONE value
+            // in the uniform the shader reads.
             const Bool batched = tier == GLESMultiDrawMode::Ext || tier == GLESMultiDrawMode::MultiIndirect ||
                                  tier == GLESMultiDrawMode::Compute;
-            if (batched && programReadsDrawID) {
+            if (batched && (programReadsDrawID || perSubDrawBaseVertex)) {
                 tier = SupportsTier(GLESMultiDrawMode::BaseVertex) ? GLESMultiDrawMode::BaseVertex
                                                                    : GLESMultiDrawMode::DrawElements;
             }
@@ -371,7 +376,8 @@ namespace MobileGL::MG_Backend::DirectGLES::MultiDrawImpl {
         // ---------------------------------------------------------------------------
 
         Bool RunIndirect(GLenum mode, const GLsizei* count, GLenum type, const GLvoid* const* indices,
-                         GLsizei drawcount, const GLint* basevertex, Bool batched, Bool feedDrawID) {
+                         GLsizei drawcount, const GLint* basevertex, Bool batched, Bool feedDrawID,
+                         Bool feedBaseVertex) {
             if (!SupportsTier(batched ? GLESMultiDrawMode::MultiIndirect : GLESMultiDrawMode::Indirect)) return false;
             const SizeT indexSize = IndexTypeSize(type);
             if (indexSize == 0) return false;
@@ -413,10 +419,12 @@ namespace MobileGL::MG_Backend::DirectGLES::MultiDrawImpl {
             } else {
                 for (GLsizei i = 0; i < drawcount; ++i) {
                     if (feedDrawID) SetCurrentDrawID(static_cast<Uint32>(i));
+                    if (feedBaseVertex) SetCurrentBaseVertex(basevertex ? basevertex[i] : 0);
                     const SizeT commandOffset = commandBase + static_cast<SizeT>(i) * sizeof(DrawElementsIndirectCommand);
                     g_GLESFuncs.glDrawElementsIndirect(mode, type, reinterpret_cast<const void*>(commandOffset));
                 }
                 if (feedDrawID) SetCurrentDrawID(0);
+                if (feedBaseVertex) SetCurrentBaseVertex(0);
             }
             BufferImpl::BindBufferId(GL_DRAW_INDIRECT_BUFFER, previousIndirectBinding);
             NoteTierExecuted(batched ? GLESMultiDrawMode::MultiIndirect : GLESMultiDrawMode::Indirect);
@@ -428,15 +436,17 @@ namespace MobileGL::MG_Backend::DirectGLES::MultiDrawImpl {
         // ---------------------------------------------------------------------------
 
         Bool RunBaseVertexLoop(GLenum mode, const GLsizei* count, GLenum type, const GLvoid* const* indices,
-                               GLsizei drawcount, const GLint* basevertex, Bool feedDrawID) {
+                               GLsizei drawcount, const GLint* basevertex, Bool feedDrawID, Bool feedBaseVertex) {
             if (!SupportsTier(GLESMultiDrawMode::BaseVertex)) return false;
             for (GLsizei i = 0; i < drawcount; ++i) {
                 if (count[i] <= 0) continue;
                 if (feedDrawID) SetCurrentDrawID(static_cast<Uint32>(i));
+                if (feedBaseVertex) SetCurrentBaseVertex(basevertex ? basevertex[i] : 0);
                 g_GLESFuncs.glDrawElementsBaseVertex(mode, count[i], type, indices[i],
                                                      basevertex ? basevertex[i] : 0);
             }
             if (feedDrawID) SetCurrentDrawID(0);
+            if (feedBaseVertex) SetCurrentBaseVertex(0);
             NoteTierExecuted(GLESMultiDrawMode::BaseVertex);
             return true;
         }
@@ -446,7 +456,8 @@ namespace MobileGL::MG_Backend::DirectGLES::MultiDrawImpl {
         // ---------------------------------------------------------------------------
 
         Bool RunRebasedDrawElements(GLenum mode, const GLsizei* count, GLenum type, const GLvoid* const* indices,
-                                    GLsizei drawcount, const GLint* basevertex, Bool feedDrawID) {
+                                    GLsizei drawcount, const GLint* basevertex, Bool feedDrawID,
+                                    Bool feedBaseVertex) {
             const SizeT indexSize = IndexTypeSize(type);
             if (indexSize == 0) return false;
 
@@ -479,7 +490,7 @@ namespace MobileGL::MG_Backend::DirectGLES::MultiDrawImpl {
                 const Uint8* source = ResolveSubDrawIndices(indexBuffer, indexBufferBytes, indexBufferSize, indices[i],
                                                             subDrawCount, indexSize);
                 if (!source) {
-                    MGLOG_E("DirectGLES multi-draw (drawelements tier): sub-draw %d reads outside the bound index "
+                    MGLOG_E_ONCE("DirectGLES multi-draw (drawelements tier): sub-draw %d reads outside the bound index "
                             "buffer; skipping the batch",
                             i);
                     return false;
@@ -500,11 +511,16 @@ namespace MobileGL::MG_Backend::DirectGLES::MultiDrawImpl {
             for (GLsizei i = 0; i < drawcount; ++i) {
                 if (count[i] <= 0) continue;
                 if (feedDrawID) SetCurrentDrawID(static_cast<Uint32>(i));
+                // The base vertex is folded into the rewritten index stream here, so the
+                // driver sees none - but gl_BaseVertex still has to report the value the
+                // application passed for this sub-draw.
+                if (feedBaseVertex) SetCurrentBaseVertex(basevertex ? basevertex[i] : 0);
                 g_GLESFuncs.glDrawElements(mode, count[i], GL_UNSIGNED_INT,
                                            reinterpret_cast<const void*>(indexBase + cursor * sizeof(Uint32)));
                 cursor += static_cast<SizeT>(count[i]);
             }
             if (feedDrawID) SetCurrentDrawID(0);
+            if (feedBaseVertex) SetCurrentBaseVertex(0);
             BufferImpl::BindBufferId(GL_ELEMENT_ARRAY_BUFFER, previousIndexBinding);
             NoteTierExecuted(GLESMultiDrawMode::DrawElements);
             return true;
@@ -580,7 +596,7 @@ void main() {
 
             const GLuint shader = g_GLESFuncs.glCreateShader(GL_COMPUTE_SHADER);
             if (shader == 0) {
-                MGLOG_E("DirectGLES multi-draw (compute tier): glCreateShader(GL_COMPUTE_SHADER) failed");
+                MGLOG_E_ONCE("DirectGLES multi-draw (compute tier): glCreateShader(GL_COMPUTE_SHADER) failed");
                 return false;
             }
             const char* source = kFlattenComputeSource;
@@ -591,14 +607,14 @@ void main() {
             if (status != GL_TRUE) {
                 char log[1024] = {};
                 g_GLESFuncs.glGetShaderInfoLog(shader, sizeof(log) - 1, nullptr, log);
-                MGLOG_E("DirectGLES multi-draw (compute tier): index-flattening shader failed to compile: %s", log);
+                MGLOG_E_ONCE("DirectGLES multi-draw (compute tier): index-flattening shader failed to compile: %s", log);
                 g_GLESFuncs.glDeleteShader(shader);
                 return false;
             }
 
             const GLuint program = g_GLESFuncs.glCreateProgram();
             if (program == 0) {
-                MGLOG_E("DirectGLES multi-draw (compute tier): glCreateProgram failed");
+                MGLOG_E_ONCE("DirectGLES multi-draw (compute tier): glCreateProgram failed");
                 g_GLESFuncs.glDeleteShader(shader);
                 return false;
             }
@@ -609,7 +625,7 @@ void main() {
             if (status != GL_TRUE) {
                 char log[1024] = {};
                 g_GLESFuncs.glGetProgramInfoLog(program, sizeof(log) - 1, nullptr, log);
-                MGLOG_E("DirectGLES multi-draw (compute tier): index-flattening program failed to link: %s", log);
+                MGLOG_E_ONCE("DirectGLES multi-draw (compute tier): index-flattening program failed to link: %s", log);
                 g_GLESFuncs.glDeleteProgram(program);
                 return false;
             }
@@ -619,7 +635,7 @@ void main() {
             g_uDrawCount = g_GLESFuncs.glGetUniformLocation(program, "uDrawCount");
             g_uTotalIndices = g_GLESFuncs.glGetUniformLocation(program, "uTotalIndices");
             g_computeProgramFailed = false;
-            MGLOG_I("DirectGLES multi-draw: index-flattening compute program ready (id %u)", program);
+            MGLOG_D("DirectGLES multi-draw: index-flattening compute program ready (id %u)", program);
             return true;
         }
 
@@ -837,8 +853,15 @@ void main() {
         // afterwards would mean unpicking the program, SSBO and index bindings
         // PrepareForDraw just made, and a dispatch inside an open transform feedback
         // span is not legal at all. On success it hands back a flattened index stream.
+        // A batch whose sub-draws carry their own base vertices cannot be flattened either
+        // when the program reads gl_BaseVertex: one draw call leaves one uniform value.
+        // Asked conservatively because this decision precedes PrepareForDraw - see
+        // CurrentProgramMayNeedPerSubDrawBuiltins. Flattening is the irreversible half:
+        // once the batch is one draw the values are gone, whereas declining to flatten only
+        // costs the unrolled tier.
         FlattenedStream flattened;
-        if (ResolvedTier() == GLESMultiDrawMode::Compute && !CurrentProgramReadsDrawID()) {
+        if (ResolvedTier() == GLESMultiDrawMode::Compute &&
+            !CurrentProgramMayNeedPerSubDrawBuiltins(basevertex != nullptr)) {
             FlattenWithCompute(mode, count, type, indices, drawcount, basevertex, flattened);
         }
 
@@ -852,8 +875,11 @@ void main() {
             return;
         }
 
+        // Now that PrepareForDraw has synced the program, both questions have real answers;
+        // the tier choice and the per-sub-draw feeds use those, not the guess above.
         const Bool feedDrawID = CurrentProgramReadsDrawID();
-        const GLESMultiDrawMode tier = ResolveTierForBatch(feedDrawID, hasIndexBuffer);
+        const Bool feedBaseVertex = basevertex != nullptr && CurrentProgramReadsBaseVertex();
+        const GLESMultiDrawMode tier = ResolveTierForBatch(feedDrawID, feedBaseVertex, hasIndexBuffer);
 
         Bool drawn = false;
         switch (tier) {
@@ -861,16 +887,19 @@ void main() {
             drawn = RunExt(mode, count, type, indices, drawcount, basevertex);
             break;
         case GLESMultiDrawMode::MultiIndirect:
-            drawn = RunIndirect(mode, count, type, indices, drawcount, basevertex, /*batched=*/true, feedDrawID);
+            drawn = RunIndirect(mode, count, type, indices, drawcount, basevertex, /*batched=*/true, feedDrawID,
+                                feedBaseVertex);
             break;
         case GLESMultiDrawMode::Indirect:
-            drawn = RunIndirect(mode, count, type, indices, drawcount, basevertex, /*batched=*/false, feedDrawID);
+            drawn = RunIndirect(mode, count, type, indices, drawcount, basevertex, /*batched=*/false, feedDrawID,
+                                feedBaseVertex);
             break;
         case GLESMultiDrawMode::BaseVertex:
-            drawn = RunBaseVertexLoop(mode, count, type, indices, drawcount, basevertex, feedDrawID);
+            drawn = RunBaseVertexLoop(mode, count, type, indices, drawcount, basevertex, feedDrawID, feedBaseVertex);
             break;
         case GLESMultiDrawMode::DrawElements:
-            drawn = RunRebasedDrawElements(mode, count, type, indices, drawcount, basevertex, feedDrawID);
+            drawn = RunRebasedDrawElements(mode, count, type, indices, drawcount, basevertex, feedDrawID,
+                                           feedBaseVertex);
             break;
         case GLESMultiDrawMode::Compute:
             // Its pre-pass ran above; reaching here means it declined this batch's shape.
@@ -883,10 +912,15 @@ void main() {
         // below are the floor: a base-vertex replay where the driver has one, and the
         // rewritten index stream where it does not. Both are safe for any batch these
         // entry points can receive.
-        if (!drawn) drawn = RunBaseVertexLoop(mode, count, type, indices, drawcount, basevertex, feedDrawID);
-        if (!drawn) drawn = RunRebasedDrawElements(mode, count, type, indices, drawcount, basevertex, feedDrawID);
         if (!drawn) {
-            MGLOG_E("DirectGLES multi-draw: no usable tier for a %d sub-draw batch (mode 0x%x, type 0x%x); "
+            drawn = RunBaseVertexLoop(mode, count, type, indices, drawcount, basevertex, feedDrawID, feedBaseVertex);
+        }
+        if (!drawn) {
+            drawn = RunRebasedDrawElements(mode, count, type, indices, drawcount, basevertex, feedDrawID,
+                                           feedBaseVertex);
+        }
+        if (!drawn) {
+            MGLOG_E_ONCE("DirectGLES multi-draw: no usable tier for a %d sub-draw batch (mode 0x%x, type 0x%x); "
                     "the batch was dropped",
                     drawcount, mode, type);
         }

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>
@@ -416,10 +417,27 @@ namespace MobileGL::MG_Backend::DirectGLES {
                 result = std::regex_replace(result, pattern, "$1flat $2");
             };
 
+            // Every stage that has an integer interface at all, on BOTH sides. Interpolation is
+            // only ever consumed at a fragment input, so the qualifier is semantically inert on
+            // a tessellation or geometry interface - but an ES linker still compares the two
+            // sides of every interface and rejects a program whose producer says `flat` and
+            // whose consumer does not. Covering only the stages that "need" it left exactly two
+            // holes, and a program that used tessellation fell into both:
+            //   vertex `flat out uint` -> tess-control `in uint`   (producer flat, consumer not)
+            //   tess-eval `out uint`   -> geometry `flat in uint`  (consumer flat, producer not)
+            // Adreno answers "output ... interpolation mismatch with other stage" and the whole
+            // program fails to link, which is a draw that silently paints nothing.
+            //
+            // Adding rather than stripping, because a fragment input's `flat` is load-bearing
+            // (ESSL forbids an interpolated integer) and would have to be put back for the last
+            // stage before the fragment shader anyway - so "everything integer is flat" is the
+            // one rule that is consistent no matter which stages a program happens to have.
             switch (shaderType) {
             case GL_VERTEX_SHADER:
                 addFlatQualifier("out");
                 break;
+            case GL_TESS_CONTROL_SHADER:
+            case GL_TESS_EVALUATION_SHADER:
             case GL_GEOMETRY_SHADER:
                 addFlatQualifier("in");
                 addFlatQualifier("out");
@@ -434,6 +452,175 @@ 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 RequestExtendedImageFormats(String glslCode, Bool needed) {
+#ifdef TRACY_ENABLE
+            ZoneScopedC(TRACY_ZONECOLOR_BACKEND);
+#endif
+            // GLSL ES core has thirteen image formats; GL has forty. SPIRV-Cross prints whatever
+            // format the OpTypeImage carries and asks for no extension for it, so an r8ui or
+            // rg16f image - declared as such, or baked from the bound one - reaches the driver as
+            // a format its core language does not know. GL_NV_image_formats is the only thing
+            // that adds them, and it has to be requested by name.
+            //
+            // The caller decides `needed`: it knows which formats are in play (from the uniform
+            // reflection and the image-unit bindings) and whether the driver advertises the
+            // extension at all - `#extension` on an unadvertised name is itself a hard error, so
+            // this must never be emitted speculatively.
+            static constexpr const char* kDirective = "#extension GL_NV_image_formats : require\n";
+            static constexpr const char* kExtName = "GL_NV_image_formats";
+            if (!needed || glslCode.find(kExtName) != String::npos) {
+                return glslCode;
+            }
+            // After the #version line, which must stay first. Everything else about the header is
+            // order-insensitive, and ForceSupporterOutput's scan for the LAST #extension
+            // directive still finds whichever one that is.
+            const SizeT versionPos = glslCode.find("#version");
+            if (versionPos == String::npos) {
+                return kDirective + glslCode;
+            }
+            const SizeT lineEnd = glslCode.find('\n', versionPos);
+            if (lineEnd == String::npos) {
+                return glslCode + "\n" + kDirective;
+            }
+            glslCode.insert(lineEnd + 1, kDirective);
+            return glslCode;
+        }
+
+        String BakeImageFormatQualifiers(String glslCode,
+                                         const UnorderedMap<String, String>& esslFormatByUniformName) {
+#ifdef TRACY_ENABLE
+            ZoneScopedC(TRACY_ZONECOLOR_BACKEND);
+#endif
+            if (esslFormatByUniformName.empty() || glslCode.find("image") == String::npos) {
+                return glslCode;
+            }
+            // Same declaration shape RebindImageUniformsToFrontendUnits matches, and for the same
+            // reason: one line, one image uniform, the name in group 3.
+            static const std::regex imageDeclRegex(
+                R"((layout\s*\(([^)]*)\)\s*)?uniform\s+(?:(?:readonly|writeonly|coherent|volatile|restrict|highp|mediump|lowp)\s+)*[iu]?image[A-Za-z0-9]+\s+([A-Za-z_][A-Za-z0-9_]*)\s*(\[[^\]]*\])?\s*;)");
+            // Every image format spelling GLSL has, so a declaration that already carries one is
+            // recognised whatever it says - the caller's map is consulted only for declarations
+            // with NO format, never to override a written one.
+            static const std::regex existingFormatRegex(
+                R"(\b(rgba32f|rgba16f|rg32f|rg16f|r11f_g11f_b10f|r32f|r16f|rgba16|rgb10_a2|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|rgba8ui|rgb10_a2ui|rg32ui|rg16ui|rg8ui|r32ui|r16ui|r8ui)\b)");
+
+            String result;
+            result.reserve(glslCode.size());
+            SizeT lineStart = 0;
+            while (lineStart <= glslCode.size()) {
+                const SizeT lineEnd = glslCode.find('\n', lineStart);
+                const Bool lastLine = lineEnd == String::npos;
+                String line = glslCode.substr(lineStart, lastLine ? String::npos : lineEnd - lineStart);
+
+                std::smatch match;
+                if (std::regex_search(line, match, imageDeclRegex)) {
+                    const String name = match[3].str();
+                    const auto formatIt = esslFormatByUniformName.find(name);
+                    const String layoutContents = match[2].matched ? match[2].str() : String();
+                    if (formatIt != esslFormatByUniformName.end() && !formatIt->second.empty() &&
+                        !std::regex_search(layoutContents, existingFormatRegex)) {
+                        if (match[1].matched) {
+                            const SizeT layoutOpen = line.find('(', match.position(1));
+                            line.insert(layoutOpen + 1, formatIt->second + ", ");
+                        } else {
+                            line.insert(match.position(0), "layout(" + formatIt->second + ") ");
+                        }
+                    }
+                }
+
+                result += line;
+                if (lastLine) {
+                    break;
+                }
+                result += '\n';
+                lineStart = lineEnd + 1;
+            }
+            return result;
+        }
+
         String RemoveLayoutBinding(const String& glslCode) {
 #ifdef TRACY_ENABLE
             ZoneScopedC(TRACY_ZONECOLOR_BACKEND);
@@ -470,6 +657,352 @@ 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 +1060,7 @@ 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 +1121,11 @@ 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.
@@ -638,7 +1176,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
             ZoneScopedC(TRACY_ZONECOLOR_BACKEND);
 #endif
             for (GLenum err = g_GLESFuncs.glGetError(); err != GL_NO_ERROR; err = g_GLESFuncs.glGetError()) {
-                MGLOG_E("-> GLES Error: %s", MG_Util::ConvertGLEnumToString(err).c_str());
+                MGLOG_D("-> GLES Error: %s", MG_Util::ConvertGLEnumToString(err).c_str());
             }
         }
 
@@ -1058,26 +1596,18 @@ namespace MobileGL::MG_Backend::DirectGLES {
             return (rowBytes + align - 1) / align * align;
         }
 
-    // Repacks wide RGBA(_INTEGER) rows into the client's (format, type) layout, honoring the
+        // Walks the client-side destination the PACK parameters describe and hands each row to
-        // client-side PACK parameters and the bound pixel-pack buffer. `wide` holds
+        // `fillRow(slice, row, dstRow)`, which writes width * dstPixelBytes bytes of finished client
-        // `sliceHeight * sliceCount` rows of `width` texels (slice-major, tightly stacked),
+        // texels. Shared by the converting and the raw-word stores so both address the destination -
-        // 4 components x GetReadbackComponentSize(wideType) bytes each.
+        // and feed the bound pixel-pack buffer - identically.
         // applyPackImageParams: GL_PACK_IMAGE_HEIGHT / GL_PACK_SKIP_IMAGES apply only to GetTexImage
         // of 3D/array images; ReadPixels and 2D GetTexImage ignore them (GL 3.3 sections 4.3.1, 6.1.4).
         // Per the GL addressing rules, slice k row j lands at
         // SKIP_IMAGES*imageStride + SKIP_ROWS*rowStride + SKIP_PIXELS*pixelBytes
         //   + k*imageStride + j*rowStride, with imageStride = max(IMAGE_HEIGHT, sliceHeight)*rowStride.
-        Bool StoreWideRowsToClient(const Uint8* wide, GLenum wideType, GLsizei width, GLsizei sliceHeight,
+        template <typename FillRow>
-                                   GLsizei sliceCount, const ReadbackChannelMapping& mapping, GLenum type,
+        static Bool StoreClientRows(SizeT dstPixelBytes, SizeT swapGroupSize, GLsizei width, GLsizei sliceHeight,
-                                   void* pixels, Bool applyPackImageParams) {
+                                    GLsizei sliceCount, void* pixels, Bool applyPackImageParams, FillRow&& fillRow) {
-        const SizeT dstPixelBytes = GetReadbackDstPixelSize(mapping, type);
-        if (dstPixelBytes == 0) {
-            return false;
-        }
-        PackedReadbackLayout packedLayout{};
-        const Bool isPackedType = GetPackedReadbackLayout(type, packedLayout);
-        const SizeT dstComponentSize = GetReadbackComponentSize(type);
-
             const auto& pixelPackBufferObject =
                 MG_State::pGLContext->GetBufferBindingSlot(BufferTarget::PixelPack).GetBoundObject();
 
@@ -1104,29 +1634,20 @@ namespace MobileGL::MG_Backend::DirectGLES {
                                         static_cast<SizeT>(sliceCount - 1) * dstImageStride +
                                         static_cast<SizeT>(sliceHeight - 1) * dstRowStride + dstRowBytes;
                 if (requiredSize > pixelPackBufferObject->GetSize()) {
-                MGLOG_E("Readback conversion: pixel pack buffer is too small");
+                    MGLOG_E_ONCE("Readback conversion: pixel pack buffer is too small");
                     return true;
                 }
             }
 
-        const SizeT srcComponentSize = GetReadbackComponentSize(wideType);
-        const SizeT srcPixelBytes = 4 * srcComponentSize;
             Vector<Uint8> convertedRow(dstRowBytes);
 
             for (GLsizei slice = 0; slice < sliceCount; ++slice) {
                 for (GLsizei row = 0; row < sliceHeight; ++row) {
-                const SizeT flatRow = static_cast<SizeT>(slice) * static_cast<SizeT>(sliceHeight) +
+                    fillRow(slice, row, convertedRow.data());
-                                   static_cast<SizeT>(row);
-                const Uint8* srcRow = wide + flatRow * static_cast<SizeT>(width) * srcPixelBytes;
-                ConvertWideReadbackRow(srcRow, convertedRow.data(), static_cast<SizeT>(width), wideType,
-                                                  mapping, type);
 
-                if (packParams.SwapBytes) {
+                    if (packParams.SwapBytes && swapGroupSize > 1) {
-                    const SizeT groupSize = isPackedType ? packedLayout.byteSize : dstComponentSize;
+                        for (SizeT offset = 0; offset + swapGroupSize <= dstRowBytes; offset += swapGroupSize) {
-                    if (groupSize > 1) {
+                            std::reverse(convertedRow.data() + offset, convertedRow.data() + offset + swapGroupSize);
-                        for (SizeT offset = 0; offset + groupSize <= dstRowBytes; offset += groupSize) {
-                            std::reverse(convertedRow.data() + offset, convertedRow.data() + offset + groupSize);
-                        }
                         }
                     }
 
@@ -1147,5 +1668,52 @@ namespace MobileGL::MG_Backend::DirectGLES {
             }
             return true;
         }
+
+        // Repacks wide RGBA(_INTEGER) rows into the client's (format, type) layout, honoring the
+        // client-side PACK parameters and the bound pixel-pack buffer. `wide` holds
+        // `sliceHeight * sliceCount` rows of `width` texels (slice-major, tightly stacked),
+        // 4 components x GetReadbackComponentSize(wideType) bytes each.
+        Bool StoreWideRowsToClient(const Uint8* wide, GLenum wideType, GLsizei width, GLsizei sliceHeight,
+                                   GLsizei sliceCount, const ReadbackChannelMapping& mapping, GLenum type,
+                                   void* pixels, Bool applyPackImageParams) {
+            const SizeT dstPixelBytes = GetReadbackDstPixelSize(mapping, type);
+            if (dstPixelBytes == 0) {
+                return false;
+            }
+            PackedReadbackLayout packedLayout{};
+            const Bool isPackedType = GetPackedReadbackLayout(type, packedLayout);
+            const SizeT swapGroupSize = isPackedType ? packedLayout.byteSize : GetReadbackComponentSize(type);
+            const SizeT srcPixelBytes = 4 * GetReadbackComponentSize(wideType);
+
+            return StoreClientRows(dstPixelBytes, swapGroupSize, width, sliceHeight, sliceCount, pixels,
+                                   applyPackImageParams,
+                                   [&](GLsizei slice, GLsizei row, Uint8* dstRow) {
+                                       const SizeT flatRow = static_cast<SizeT>(slice) *
+                                                                 static_cast<SizeT>(sliceHeight) +
+                                                             static_cast<SizeT>(row);
+                                       const Uint8* srcRow =
+                                           wide + flatRow * static_cast<SizeT>(width) * srcPixelBytes;
+                                       ConvertWideReadbackRow(srcRow, dstRow, static_cast<SizeT>(width), wideType,
+                                                              mapping, type);
+                                   });
+        }
+
+        Bool StorePackedWordsToClient(const Uint8* srcWords, GLsizei width, GLsizei sliceHeight, GLsizei sliceCount,
+                                      GLenum type, void* pixels, Bool applyPackImageParams) {
+            PackedReadbackLayout packedLayout{};
+            if (!GetPackedReadbackLayout(type, packedLayout) || packedLayout.byteSize != 4) {
+                return false;
+            }
+            const SizeT srcRowBytes = static_cast<SizeT>(width) * 4;
+
+            return StoreClientRows(4, packedLayout.byteSize, width, sliceHeight, sliceCount, pixels,
+                                   applyPackImageParams,
+                                   [&](GLsizei slice, GLsizei row, Uint8* dstRow) {
+                                       const SizeT flatRow = static_cast<SizeT>(slice) *
+                                                                 static_cast<SizeT>(sliceHeight) +
+                                                             static_cast<SizeT>(row);
+                                       Memcpy(dstRow, srcWords + flatRow * srcRowBytes, srcRowBytes);
+                                   });
+        }
     } // namespace ReadbackImpl
 } // namespace MobileGL::MG_Backend::DirectGLES

MobileGL/MG_Backend/DirectGLES/Utils.h

@@ -115,6 +115,16 @@ namespace MobileGL::MG_Backend::DirectGLES {
         Bool StoreWideRowsToClient(const Uint8* wide, GLenum wideType, GLsizei width, GLsizei sliceHeight,
                                    GLsizei sliceCount, const ReadbackChannelMapping& mapping, GLenum type,
                                    void* pixels, Bool applyPackImageParams);
+
+        // Stores packed 32-bit source words verbatim, with the same destination addressing, PACK
+        // parameters and pixel-pack-buffer handling as StoreWideRowsToClient. For the sources whose
+        // storage word already IS the client word (MG_Util::IsRawPackedPixelTransfer): routing those
+        // through the wide float intermediate re-encodes them, and the RGB9_E5 encoder canonicalizes
+        // the shared exponent, so glGetTexImage would answer with different bits than were stored.
+        // `srcWords` holds sliceHeight * sliceCount tightly stacked rows of `width` 32-bit words.
+        // False when `type` is not a 4-byte packed type.
+        Bool StorePackedWordsToClient(const Uint8* srcWords, GLsizei width, GLsizei sliceHeight, GLsizei sliceCount,
+                                      GLenum type, void* pixels, Bool applyPackImageParams);
     } // namespace ReadbackImpl
 
     namespace PrgramImpl {
@@ -130,7 +140,68 @@ 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);
+        // Adds `#extension GL_NV_image_formats : require` when the shader carries an image
+        // format qualifier GLSL ES has no core spelling for. SPIRV-Cross prints the format and
+        // asks for nothing, so the request has to be made here. `needed` is the caller's answer,
+        // because only it knows which formats are in play AND whether the driver advertises the
+        // extension - requesting an unadvertised extension is itself a compile error, so this is
+        // never emitted speculatively. A no-op when not needed or already present.
+        String RequestExtendedImageFormats(String glslCode, Bool needed);
+        // Writes a format layout qualifier into the image declarations named in
+        // `esslFormatByUniformName` that still have none. The completion half of the image-format
+        // bake, and ONLY that: the SPIR-V pass (BakeImageFormatsPass) is what normally puts the
+        // format in, but SPIRV-Cross throws rather than printing the formats it calls
+        // desktop-only when it targets ESSL - r8ui among them, which is what the stencil half of
+        // KHR-GL4x.packed_depth_stencil.stencil_texturing binds - and a throw loses the whole
+        // stage. So those formats stay out of the module and are spelled here instead, on the
+        // emitted text, where nothing can refuse them.
+        //
+        // Declarations that already carry a format are left exactly as they are, whoever wrote
+        // it. Must run before RemoveLayoutBinding, which is where an image's layout qualifier
+        // stops being safe to edit by hand.
+        String BakeImageFormatQualifiers(String glslCode, const UnorderedMap<String, String>& esslFormatByUniformName);
         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 +213,12 @@ 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

@@ -9,7 +9,9 @@
 #include "BackendObject_DirectVulkan.h"
 #include "MG_Backend/BackendObject.h"
 #include "DirectVulkan.h"
+#include "SubgroupSupportPolicy.h"
 #include "MG_State/GLState/FramebufferState/FramebufferObject.h"
+#include "MG_State/GLState/Core.h"
 #include "MG_State/GLState/TextureState/TextureState.h"
 #include "MG_Util/Classifiers/TextureEnumClassifier.h"
 #include "MG_Util/Converters/MGToGL/TextureEnumConverter.h"
@@ -383,6 +385,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         }
         UpdateDynamicBackendParameters();
         UpdateAdvertisedExtensions();
+        if (MG_State::pGLContext) {
+            MG_State::pGLContext->InvalidateCompileEnv();
+        }
         PopulateFormatCapabilities(physicalDevice.handle, vkGetPhysicalDeviceFormatProperties, m_vulkanCaps,
                                    MutableFormatCapabilities());
         PrintFormatCapabilities(GetFormatCapabilities());
@@ -497,30 +502,48 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             .ExtraVendor = Nullopt,
             .RendererGLInfo = {.TargetGLVersion = {4, 0, 0},
                                .TargetGLSLVersion = {4, 6, 0},
-                               // Baseline advertisement (no shader subgroup, no timer queries); a
+                               // Baseline advertisement (no runtime-gated capabilities); a live
-                               // live backend reconciles its copy in UpdateAdvertisedExtensions.
+                               // backend reconciles its copy in UpdateAdvertisedExtensions.
-                               .Extensions = BuildAdvertisedExtensions(false, false, false),
+                               .Extensions = BuildAdvertisedExtensions(false, false, false, false),
                                .IsCompatibilityProfile = false},
             .StaticBackendCapability = {.AllowVSOnlyPrograms = false}};
         return rendererInfo;
     }
 
     Vector<GLExtension> BuildAdvertisedExtensions(Bool shaderSubgroupSupported, Bool timerQueriesSupported,
-                                                  Bool anisotropicFilteringSupported) {
+                                                  Bool anisotropicFilteringSupported,
+                                                  Bool nonZeroIndirectBaseInstanceSupported) {
         Vector<GLExtension> extensions = {
             V_OpenGL30, V_OpenGL31, V_OpenGL32, V_OpenGL33, V_OpenGL40, E_GL_ARB_draw_buffers_blend,
             E_GL_ARB_compute_shader, E_GL_ARB_shader_storage_buffer_object, E_GL_ARB_shader_image_load_store,
-            E_GL_ARB_program_interface_query, E_GL_ARB_framebuffer_object, E_GL_ARB_multi_draw_indirect,
+            E_GL_ARB_clear_buffer_object, E_GL_ARB_program_interface_query, E_GL_ARB_framebuffer_object, E_GL_ARB_draw_indirect,
+            E_GL_ARB_multi_draw_indirect,
             E_GL_ARB_indirect_parameters, E_GL_EXT_framebuffer_object, E_GL_ARB_depth_texture, E_GL_ARB_buffer_storage,
             E_GL_ARB_texture_storage, E_GL_ARB_texture_storage_multisample, E_GL_ARB_texture_multisample,
             E_GL_ARB_clear_texture, E_GL_ARB_direct_state_access, E_GL_ARB_shader_draw_parameters,
             E_GL_ARB_gpu_shader_int64, E_GL_KHR_debug, E_GL_ARB_gpu_shader5, E_GL_ARB_multi_bind,
             E_GL_ARB_shading_language_420pack, E_GL_ARB_vertex_attrib_binding, E_GL_ARB_shader_image_size,
             E_GL_ARB_explicit_attrib_location,
+            // Core since GL 3.1 and implemented for every version advertised here. The string
+            // matters because applications gate the ENTRY POINTS on it rather than on the
+            // version: a caller that finds the extension missing never resolves
+            // glGetUniformBlockIndex / glUniformBlockBinding, and one that then uses uniform
+            // blocks anyway calls through a null pointer.
+            E_GL_ARB_uniform_buffer_object,
+            // Sampling the stencil aspect through DEPTH_STENCIL_TEXTURE_MODE. Core from 4.3,
+            // so on a 4.0 context the string is the only way to reach it.
+            E_GL_ARB_stencil_texturing,
             // Advertised with GL_NUM_PROGRAM_BINARY_FORMATS = 0, which the
             // extension explicitly permits. It is also the only thing that
             // exposes glProgramParameteri before GL 4.1.
             E_GL_ARB_get_program_binary};
+        // Vulkan's drawIndirectFirstInstance feature is optional. Direct base-instance calls work
+        // without it, but ARB_base_instance also promises non-zero firstInstance in GPU indirect
+        // commands; the renderer supplies true only when that word is legal and gl_InstanceID can
+        // be rebased to OpenGL's zero-based semantics.
+        if (nonZeroIndirectBaseInstanceSupported) {
+            extensions.push_back(E_GL_ARB_base_instance);
+        }
         if (shaderSubgroupSupported && !MG_Config::Features.DisableSubgroup) {
             extensions.push_back(E_GL_KHR_shader_subgroup);
         }
@@ -539,6 +562,15 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         if (MG_Util::Async::AsyncShaderCompileEnabled()) {
             extensions.push_back(E_GL_KHR_parallel_shader_compile);
         }
+        // GL_ARB_gpu_shader_fp64 is opt-in (MOBILEGL_ADVERTISE_FP64). Every `double` in a
+        // shader compiles and runs already - it is narrowed to 32 bits before the module
+        // reaches this backend - so an application that simply uses doubles needs nothing
+        // advertised. What the extension additionally promises is 64-bit PRECISION, which no
+        // mobile GPU has and the narrowing cannot fake, so advertising it by default would
+        // make an application that checks the string take a path MobileGL cannot honour.
+        if (MG_Config::Features.AdvertiseFp64) {
+            extensions.push_back(E_GL_ARB_gpu_shader_fp64);
+        }
         // GL_ARB_timer_query gates MC's F3 GPU% (LWJGL checks the extension string);
         // only advertised when the device actually supports timestamp queries and the
         // MOBILEGL_DISABLE_TIMERQUERY escape hatch is off.
@@ -660,6 +692,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         m_vulkanCaps = capabilities;
         UpdateDynamicBackendParameters();
         UpdateAdvertisedExtensions();
+        if (MG_State::pGLContext) {
+            MG_State::pGLContext->InvalidateCompileEnv();
+        }
         MutableFormatCapabilities().Clear();
     }
 
@@ -670,9 +705,16 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // real device timestamp support. ApplyVulkanCapabilitiesForTesting may
         // run without a renderer; no timer query is advertised then. Rebuilding
         // the whole list keeps re-runs idempotent.
+        // The opt-in emulated compute path (SubgroupSupportPolicy.h) carries the
+        // extension by itself on devices with no native subgroup support at all; a
+        // device with native subgroups always advertises - and uses - those.
+        const Bool subgroupSupportAdvertised =
+            m_vulkanCaps.SupportsShaderSubgroup ||
+            ShouldEmulateSubgroups(m_vulkanCaps.SupportsShaderSubgroup);
         m_rendererInfo.RendererGLInfo.Extensions = BuildAdvertisedExtensions(
-            m_vulkanCaps.SupportsShaderSubgroup, pVulkanRenderer && pVulkanRenderer->IsTimerQuerySupported(),
+            subgroupSupportAdvertised, pVulkanRenderer && pVulkanRenderer->IsTimerQuerySupported(),
-            pVulkanRenderer && pVulkanRenderer->IsSamplerAnisotropySupported());
+            pVulkanRenderer && pVulkanRenderer->IsSamplerAnisotropySupported(),
+            pVulkanRenderer && pVulkanRenderer->IsNonZeroIndirectBaseInstanceSupported());
     }
 
     void BackendObject_DirectVulkan::UpdateDynamicBackendParameters() {
@@ -768,14 +810,48 @@ 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;
+        // Vulkan descriptor limits are not GL limits, and a GL application reads an advertised
-        m_dynamicParameters.MaxCombinedShaderStorageBlocks = m_vulkanCaps.MaxCombinedShaderStorageBlocks;
+        // limit as an amount it may actually USE. Adreno answers the per-stage/per-set descriptor
-        m_dynamicParameters.MaxComputeUniformBlocks = m_vulkanCaps.MaxComputeUniformBlocks;
+        // 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.MaxShaderStorageBufferBindings =
-        m_dynamicParameters.MaxTextureBufferSize = m_vulkanCaps.MaxTextureBufferSize;
+            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);
@@ -843,7 +919,27 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                     DynParams::PerLayerFramebufferAttachmentBit(TextureTarget::TextureCubeMapArray);
             }
         }
-        m_dynamicParameters.SupportsFloat64VertexAttributes = m_vulkanCaps.SupportsShaderFloat64;
+        // Never, on any device, and no longer for the reason it used to be. It used to track
+        // shaderFloat64 because a `dvec3` input needed the Float64 capability to exist in the
+        // module at all; a 64-bit vertex FETCH was already impossible (VK_FORMAT_R64*_SFLOAT is
+        // optional and lavapipe reports zero bufferFeatures for all four), so the attribute
+        // arrived as its 32-bit word pair and PackDoubleVertexInputsPass bitcast it back.
+        //
+        // The shader half of that is gone: every 64-bit float is narrowed before any module
+        // reaches a backend (ShaderTranspiler::DemoteFloat64Pass), so there is no `double` input
+        // left to bitcast INTO, and feeding a UINT-formatted attribute to what is now a `float`
+        // input would be silent garbage. Reconstructing the value would mean decoding the
+        // IEEE-754 double bit pattern in the shader - software fp64, which is precisely what the
+        // demotion exists to avoid - and on Espryt it would additionally need the ES driver to
+        // fetch 2N uint components where the application declared N doubles, which a dvec3 or
+        // dvec4 cannot even express within one attribute location.
+        //
+        // So glVertexAttribLFormat / glVertexAttribLPointer are declined here exactly as they
+        // already were on Espryt and on every real mobile device (Adreno and Mali both report
+        // shaderFloat64 == VK_FALSE), and for the same visible reason. A `dvec3` INPUT still
+        // compiles and draws - it is a `vec3` after demotion - as long as the application feeds
+        // it with glVertexAttribPointer(GL_FLOAT) rather than 64-bit data.
+        m_dynamicParameters.SupportsFloat64VertexAttributes = false;
         m_dynamicParameters.MaxShaderStorageBlockSize =
             std::min(m_vulkanCaps.MaxShaderStorageBlockSize, kMaxAdvertisedShaderStorageBlockSize);
         if (m_vulkanCaps.SupportsShaderSubgroup) {
@@ -852,6 +948,18 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             m_dynamicParameters.SubgroupSupportedFeatures =
                 mapSubgroupFeatures(m_vulkanCaps.SubgroupSupportedOperations);
             m_dynamicParameters.SubgroupQuadOperationsInAllStages = m_vulkanCaps.SubgroupQuadOperationsInAllStages;
+        } else if (ShouldEmulateSubgroups(m_vulkanCaps.SupportsShaderSubgroup)) {
+            // MOBILEGL_MAGMA_EMULATE_SUBGROUP on a device with no native subgroups: the
+            // advertised values describe the 32-lane virtual subgroup the compute
+            // lowering implements (SubgroupSupportPolicy.h / EmulateSubgroupsPass).
+            // GL requires the advertisement and the execution to agree, and on this
+            // path the emulation is what executes; only the compute stage is offered.
+            m_dynamicParameters.SubgroupSize = kEmulatedSubgroupSize;
+            m_dynamicParameters.SubgroupSupportedStages = kEmulatedSubgroupStages;
+            m_dynamicParameters.SubgroupSupportedFeatures = kEmulatedSubgroupFeatures;
+            m_dynamicParameters.SubgroupQuadOperationsInAllStages = false;
+            MGLOG_I("DirectVulkan: emulating 32-lane compute subgroups "
+                    "(MOBILEGL_MAGMA_EMULATE_SUBGROUP, no native subgroup support)");
         } else {
             m_dynamicParameters.SubgroupSize = 0;
             m_dynamicParameters.SubgroupSupportedStages = 0;

MobileGL/MG_Backend/DirectVulkan/BackendObject_DirectVulkan.h

@@ -62,8 +62,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
     // POST screen shows.
 
     // Static identity of the Magma renderer (renderer/backend names, target GL/GLSL
-    // versions, ExtraVendor) with the baseline extension advertisement (no shader
+    // versions, ExtraVendor) with the baseline extension advertisement (no runtime-gated
-    // subgroup, no timer queries). A live backend copies this in its constructor and
+    // capabilities). A live backend copies this in its constructor and
     // reconciles the Extensions in UpdateAdvertisedExtensions once real capabilities
     // exist; callers that need the advertised list for a known capability set must
     // use BuildAdvertisedExtensions instead.
@@ -74,7 +74,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
     // MOBILEGL_DISABLE_TIMERQUERY escape hatches are applied inside, so callers pass
     // the detected device support (passing an already-gated value is harmless).
     Vector<GLExtension> BuildAdvertisedExtensions(Bool shaderSubgroupSupported, Bool timerQueriesSupported,
-                                                  Bool anisotropicFilteringSupported);
+                                                  Bool anisotropicFilteringSupported,
+                                                  Bool nonZeroIndirectBaseInstanceSupported);
 
     // Format: <GPU Name>, Vulkan <Vulkan Version>, Driver <Driver Version> — the exact
     // string an initialized backend returns from GetBackendAPIVersionString (and that

MobileGL/MG_Backend/DirectVulkan/DirectVulkan.cpp

@@ -69,6 +69,12 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             // slot's ownership unambiguous.
             Uint64 programLifetimeId = 0;
             Uint32 backendStateVersion = 0;
+            // glShaderStorageBlockBinding deliberately does NOT bump the backend state
+            // version, and the pipeline composite is unnamed so the in-place patch in
+            // DirectVulkan::ShaderStorageBlockBinding can never reach its slot - the
+            // mirror replay bumps only the program's block-binding version. Without this
+            // key the composite's slot kept serving the pre-rebind block.binding.
+            Uint32 blockBindingVersion = 0;
             Vector<StorageBlockResource> storageBlocks;
             Vector<BufferVariableResource> bufferVariables;
             GLint computeWorkGroupSize[3] = {1, 1, 1};
@@ -156,18 +162,33 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             auto& cache = g_programResourceCaches[program.GetExternalIndex()];
             const Uint64 programLifetimeId = program.GetLifetimeId();
             const Uint32 backendStateVersion = program.GetBackendStateVersion();
+            const Uint32 blockBindingVersion = program.GetBlockBindingVersion();
             // The lifetime id must match too: a new program that reuses a deleted
             // program's name and happens to land on the same backendStateVersion (both
             // count from zero) would otherwise be served the dead program's reflection.
             if (cache.programLifetimeId == programLifetimeId &&
                 cache.backendStateVersion == backendStateVersion &&
                 (!cache.storageBlocks.empty() || !cache.bufferVariables.empty())) {
+                if (cache.blockBindingVersion != blockBindingVersion) {
+                    // Only the block bindings moved (glShaderStorageBlockBinding, or the
+                    // pipeline composite's mirror replay - neither touches the backend
+                    // state version): the reflection itself is unchanged, so re-apply the
+                    // overrides by name instead of re-running spirv-reflect. Overrides
+                    // only ever accumulate, so a block without one still holds its
+                    // declared binding.
+                    for (auto& block : cache.storageBlocks) {
+                        const Int rebound = program.GetShaderStorageBlockBindingOverride(block.name);
+                        if (rebound >= 0) block.binding = static_cast<Uint32>(rebound);
+                    }
+                    cache.blockBindingVersion = blockBindingVersion;
+                }
                 return cache;
             }
 
             cache = {};
             cache.programLifetimeId = programLifetimeId;
             cache.backendStateVersion = backendStateVersion;
+            cache.blockBindingVersion = blockBindingVersion;
 
             Vector<SpvReflectShaderModule> modules;
             Vector<Bool> validModules;
@@ -269,14 +290,14 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 drawBuffer->SyncPersistentMappedRange();
                 const SizeT commandOffset = reinterpret_cast<SizeT>(indirect);
                 if (drawBuffer->MappedData() == nullptr || commandOffset + requiredBytes > drawBuffer->GetSize()) {
-                    MGLOG_E("%s skipped: invalid GL_DRAW_INDIRECT_BUFFER binding or range", label);
+                    MGLOG_E_ONCE("%s skipped: invalid GL_DRAW_INDIRECT_BUFFER binding or range", label);
                     return nullptr;
                 }
                 return drawBuffer->MappedData() + commandOffset;
             }
 
             if (!indirect) {
-                MGLOG_E("%s skipped: indirect pointer is null", label);
+                MGLOG_E_ONCE("%s skipped: indirect pointer is null", label);
                 return nullptr;
             }
 
@@ -398,7 +419,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             stride = sizeof(DrawArraysIndirectCommand);
         }
         if (stride < static_cast<GLsizei>(sizeof(DrawArraysIndirectCommand))) {
-            MGLOG_E("MultiDrawArraysIndirect skipped: stride %d is smaller than command size %zu",
+            MGLOG_E_ONCE("MultiDrawArraysIndirect skipped: stride %d is smaller than command size %zu",
                     stride, sizeof(DrawArraysIndirectCommand));
             return;
         }
@@ -446,20 +467,20 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             stride = sizeof(DrawArraysIndirectCommand);
         }
         if (stride < static_cast<GLsizei>(sizeof(DrawArraysIndirectCommand))) {
-            MGLOG_E("MultiDrawArraysIndirectCount skipped: stride %d is smaller than command size %zu",
+            MGLOG_E_ONCE("MultiDrawArraysIndirectCount skipped: stride %d is smaller than command size %zu",
                     stride, sizeof(DrawArraysIndirectCommand));
             return;
         }
 
         auto parameterBuffer = MG_State::pGLContext->GetBufferBindingSlot(BufferTarget::Parameter).GetBoundObject();
         if (!parameterBuffer || drawcount < 0 || static_cast<SizeT>(drawcount) + sizeof(Uint32) > parameterBuffer->GetSize()) {
-            MGLOG_E("MultiDrawArraysIndirectCount skipped: invalid GL_PARAMETER_BUFFER binding or range");
+            MGLOG_E_ONCE("MultiDrawArraysIndirectCount skipped: invalid GL_PARAMETER_BUFFER binding or range");
             return;
         }
 
         parameterBuffer->SyncPersistentMappedRange();
         if (parameterBuffer->MappedData() == nullptr) {
-            MGLOG_E("MultiDrawArraysIndirectCount skipped: CPU fallback cannot read parameter buffer");
+            MGLOG_E_ONCE("MultiDrawArraysIndirectCount skipped: CPU fallback cannot read parameter buffer");
             return;
         }
 
@@ -513,7 +534,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
         const SizeT indexSize = MG_Util::GetGLTypeSize(type);
         if (indexSize == 0) {
-            MGLOG_E("DrawElementsIndirect skipped: unsupported index type 0x%x", type);
+            MGLOG_E_ONCE("DrawElementsIndirect skipped: unsupported index type 0x%x", type);
             return;
         }
 
@@ -801,9 +822,18 @@ 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 +996,29 @@ 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;
@@ -977,7 +1030,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // shift - the hardware divide was the hottest instruction of this loop.
         const SizeT indexSize = MG_Util::GetGLTypeSize(type);
         if (indexSize == 0) {
-            MGLOG_E("MultiDrawElements skipped: unsupported index type 0x%x", type);
+            MGLOG_E_ONCE("MultiDrawElements skipped: unsupported index type 0x%x", type);
             return;
         }
         const Uint32 indexSizeShift = static_cast<Uint32>(std::countr_zero(indexSize));

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/FrameContext.cpp

@@ -205,7 +205,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // commands away. The device is gone on that path anyway - stay silent-safe
         // rather than trade a lost device for a barrier into a closed buffer.
         if (frame.hasCommandBufferRecorded) {
-            MGLOG_E("TransitionToPresent: command buffer already closed; skipping the present barrier");
+            MGLOG_E_ONCE("TransitionToPresent: command buffer already closed; skipping the present barrier");
             return false;
         }
 

MobileGL/MG_Backend/DirectVulkan/Renderer/PipelineFactory.cpp

@@ -206,6 +206,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         XXHASH_VERIFY(
             XXH64_update(m_hashState, &payload.primitiveRestartEnable, sizeof(payload.primitiveRestartEnable)));
         XXHASH_VERIFY(XXH64_update(m_hashState, &payload.patchControlPoints, sizeof(payload.patchControlPoints)));
+        XXHASH_VERIFY(XXH64_update(m_hashState, &payload.viewportCount, sizeof(payload.viewportCount)));
         XXHASH_VERIFY(XXH64_update(m_hashState, &payload.polygonMode, sizeof(payload.polygonMode)));
         XXHASH_VERIFY(XXH64_update(m_hashState, &payload.cullMode, sizeof(payload.cullMode)));
         XXHASH_VERIFY(XXH64_update(m_hashState, &payload.frontFace, sizeof(payload.frontFace)));
@@ -252,6 +253,23 @@ 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) {
+            // Unlatched, like the CreatePipeline report it accompanies: a pipeline MobileGL
+            // assembled and the driver refused is a broken invariant, not an expected failure,
+            // so it stays loud for as long as it is reachable. Raised from MGLOG_I once the
+            // Log.h ordering fix made MGLOG_E live in INFO builds.
+            MGLOG_E("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;
@@ -389,8 +407,12 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         tessellation.patchControlPoints = payload.patchControlPoints;
 
         VkPipelineViewportStateCreateInfo vpci{VK_STRUCTURE_TYPE_PIPELINE_VIEWPORT_STATE_CREATE_INFO};
-        vpci.viewportCount = 1;
+        // Both counts move together: GL has one scissor rectangle per viewport, and Vulkan
-        vpci.scissorCount = 1;
+        // requires viewportCount == scissorCount whenever both are dynamic
+        // (VUID-VkPipelineViewportStateCreateInfo-scissorCount-04136). The caller has already
+        // clamped this to the device's multiViewport capability.
+        vpci.viewportCount = std::max<Uint32>(payload.viewportCount, 1u);
+        vpci.scissorCount = vpci.viewportCount;
 
         VkPipelineRasterizationStateCreateInfo raster{VK_STRUCTURE_TYPE_PIPELINE_RASTERIZATION_STATE_CREATE_INFO};
         raster.polygonMode = payload.polygonMode;
@@ -458,9 +480,56 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         blend.attachmentCount = payload.colorAttachmentCount;
         blend.pAttachments = colorAttachments.empty() ? nullptr : colorAttachments.data();
 
+        // A GL program may have a tessellation EVALUATION stage and no CONTROL stage: GL 4.6 core
+        // 11.2.2 gives it a fixed-function pass-through instead. Vulkan has no such stage, and
+        // VUID-VkGraphicsPipelineCreateInfo-pStages-00730 requires both tessellation stages or
+        // neither - so the renderer synthesizes the pass-through GL describes and hands it in
+        // here (see ProgramFactory::GetOrCreatePassthroughTessControlStage).
+        //
+        // The refusal below is what keeps the half-tessellated shape away from the driver when
+        // there is no synthesized stage to add - because Mali does not reject it, it dereferences
+        // null INSIDE vkCreateGraphicsPipelines and takes the process down (SIGSEGV, fault addr
+        // 0x34, on Mali-G715/r54p2 and Mali-G925/r49p1 alike; Adreno and lavapipe merely render
+        // wrong). Returning VK_NULL_HANDLE routes this through the same path a driver rejection
+        // takes: the draw is skipped, nothing is memoised, and the process survives.
+        const Vector<VkPipelineShaderStageCreateInfo>* effectiveStages = payload.stages;
+        Vector<VkPipelineShaderStageCreateInfo> stagesWithPassthrough;
+        if (payload.passthroughTessControlStage.module != VK_NULL_HANDLE) {
+            stagesWithPassthrough = *payload.stages;
+            stagesWithPassthrough.push_back(payload.passthroughTessControlStage);
+            effectiveStages = &stagesWithPassthrough;
+        }
+        {
+            VkShaderStageFlags stagesPresent = 0;
+            for (const auto& stageInfo : *effectiveStages) {
+                stagesPresent |= stageInfo.stage;
+            }
+            const Bool hasTessControl = (stagesPresent & VK_SHADER_STAGE_TESSELLATION_CONTROL_BIT) != 0;
+            const Bool hasTessEval = (stagesPresent & VK_SHADER_STAGE_TESSELLATION_EVALUATION_BIT) != 0;
+            if (hasTessControl != hasTessEval) {
+                // Latched, and the latch is the point: a failed creation is deliberately never
+                // memoised (see GetOrCreatePipeline), so a program in this state re-enters here
+                // once per draw, every frame - and a refusal diagnostic that repeats per draw is
+                // noise, not a diagnostic. One line names the program; the draws it explains are
+                // all the same draw.
+                static Bool s_warnedHalfTessellatedPipeline = false;
+                if (!s_warnedHalfTessellatedPipeline) {
+                    s_warnedHalfTessellatedPipeline = true;
+                    MGLOG_E_ONCE("PipelineFactory::CreatePipeline: refusing a pipeline with %s tessellation stage and "
+                            "no %s stage (VUID-VkGraphicsPipelineCreateInfo-pStages-00730). programHash=0x%llx "
+                            "patchControlPoints=%u. Its draws are skipped; logged once.",
+                            hasTessEval ? "an evaluation" : "a control",
+                            hasTessEval ? "control" : "evaluation",
+                            static_cast<unsigned long long>(payload.programHash),
+                            payload.patchControlPoints);
+                }
+                return VK_NULL_HANDLE;
+            }
+        }
+
         VkGraphicsPipelineCreateInfo gpi{VK_STRUCTURE_TYPE_GRAPHICS_PIPELINE_CREATE_INFO};
-        gpi.stageCount = static_cast<Uint32>(payload.stages->size());
+        gpi.stageCount = static_cast<Uint32>(effectiveStages->size());
-        gpi.pStages = payload.stages->data();
+        gpi.pStages = effectiveStages->data();
         gpi.pVertexInputState = payload.vertexInputState;
         gpi.pInputAssemblyState = &ia;
         gpi.pTessellationState =
@@ -477,6 +546,13 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
         VkPipeline pipeline = VK_NULL_HANDLE;
         const VkResult result = vkCreateGraphicsPipelines(m_device, m_pipelineCache, 1, &gpi, nullptr, &pipeline);
+        // Loud, at MGLOG_F, and deliberately NOT latched. vkCreateGraphicsPipelines refusing a
+        // pipeline MobileGL assembled is a should-never-happen state, and the driver's own
+        // answer is VK_ERROR_UNKNOWN - no information at all - so this dump is the entire
+        // diagnosis. It is not an expected failure mode, so the one-shot rule that quiets W/E
+        // does not apply: while this is reachable it should keep saying so on every draw.
+        // GetOrCreatePipeline deliberately does not cache the failure, which is what makes that
+        // repetition happen; if the repetition ever needs to stop, fix the pipeline, not the log.
         if (result != VK_SUCCESS) {
             MGLOG_F("PipelineFactory::CreatePipeline failed: result=%s (%d) programHash=0x%llx vertexInputHash=0x%llx stageCount=%u topology=%s(%d) colorAttachmentCount=%u samples=%s(%d) subpass=%u",
                     VkResultToString(result),
@@ -507,6 +583,36 @@ 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: this is part of a
+            // should-never-happen report and must survive in the INFO-level builds that CTS
+            // actually runs against, alongside the MGLOG_F lines above.
+            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;
@@ -32,6 +42,14 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             Bool primitiveRestartEnable = false;
             // GL_PATCH_VERTICES; only read for a PATCH_LIST topology.
             Uint32 patchControlPoints = 3;
+            // How many of ARB_viewport_array's viewports this pipeline rasterizes into. 1 for
+            // every program that never assigns gl_ViewportIndex, which is all of them outside the
+            // conformance suite - the wide shape costs a longer vkCmdSetViewport/Scissor per state
+            // change and can cost hardware fast paths, so it is opt-in per program. Baked into the
+            // pipeline (viewportCount is not dynamic without VK_EXT_extended_dynamic_state) and
+            // therefore hashed; the DYNAMIC viewport/scissor arrays the draw pushes must have
+            // exactly this many elements (VUID-vkCmdDraw-viewportCount-03417/-03418).
+            Uint32 viewportCount = 1;
             VkPolygonMode polygonMode = VK_POLYGON_MODE_FILL;
             VkCullModeFlags cullMode = VK_CULL_MODE_BACK_BIT;
             VkFrontFace frontFace = VK_FRONT_FACE_CLOCKWISE;
@@ -61,7 +79,20 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             Bool fragmentReplacesDepth = false;
             Array<VkPipelineColorBlendAttachmentState, kMaxColorAttachments> colorBlendAttachments{};
             const Vector<VkPipelineShaderStageCreateInfo>* stages = nullptr;
+            // The tessellation control stage this renderer synthesized for a program that has
+            // an evaluation stage and none of its own (GL 4.6 core 11.2.2 gives such a program a
+            // fixed-function pass-through; Vulkan has no such thing and
+            // VUID-VkGraphicsPipelineCreateInfo-pStages-00730 forbids the half-tessellated
+            // pipeline outright). Appended to `stages` at creation. A null module means the
+            // renderer could not build one, and CreatePipeline refuses the pipeline - the same
+            // refusal it applies when `stages` itself is half-tessellated.
+            //
+            // NOT hashed: it is a pure function of the program and of patchControlPoints, both
+            // of which ComputeHash already mixes in.
+            VkPipelineShaderStageCreateInfo passthroughTessControlStage{};
             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.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"
@@ -32,7 +33,13 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             CombinedImageSampler,
             UniformTexelBuffer,
             StorageBuffer,
-            StorageImage
+            StorageImage,
+            // GLSL `imageBuffer` - a buffer texture reached through an IMAGE unit rather than a
+            // texture unit. Vulkan spells it VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER, which is a
+            // VkBufferView like UniformTexelBuffer and not a VkImageView like StorageImage: it is
+            // the one image uniform whose descriptor is a buffer. Appended, never inserted -
+            // DescriptorKeyHash mixes the enumerator's value.
+            StorageTexelBuffer
         };
 
         enum class CompileOptionBit : Uint {
@@ -52,19 +59,56 @@ 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,
+            // Replaces the vertex stage's gl_BaseVertex reads with zero. GL defines the builtin
+            // as zero for every drawing command that has no baseVertex parameter - all the
+            // DrawArrays forms - while Vulkan's BaseVertex reports firstVertex there. Set only
+            // for a non-indexed draw whose program actually reads the builtin, so nothing else
+            // acquires a second program/pipeline variant. See ZeroBaseVertexPass.
+            ZeroBaseVertex = 1 << 8,
         };
         using CompileOptionFlags = Flags<CompileOptionBit>;
         using HashType = Uint64;
 
+        struct UpdateAfterBindLimits {
+            Bool enabled = false;
+            Uint32 maxPerStageSamplers = 0;
+            Uint32 maxPerStageUniformBuffers = 0;
+            Uint32 maxPerStageStorageBuffers = 0;
+            Uint32 maxPerStageSampledImages = 0;
+            Uint32 maxPerStageStorageImages = 0;
+            Uint32 maxPerStageResources = 0;
+            Uint32 maxSetSamplers = 0;
+            Uint32 maxSetUniformBuffers = 0;
+            Uint32 maxSetUniformBuffersDynamic = 0;
+            Uint32 maxSetStorageBuffers = 0;
+            Uint32 maxSetStorageBuffersDynamic = 0;
+            Uint32 maxSetSampledImages = 0;
+            Uint32 maxSetStorageImages = 0;
+        };
+
         struct VkProgramObject {
             static constexpr Uint32 kMaxVertexInputLocations = 32;
 
             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;
+            // True only when this layout passed every descriptor-indexing feature and
+            // update-after-bind limit gate at reflection time. It controls both the
+            // layout/binding flags and the pool class used by UniformManager.
+            Bool usesUpdateAfterBind = false;
             VkPipelineLayout pipelineLayout = VK_NULL_HANDLE;
             Vector<DescriptorBindingKind> bindingKinds;
             // The bindings this program actually declares, ascending. bindingKinds is sized to the
@@ -75,8 +119,9 @@ 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
+            // Descriptor count per binding (1 except for a descriptor ARRAY - a UBO or storage
-            // binding with descriptorCount = N).
+            // 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.
@@ -85,6 +130,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             Vector<Int> samplerUniformLocationByBinding;
             Vector<TextureTarget> samplerTextureTargetByBinding;
             Vector<SamplerNumericDomain> samplerNumericDomainByBinding;
+            // Shared by StorageImage and StorageTexelBuffer bindings: a binding is one kind or
+            // the other, never both, and both need exactly the same thing - the format the
+            // shader declared, so the per-draw resolve can tell a typed declaration from a
+            // formatless one. Kept as one pair rather than two so the move operations below
+            // cannot drift out of sync with a field that only one kind populates.
             Vector<VkFormat> storageImageFormatByBinding;
             Vector<Bool> storageImageUsesBindingFormatByBinding;
             Vector<String> storageBlockNameByBinding;
@@ -92,6 +142,19 @@ 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{};
@@ -104,6 +167,33 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             // gl_FragDepth); shader-computed depth is immune to the cross-pipeline
             // position-invariance quirk (see PipelineFactory::ShouldSuppressDepthWrite).
             Bool fragmentReplacesDepth = false;
+            // The vertex module declares the BaseVertex builtin. Selects the ZeroBaseVertex
+            // program variant for non-indexed draws, and is deliberately a property of the
+            // PROGRAM rather than of the variant: the zeroed variant leaves the variable
+            // declared, so both variants answer the same and the draw path can ask either.
+            Bool readsBaseVertexBuiltin = false;
+            // Some pre-rasterization stage assigns gl_ViewportIndex. Its pipeline declares
+            // viewportCount = the renderer's rasterizable viewport count instead of 1, and its
+            // draws push the whole viewport/scissor array; every other program keeps the
+            // single-viewport fast path untouched. Part of the program's identity (folded into
+            // the pipeline hash through programHash), so no memo can serve the wrong shape.
+            Bool writesViewportIndexBuiltin = false;
+            // This program has a tessellation EVALUATION stage and no tessellation CONTROL
+            // stage. GL allows that (4.6 core 11.2.2: with no control shader the input patch
+            // is passed through unmodified, the output patch size is PATCH_VERTICES, and the
+            // levels come from the PATCH_DEFAULT_*_LEVEL state); Vulkan does not - either both
+            // tessellation stages are present or neither
+            // (VUID-VkGraphicsPipelineCreateInfo-pStages-00730). So the draw path has to supply
+            // the pass-through stage GL describes; see GetOrCreatePassthroughTessControlStage.
+            Bool needsPassthroughTessControl = false;
+            // ...and the pass-through this renderer can synthesize carries gl_Position and
+            // nothing else, so it is only correct when the evaluation stage's inputs are
+            // built-ins. A user-defined varying would arrive at the evaluation stage
+            // UNWRITTEN once a control stage sits between it and the vertex stage, which is
+            // silently wrong pixels rather than a crash - so those programs are declined
+            // instead (PipelineFactory::CreatePipeline refuses the pipeline and the draw is
+            // skipped). See ReflectPassthroughTessControlNeed.
+            Bool passthroughTessControlEmulatable = false;
             // Frame-boundary counter value of the last GetOrCreateProgram hit; drives
             // cache eviction (see OnFrameBoundary). Mutable: the draw snapshot's memoised
             // entry pointer re-stamps use through a const reference (StampProgramUse).
@@ -118,7 +208,16 @@ 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;
+                usesUpdateAfterBind = other.usesUpdateAfterBind;
                 pipelineLayout = other.pipelineLayout;
                 bindingKinds = std::move(other.bindingKinds);
                 activeBindings = std::move(other.activeBindings);
@@ -136,6 +235,7 @@ 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;
@@ -145,11 +245,17 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 producerOutputComponentCount = other.producerOutputComponentCount;
                 fragmentInputComponentCount = other.fragmentInputComponentCount;
                 fragmentReplacesDepth = other.fragmentReplacesDepth;
+                readsBaseVertexBuiltin = other.readsBaseVertexBuiltin;
+                writesViewportIndexBuiltin = other.writesViewportIndexBuiltin;
+                needsPassthroughTessControl = other.needsPassthroughTessControl;
+                passthroughTessControlEmulatable = other.passthroughTessControlEmulatable;
                 lastUsedFrame = other.lastUsedFrame;
                 other.hash = 0;
                 other.descriptorSetLayout = VK_NULL_HANDLE;
+                other.usesUpdateAfterBind = false;
                 other.pipelineLayout = VK_NULL_HANDLE;
                 other.hasStorageImages = false;
+                other.declinedDescriptors = false;
                 other.globalUboBinding = -1;
                 other.activeVertexInputLocationMask = 0;
                 other.activeFragmentOutputLocationMask = 0;
@@ -157,6 +263,10 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 other.producerOutputComponentCount = 0;
                 other.fragmentInputComponentCount = 0;
                 other.fragmentReplacesDepth = false;
+                other.readsBaseVertexBuiltin = false;
+                other.writesViewportIndexBuiltin = false;
+                other.needsPassthroughTessControl = false;
+                other.passthroughTessControlEmulatable = false;
                 other.lastUsedFrame = 0;
             }
             VkProgramObject& operator=(VkProgramObject&& other) noexcept {
@@ -167,7 +277,9 @@ 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;
+                usesUpdateAfterBind = other.usesUpdateAfterBind;
                 pipelineLayout = other.pipelineLayout;
                 bindingKinds = std::move(other.bindingKinds);
                 activeBindings = std::move(other.activeBindings);
@@ -185,6 +297,7 @@ 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;
@@ -194,11 +307,17 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 producerOutputComponentCount = other.producerOutputComponentCount;
                 fragmentInputComponentCount = other.fragmentInputComponentCount;
                 fragmentReplacesDepth = other.fragmentReplacesDepth;
+                readsBaseVertexBuiltin = other.readsBaseVertexBuiltin;
+                writesViewportIndexBuiltin = other.writesViewportIndexBuiltin;
+                needsPassthroughTessControl = other.needsPassthroughTessControl;
+                passthroughTessControlEmulatable = other.passthroughTessControlEmulatable;
                 lastUsedFrame = other.lastUsedFrame;
                 other.hash = 0;
                 other.descriptorSetLayout = VK_NULL_HANDLE;
+                other.usesUpdateAfterBind = false;
                 other.pipelineLayout = VK_NULL_HANDLE;
                 other.hasStorageImages = false;
+                other.declinedDescriptors = false;
                 other.globalUboBinding = -1;
                 other.activeVertexInputLocationMask = 0;
                 other.activeFragmentOutputLocationMask = 0;
@@ -206,6 +325,10 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 other.producerOutputComponentCount = 0;
                 other.fragmentInputComponentCount = 0;
                 other.fragmentReplacesDepth = false;
+                other.readsBaseVertexBuiltin = false;
+                other.writesViewportIndexBuiltin = false;
+                other.needsPassthroughTessControl = false;
+                other.passthroughTessControlEmulatable = false;
                 other.lastUsedFrame = 0;
                 return *this;
             }
@@ -233,6 +356,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 }
                 modules.clear();
                 stages.clear();
+                stageSpirvDigests.clear(); // the modules they describe are gone
             }
         };
 
@@ -248,21 +372,62 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             virtual void OnProgramEvicted(HashType programHash, VkDescriptorSetLayout descriptorSetLayout) = 0;
         };
 
-        explicit ProgramFactory(VkDevice device, const VulkanRendererConfig& config, Uint32 maxBindings = 16,
+        // How this factory's compute modules implement GL_KHR_shader_subgroup. Computed
-                                Bool shaderDrawParametersEnabled = false,
+        // once at renderer initialization (SubgroupSupportPolicy.h + the device's
-                                Bool unformattedFloatStorageImagesEnabled = false)
+        // subgroup properties) so lowering can never disagree with the advertised
+        // capabilities. Native subgroup operations always execute natively; the two
+        // repair passes patch modules AROUND them, and the emulation only replaces them
+        // on opted-in devices with no subgroup support at all.
+        struct SubgroupLoweringPolicy {
+            Bool emulateSubgroups = false;      // MOBILEGL_MAGMA_EMULATE_SUBGROUP, no-native-support devices
+            Bool fixIterationRPSubgroupScratch = false; // patch iterationRP's under-declared scratch
+            Bool fixIterationRPBarrier = false; // repair Program 203's shared-scratch race
+            Bool deriveNumSubgroups = false;    // repair the NumSubgroups builtin
+            Bool requireFullSubgroups = false;  // computeFullSubgroups enabled on the device
+            Uint32 nativeSubgroupSize = 0;
+            // Full-subgroup launches are bounded by this device limit; a dispatch whose
+            // workgroup needs more subgroups than this cannot request the flag.
+            Uint32 maxComputeWorkgroupSubgroups = 0;
+            // VkPhysicalDeviceLimits::maxComputeSharedMemorySize; bounds the scratch the
+            // emulation pass may add (0 falls back to the Vulkan minimum, 16384).
+            Uint32 maxComputeSharedMemoryBytes = 0;
+        };
+
+        explicit ProgramFactory(VkDevice device, const VulkanRendererConfig& config, Uint32 maxBindings,
+                                Bool shaderDrawParametersEnabled,
+                                Bool unformattedFloatStorageImagesEnabled,
+                                Bool enableSpirvValidation,
+                                UpdateAfterBindLimits updateAfterBindLimits,
+                                SubgroupLoweringPolicy subgroupPolicy)
             : m_device(device), m_maxBindings(maxBindings), m_config(config),
               m_shaderDrawParametersEnabled(shaderDrawParametersEnabled),
-              m_unformattedFloatStorageImagesEnabled(unformattedFloatStorageImagesEnabled) {
+              m_unformattedFloatStorageImagesEnabled(unformattedFloatStorageImagesEnabled),
+              m_enableSpirvValidation(enableSpirvValidation),
+              m_updateAfterBindLimits(updateAfterBindLimits),
+              m_subgroupPolicy(subgroupPolicy) {
             VkProgramObject::s_device = device;
         }
-        ~ProgramFactory() = default;
+        // Destroys the pass-through tessellation control modules. Runs while the device is
+        // still alive for the same reason ~VkProgramObject's does: this factory outlives
+        // nothing that owns the device.
+        ~ProgramFactory();
         ProgramFactory(const ProgramFactory&) = delete;
 
         HashType ComputeHash(const MG_State::GLState::ProgramObject& program, CompileOptionFlags flags) const;
         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
@@ -291,6 +456,39 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // True when an entry point reads the InstanceIndex builtin. Only gates a diagnostic:
         // without shaderDrawParameters such a shader cannot have gl_InstanceID rebased.
         static Bool ReflectedReadsInstanceIndexBuiltin(const SpvReflectShaderModule& reflectModule);
+        // True when an entry point declares the BaseVertex builtin, i.e. when a non-indexed
+        // draw with this program has to take the ZeroBaseVertex variant.
+        static Bool ReflectedReadsBaseVertexBuiltin(const SpvReflectShaderModule& reflectModule);
+        // Shared by the two above: does any entry point list an input variable decorated with
+        // this builtin?
+        static Bool ReflectedDeclaresInputBuiltin(const SpvReflectShaderModule& reflectModule, SpvBuiltIn builtin);
+        // True when an entry point writes the ViewportIndex builtin (gl_ViewportIndex), i.e. when
+        // the program can route primitives to a viewport other than 0 and its pipeline therefore
+        // has to declare more than one. Asks about OUTPUT variables because that is the direction
+        // a pre-rasterization stage declares it in.
+        static Bool ReflectedWritesViewportIndexBuiltin(const SpvReflectShaderModule& reflectModule);
+        static Bool ReflectedDeclaresOutputBuiltin(const SpvReflectShaderModule& reflectModule, SpvBuiltIn builtin);
+
+        // The pass-through tessellation control stage GL 4.6 core 11.2.2 describes for a
+        // program that has an evaluation stage and no control stage, for an input patch of
+        // `patchVertices` control points. Returned BY VALUE (a stage description is a POD, and
+        // the cache below is a rehashing map, so a pointer into it would not survive the next
+        // distinct patch size). `.module == VK_NULL_HANDLE` means the stage could not be built:
+        // the caller then has no control stage to inject, and CreatePipeline refuses the
+        // pipeline rather than handing the driver a half-tessellated one.
+        //
+        // Keyed on the patch size because GL takes the output patch size from PATCH_VERTICES,
+        // which is draw state, not link state - the CTS case that motivated this links at the
+        // default 3 and draws at 4. The pipeline cache already re-keys on patchControlPoints,
+        // so the module a pipeline was built with is part of that pipeline's identity.
+        // Compiling is bounded by the number of distinct patch sizes a program draws with
+        // (MAX_PATCH_VERTICES = 32 in the worst case, one or two in practice) and only ever
+        // happens for the rare program that has no control stage at all.
+        VkPipelineShaderStageCreateInfo GetOrCreatePassthroughTessControlStage(Uint32 patchVertices);
+
+        // Source of the module above. Exposed for tests: the generated GLSL is the whole
+        // contract with the evaluation stage, so it is worth pinning independently of a device.
+        static String BuildPassthroughTessControlSource(Uint32 patchVertices);
 
     private:
         struct ProgramLookupCache {
@@ -304,11 +502,20 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         void ReflectVertexInputs(const Vector<SharedPtr<MG_State::GLState::ShaderObject>>& shaders,
                      const Vector<Vector<Uint>>& spirv,
                      VkProgramObject& entry) const;
+        void ReflectViewportIndexUsage(const Vector<SharedPtr<MG_State::GLState::ShaderObject>>& shaders,
+                                       const Vector<Vector<Uint>>& spirv,
+                                       VkProgramObject& entry) const;
         void ReflectFragmentOutputs(const Vector<SharedPtr<MG_State::GLState::ShaderObject>>& shaders,
                         const Vector<Vector<Uint>>& spirv,
                         VkProgramObject& entry) const;
         void ReflectLayout(const MG_State::GLState::ProgramObject& program, const Vector<Vector<Uint>>& spirv,
                            VkProgramObject& entry) const;
+        // Fills needsPassthroughTessControl / passthroughTessControlEmulatable off the linked
+        // modules. Const and reflection-only: it decides nothing about the pipeline, it only
+        // records what the evaluation stage's input interface is made of.
+        void ReflectPassthroughTessControlNeed(const Vector<SharedPtr<MG_State::GLState::ShaderObject>>& shaders,
+                                               const Vector<Vector<Uint>>& spirv,
+                                               VkProgramObject& entry) const;
 
         VkDevice m_device = VK_NULL_HANDLE;
         Uint32 m_maxBindings = 0;
@@ -320,12 +527,28 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // True only when the logical device enabled both
         // shaderStorageImageReadWithoutFormat and shaderStorageImageWriteWithoutFormat.
         Bool m_unformattedFloatStorageImagesEnabled = false;
+        // Startup snapshot used only by internally synthesized shader modules, which do not
+        // originate from a ProgramLinkTask.
+        Bool m_enableSpirvValidation = false;
+        // Device feature and limit gate resolved before vkCreateDevice. Keeping it in
+        // the factory lets each reflected layout choose ordinary descriptors when its
+        // own counts would exceed the update-after-bind budget.
+        UpdateAfterBindLimits m_updateAfterBindLimits{};
+        SubgroupLoweringPolicy m_subgroupPolicy{};
+        // 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;
         // See GetCacheStructureEpoch(). Starts at 1 so a zero-initialized memo can never match.
         Uint64 m_cacheStructureEpoch = 1;
         IEvictionObserver* m_evictionObserver = nullptr;
+        // Pass-through tessellation control stages by input patch size. Never evicted: at most
+        // MAX_PATCH_VERTICES entries exist for the lifetime of the device, and every pipeline
+        // ever built from one keeps referencing its module. A failed build is cached as
+        // VK_NULL_HANDLE so a broken generator costs one compile, not one per draw.
+        UnorderedMap<Uint32, VkPipelineShaderStageCreateInfo> m_passthroughTessControlStages;
         static inline XXH64_state_t* m_hashState = XXH64_createState();
     };
 } // namespace MobileGL::MG_Backend::DirectVulkan

MobileGL/MG_Backend/DirectVulkan/Renderer/SwapchainObject.cpp

@@ -157,7 +157,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
         MGLOG_I("Got %d surface formats:", swapchainCapabilities.surfaceFormats.size());
         for (const auto& sf : swapchainCapabilities.surfaceFormats) {
-            MGLOG_I("    [%s, %s]", string_VkFormat(sf.format), string_VkColorSpaceKHR(sf.colorSpace));
+            MGLOG_D("    [%s, %s]", string_VkFormat(sf.format), string_VkColorSpaceKHR(sf.colorSpace));
         }
 
         const auto pickedSurfaceFormat = ChooseSwapchainSurfaceFormat(swapchainCapabilities.surfaceFormats);
@@ -166,7 +166,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
         MGLOG_I("Got %d present modes:", swapchainCapabilities.presentModes.size());
         for (const auto& pm : swapchainCapabilities.presentModes) {
-            MGLOG_I("    %s", string_VkPresentModeKHR(pm));
+            MGLOG_D("    %s", string_VkPresentModeKHR(pm));
         }
 
         const auto presentMode = ChooseSwapchainPresentMode(swapchainCapabilities.presentModes);

MobileGL/MG_Backend/DirectVulkan/Renderer/UniformManager.h

@@ -26,9 +26,20 @@ namespace MobileGL::MG_Backend::DirectVulkan {
     public:
         struct SamplerBindingOverride {
             Uint32 binding = 0;
+            Uint32 element = 0;
             MG_State::GLState::ITextureObject* texture = nullptr;
             const MG_State::GLState::SamplerObject* sampler = nullptr;
             VkImageView imageView = VK_NULL_HANDLE;
+            VkImageLayout imageLayout = VK_IMAGE_LAYOUT_UNDEFINED;
+            Bool forceNearestFiltering = false;
+        };
+
+        struct SamplerImageFeedbackBinding {
+            Uint32 samplerBinding = 0;
+            Uint32 samplerElement = 0;
+            MG_State::GLState::ITextureObject* texture = nullptr;
+            const MG_State::GLState::SamplerObject* sampler = nullptr;
+            SamplerNumericDomain numericDomain = SamplerNumericDomain::Unknown;
         };
 
         Bool Initialize(VkDevice device, VkBufferManager* bufferManager,
@@ -53,10 +64,13 @@ 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,
+        // One record per visited CombinedImageSampler DESCRIPTOR (post fallback substitution,
-        // in binding order): the resolved texture and effective sampler, as never-reused
+        // in binding order, and within a binding in array-element order): the resolved texture
-        // lifetime ids so a freed-and-reallocated object at the same heap address can only
+        // and effective sampler, as never-reused lifetime ids so a freed-and-reallocated object
-        // MISS a comparison, never false-hit it (same ABA rule as SamplerResolveMemo).
+        // 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;
@@ -76,6 +90,12 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         Bool CollectStorageImageTextures(const MG_State::GLState::ProgramObject& program,
                                          const ProgramFactory::VkProgramObject& programObj,
                                          Vector<MG_State::GLState::ITextureObject*>& outTextures) const;
+        Bool CollectSamplerImageFeedback(
+            const MG_State::GLState::ProgramObject& program,
+            const ProgramFactory::VkProgramObject& programObj,
+            Vector<SamplerImageFeedbackBinding>& outBindings) const;
+        static Bool SamplerOverlapsWritableImageSubresource(Int samplerBaseLevel, Int samplerMaxLevel,
+                                                             GLint imageLevel, GLenum imageAccess);
         // samplerDescriptorsUnchangedHint: the caller (SetupDraw fast path) proved that
         // every input of every combined-image-sampler resolution is unchanged since the
         // previous draw's resolve - same (texture, sampler) per binding, texture params
@@ -88,7 +108,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                                        Uint32 frameIndex,
                                        VkPipelineBindPoint bindPoint = VK_PIPELINE_BIND_POINT_GRAPHICS,
                                        const SamplerBindingOverride* samplerBindingOverride = nullptr,
-                                       Bool samplerDescriptorsUnchangedHint = false);
+                                       Bool samplerDescriptorsUnchangedHint = false,
+                                       const Vector<SamplerBindingOverride>* samplerBindingOverrides = nullptr);
 
         // Pure format-policy helper kept public for host regression tests. Formatted storage
         // images use their shader qualifier; transformed float images use glBindImageTexture's
@@ -111,6 +132,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             VkDescriptorPool handle = VK_NULL_HANDLE;
             Uint32 maxSets = 0;
             Uint32 allocatedSets = 0;
+            Bool updateAfterBind = false;
         };
 
         // A cached descriptor set together with the pool it was allocated from, so a
@@ -143,8 +165,9 @@ 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 +175,45 @@ 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);
+        // GLSL `imageBuffer`: the same VkBufferView descriptor as the sampled texel buffer above,
+        // but resolved from an IMAGE unit (glBindImageTexture) rather than a texture unit, and
+        // made GPU-resident-writable because the shader may store to it. No `element` parameter:
+        // an imageBuffer ARRAY is refused at program creation, so a binding is always one
+        // descriptor (see the array gate in RemapDescriptorBindingsForVulkan).
+        Bool ResolveStorageTexelBufferDescriptor(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 {
@@ -201,8 +242,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         void BindDescriptorSetDeduped(VkCommandBuffer commandBuffer, VkPipelineBindPoint bindPoint,
                                       VkPipelineLayout pipelineLayout, VkDescriptorSet descriptorSet,
                                       const Vector<Uint32>& dynamicOffsets);
-        Bool CreateDescriptorPool(Uint32 maxSets, VkDescriptorPool& outPool) const;
+        Bool CreateDescriptorPool(Uint32 maxSets, Bool updateAfterBind, VkDescriptorPool& outPool) const;
-        Bool GrowFrameDescriptorPool(FrameResources& frame, Uint32 frameIndex);
+        Bool GrowFrameDescriptorPool(FrameResources& frame, Uint32 frameIndex, Bool updateAfterBind);
         VkResult AllocateDescriptorSetsFromActivePool(
             Uint32 frameIndex, const ProgramFactory::VkProgramObject& programObj, VkDescriptorSet& outDescriptorSet);
         VkResult AcquireDescriptorSet(Uint32 frameIndex,
@@ -319,8 +360,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // lifetime id, so a freed-and-reallocated sampler or texture at the same heap address
         // always gets a fresh id and misses (a raw pointer would false-hit that ABA) - so a
         // stale guess can only miss and fall through to the hash, never resolve wrong. Still
-        // reset each frame alongside the descriptor-set cache. Indexed by binding.
+        // reset each frame alongside the descriptor-set cache. Indexed by binding, but the
+        // whole-descriptor entry is additionally keyed by program lifetime: Vulkan binding
+        // numbers are layout-local and unrelated programs routinely reuse binding 0/1.
         struct SamplerResolveMemo {
+            Uint64 infoProgramLifetimeId = 0;
             Uint64 samplerLifetimeId = 0;
             Uint64 textureLifetimeId = 0;
             VkSampler sampler = VK_NULL_HANDLE;
@@ -341,6 +385,14 @@ 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.cpp

@@ -110,7 +110,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             const VkFormat sourceVkFormat =
                 ToVkVertexFormat(attr.Type, attr.Size, attr.Normalized, attr.IsInteger, attr.IsBgra, attr.IsLong);
             if (sourceVkFormat == VK_FORMAT_UNDEFINED) {
-                MGLOG_E("Unsupported vertex attribute layout (location=%u, type=%s, size=%d): the array is "
+                MGLOG_E_ONCE("Unsupported vertex attribute layout (location=%u, type=%s, size=%d): the array is "
                         "enabled but cannot be mapped to a VkFormat",
                         location, MG_Util::ConvertDataTypeToString(attr.Type).c_str(), attr.Size);
                 unsupportedAttribMask |= (1u << location);
@@ -125,7 +125,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                     if (fallbackFormat != VK_FORMAT_UNDEFINED && SupportsVertexBufferFormat(fallbackFormat)) {
                         vkFormat = fallbackFormat;
                         conversion = VertexStreamConversion::ScaledIntegerToFloat32;
-                        MGLOG_W("Vertex attribute location=%u format=%d lacks "
+                        MGLOG_W_ONCE("Vertex attribute location=%u format=%d lacks "
                                 "VK_FORMAT_FEATURE_VERTEX_BUFFER_BIT; using float32 stream format=%d "
                                 "(type=%s size=%d normalized=%s integer=%s)",
                                 location, static_cast<Int>(sourceVkFormat), static_cast<Int>(vkFormat),
@@ -135,7 +135,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 }
 
                 if (conversion == VertexStreamConversion::None) {
-                    MGLOG_E("Unsupported Vulkan vertex format (location=%u, format=%d, type=%s, size=%d): "
+                    MGLOG_E_ONCE("Unsupported Vulkan vertex format (location=%u, format=%d, type=%s, size=%d): "
                             "VK_FORMAT_FEATURE_VERTEX_BUFFER_BIT is unavailable and no semantic fallback exists",
                             location, static_cast<Int>(sourceVkFormat),
                             MG_Util::ConvertDataTypeToString(attr.Type).c_str(), attr.Size);
@@ -146,15 +146,21 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
             const SizeT attribByteSize = GetAttributeByteSize(attr.Type, attr.Size, attr.IsBgra);
             if (attribByteSize == 0) {
-                MGLOG_E("Vertex attribute with unknown component size (location=%u, type=%s): the array is "
+                MGLOG_E_ONCE("Vertex attribute with unknown component size (location=%u, type=%s): the array is "
                         "enabled but cannot be sized",
                         location, MG_Util::ConvertDataTypeToString(attr.Type).c_str());
                 unsupportedAttribMask |= (1u << location);
                 continue;
             }
 
-            const Uint32 sourceStride =
+            // Verbatim, zero included. The frontend already resolved a pointer call's
-                attr.Stride > 0 ? static_cast<Uint32>(attr.Stride) : static_cast<Uint32>(attribByteSize);
+            // "tightly packed" stride 0 into the element size (see VertexAttribute::Stride),
+            // so a zero here is the binding model's stride 0 - every vertex reads the same
+            // element - which is exactly what a zero VkVertexInputBindingDescription::stride
+            // means. Substituting the element size fetched a fresh element per vertex and ran
+            // off the end of the buffer (KHR-GL43.vertex_attrib_binding.basic-input-case7/8).
+            // Client-memory arrays cannot reach zero: they only exist on the pointer path.
+            const Uint32 sourceStride = static_cast<Uint32>(attr.Stride);
             const Bool packedAttribute = attr.Type == DataType::Int2101010Rev ||
                                          attr.Type == DataType::Uint2101010Rev;
             const SizeT requiredAlignment = packedAttribute ? attribByteSize : GetComponentSize(attr.Type);
@@ -169,17 +175,23 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 // unless VK_EXT_legacy_vertex_attributes is available, so deinterleave this one
                 // attribute into a tightly packed transient stream without changing its format.
                 conversion = VertexStreamConversion::Repack;
-                MGLOG_W("Vertex attribute location=%u uses Vulkan-incompatible alignment "
+                MGLOG_W_ONCE("Vertex attribute location=%u uses Vulkan-incompatible alignment "
                         "(offset=%zu stride=%u required=%zu); using a tightly packed stream",
                         location, attr.Offset, sourceStride, requiredAlignment);
             }
 
             Uint32 stride = sourceStride;
+            // A converted stream is tightly packed, so its stride is the converted element
+            // size - unless the source stride is zero, which does not describe a packing at
+            // all but "never advance". That survives the conversion unchanged: the draw path
+            // converts exactly one element and every vertex reads it.
+            if (sourceStride != 0) {
                 if (conversion == VertexStreamConversion::Repack) {
                     stride = static_cast<Uint32>(attribByteSize);
                 } else if (conversion == VertexStreamConversion::ScaledIntegerToFloat32) {
                     stride = static_cast<Uint32>(attr.Size * static_cast<Int>(sizeof(Float)));
                 }
+            }
             const VkVertexInputRate inputRate =
                 (attr.Divisor == 0) ? VK_VERTEX_INPUT_RATE_VERTEX : VK_VERTEX_INPUT_RATE_INSTANCE;
 
@@ -275,8 +287,10 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             if (m_frameBoundaryCounter - it->second->lastUsedFrameBoundary > kRetireAgeBoundaries) {
                 it = m_cache.erase(it);
                 // Invalidate every VAO's state-pointer memo: the erased node's
-                // address may be reused by a future insert.
+                // address may be reused by a future insert. Advance through the
-                ++m_evictionEpoch;
+                // process-wide source so the value stays unique across factory
+                // instances (see the member comment).
+                m_evictionEpoch = ++s_evictionEpochSource;
             } else {
                 ++it;
             }

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,
+        // Values are heap-allocated: UnorderedMap is open-addressing, so INSERT
-        // so INSERT invalidates references to stored values. The draw path (and
+        // invalidates references to stored values - and so does ERASE, which shifts
-        // the VAOs' state-pointer memos) hold entry pointers across inserts;
+        // the rest of the probe cluster into the hole and therefore moves entries
-        // only the unique_ptr cell moves, never the pointee.
+        // 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;
@@ -123,7 +125,17 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // construction); a memo is honored only while its recorded epoch
         // matches, so an evicted entry can never be dereferenced through a
         // stale memo.
-        Uint64 m_evictionEpoch = 1;
+        //
+        // Drawn from a process-wide source, never a per-instance counter: the VAO
+        // memos outlive this factory (they live on pGLContext's VAOs, the renderer
+        // is destroyed and recreated on EGL surface release/re-create), so a fresh
+        // factory restarting at a dead factory's epoch value would honor its
+        // dangling entry pointers. The constructor takes a value strictly greater
+        // than anything a predecessor ever stamped, so a dead factory's memo can
+        // never compare equal here - the same never-reused idiom as the lifetime ids.
+        // Single-threaded like the rest of the factory (renderer-thread only).
+        static inline Uint64 s_evictionEpochSource = 0;
+        Uint64 m_evictionEpoch = ++s_evictionEpochSource;
         static inline XXH64_state_t* m_hashState = XXH64_createState();
     };
 } // namespace MobileGL::MG_Backend::DirectVulkan

MobileGL/MG_Backend/DirectVulkan/Renderer/VkBufferManager.cpp

@@ -23,7 +23,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             VK_BUFFER_USAGE_VERTEX_BUFFER_BIT | VK_BUFFER_USAGE_INDEX_BUFFER_BIT |
             VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT | VK_BUFFER_USAGE_STORAGE_BUFFER_BIT |
             VK_BUFFER_USAGE_INDIRECT_BUFFER_BIT | VK_BUFFER_USAGE_UNIFORM_TEXEL_BUFFER_BIT |
-            VK_BUFFER_USAGE_TRANSFER_SRC_BIT;
+            // "Every usage" has to mean every usage: a buffer texture reached through an IMAGE
+            // unit takes a VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER descriptor, and the write is
+            // invalid unless the buffer was created with this bit. Nothing asked for it until
+            // imageBuffer support existed, so the omission was invisible.
+            VK_BUFFER_USAGE_STORAGE_TEXEL_BUFFER_BIT | VK_BUFFER_USAGE_TRANSFER_SRC_BIT;
         // Appended to kPersistentBackedUsage when VK_EXT_transform_feedback is enabled
         // (see VkBufferManagerInitInfo::transformFeedbackUsageEnabled).
         constexpr VkBufferUsageFlags kTransformFeedbackUsage =
@@ -161,12 +165,23 @@ 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() {
@@ -287,7 +302,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             .requiredFlags = requiredFlags,
         });
         if (!created || resource.buffer.Map() == nullptr) {
-            MGLOG_E("VkBufferManager::CreateResidentStorage failed (size=%llu)",
+            MGLOG_E_ONCE("VkBufferManager::CreateResidentStorage failed (size=%llu)",
                     static_cast<unsigned long long>(size));
             resource.buffer.Destroy();
             resource.storageSize = 0;
@@ -309,7 +324,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             return false;
         }
         if (!resource.buffer.Upload(bufferObject.MappedData(), size, 0)) {
-            MGLOG_E("VkBufferManager::SwapStorageAndUploadAll: upload failed");
+            MGLOG_E_ONCE("VkBufferManager::SwapStorageAndUploadAll: upload failed");
             resource.pendingFullUpload = true;
             return false;
         }
@@ -368,6 +383,12 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         BumpSliceEpoch(*resource);
         // Any cached streaming slice refers to the previous contents.
         resource->transientFrameSerial = 0;
+        // Redefining the store hands any adopted mapping back to the CPU shadow
+        // (BufferObject::RedefineStorage), so a buffer that reaches here persistent-mapped
+        // is an ordinary resident one again: it needs the busy-tracking and conditional
+        // orphan below, and the next AcquirePersistentMap has to mint storage for the new
+        // store rather than hand back a mapping of the old one.
+        resource->persistentMapped = false;
         if (!resource->buffer.IsValid()) {
             return; // streaming-only resource: shadow + serial are enough
         }
@@ -388,7 +409,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         }
 
         if (!resource->buffer.Upload(bufferObject.MappedData(), size, 0)) {
-            MGLOG_E("VkBufferManager::OnRespecify: in-place upload failed");
+            MGLOG_E_ONCE("VkBufferManager::OnRespecify: in-place upload failed");
             resource->pendingFullUpload = true;
         }
     }
@@ -413,7 +434,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         if (!IsResourceBusy(*resource)) {
             if (!resource->buffer.Upload(bufferObject.MappedData() + offset,
                                          static_cast<VkDeviceSize>(size), static_cast<VkDeviceSize>(offset))) {
-                MGLOG_E("VkBufferManager::OnSubData: host upload failed");
+                MGLOG_E_ONCE("VkBufferManager::OnSubData: host upload failed");
                 resource->pendingFullUpload = true;
             }
             return;
@@ -450,7 +471,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         if ((appAccess & BufferMappingAccessBit::Unsynchronized) || !IsResourceBusy(*resource)) {
             if (!resource->buffer.Upload(bufferObject.MappedData() + offset,
                                          static_cast<VkDeviceSize>(size), static_cast<VkDeviceSize>(offset))) {
-                MGLOG_E("VkBufferManager::OnFlushMappedRange: host upload failed");
+                MGLOG_E_ONCE("VkBufferManager::OnFlushMappedRange: host upload failed");
                 resource->pendingFullUpload = true;
             }
             return;
@@ -542,7 +563,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
         const VkDeviceSize size = static_cast<VkDeviceSize>(bufferObject->GetSize());
         if (size == 0) {
-            MGLOG_E("VkBufferManager::AcquireResidentSlice failed: buffer size is zero");
+            MGLOG_E_ONCE("VkBufferManager::AcquireResidentSlice failed: buffer size is zero");
             return false;
         }
 
@@ -564,7 +585,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 return false;
             }
             if (!resource->buffer.Upload(bufferObject->MappedData(), size, 0)) {
-                MGLOG_E("VkBufferManager::AcquireResidentSlice failed: initial upload failed");
+                MGLOG_E_ONCE("VkBufferManager::AcquireResidentSlice failed: initial upload failed");
                 resource->buffer.Destroy();
                 resource->storageSize = 0;
                 resource->usageFlags = 0;
@@ -599,7 +620,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
         const VkDeviceSize size = static_cast<VkDeviceSize>(bufferObject->GetSize());
         if (size == 0) {
-            MGLOG_E("VkBufferManager::AcquireStreamedSlice failed: buffer size is zero");
+            MGLOG_E_ONCE("VkBufferManager::AcquireStreamedSlice failed: buffer size is zero");
             return false;
         }
 
@@ -697,7 +718,13 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         case BufferKind::Uniform:
             return VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT;
         case BufferKind::TextureBuffer:
-            return VK_BUFFER_USAGE_UNIFORM_TEXEL_BUFFER_BIT;
+            // Both texel roles, for the same reason vertex/index carry both bits: one GL buffer
+            // texture can be read as a samplerBuffer and written as an imageBuffer, and which of
+            // the two it is only becomes known when a shader that uses it is bound - long after
+            // the resident buffer was created. A VkBufferView for a storage-texel descriptor is
+            // invalid unless the buffer was created with the storage bit, so a buffer that
+            // acquired only the uniform bit could never be given one.
+            return VK_BUFFER_USAGE_UNIFORM_TEXEL_BUFFER_BIT | VK_BUFFER_USAGE_STORAGE_TEXEL_BUFFER_BIT;
         case BufferKind::ShaderStorage:
             return VK_BUFFER_USAGE_STORAGE_BUFFER_BIT | VK_BUFFER_USAGE_INDIRECT_BUFFER_BIT;
         case BufferKind::Indirect:

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
+        // Drains every frame slot's deferred buffer/resource releases. Only valid when
-        // transient arena's parked superseded blocks). Only valid when the
+        // the caller has proven every queue submission complete; used by the present-less
-        // caller has proven every queue submission complete; used by the
+        // frame-boundary drain. Deliberately does NOT touch the transient arena's parked
-        // present-less frame-boundary drain.
+        // 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/VkBufferObject.cpp

@@ -76,7 +76,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         const VkResult result =
             vmaCreateBuffer(m_allocator, &bufferInfo, &allocationInfo, &m_buffer, &m_allocation, nullptr);
         if (result != VK_SUCCESS) {
-            MGLOG_E("VkBufferObject::Create failed: vmaCreateBuffer returned %d", result);
+            MGLOG_E_ONCE("VkBufferObject::Create failed: vmaCreateBuffer returned %d", result);
             m_allocator = nullptr;
             m_buffer = VK_NULL_HANDLE;
             m_allocation = nullptr;
@@ -108,7 +108,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
         const VkResult mapResult = vmaMapMemory(m_allocator, m_allocation, &m_mappedData);
         if (mapResult != VK_SUCCESS || m_mappedData == nullptr) {
-            MGLOG_E("VkBufferObject::Map failed: vmaMapMemory returned %d", mapResult);
+            MGLOG_E_ONCE("VkBufferObject::Map failed: vmaMapMemory returned %d", mapResult);
             m_mappedData = nullptr;
             return nullptr;
         }
@@ -138,14 +138,14 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         const Bool wasMapped = IsMapped();
         void* mapped = wasMapped ? m_mappedData : Map();
         if (mapped == nullptr) {
-            MGLOG_E("VkBufferObject::Upload failed: unable to map buffer");
+            MGLOG_E_ONCE("VkBufferObject::Upload failed: unable to map buffer");
             return false;
         }
 
         Memcpy(static_cast<Uint8*>(mapped) + offset, data, static_cast<SizeT>(size));
         const VkResult flushResult = vmaFlushAllocation(m_allocator, m_allocation, offset, size);
         if (flushResult != VK_SUCCESS) {
-            MGLOG_E("VkBufferObject::Upload failed: vmaFlushAllocation returned %d", flushResult);
+            MGLOG_E_ONCE("VkBufferObject::Upload failed: vmaFlushAllocation returned %d", flushResult);
             if (!wasMapped) {
                 Unmap();
             }
@@ -170,7 +170,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
         const VkResult result = vmaInvalidateAllocation(m_allocator, m_allocation, offset, resolvedSize);
         if (result != VK_SUCCESS) {
-            MGLOG_E("VkBufferObject::Invalidate failed: vmaInvalidateAllocation returned %d", result);
+            MGLOG_E_ONCE("VkBufferObject::Invalidate failed: vmaInvalidateAllocation returned %d", result);
             return false;
         }
         return true;

MobileGL/MG_Backend/DirectVulkan/Renderer/VkClearManager.cpp

@@ -166,7 +166,15 @@ namespace MobileGL::MG_Backend::DirectVulkan {
     void VkClearManager::MergeClearPayload(ClearAttachmentPayload& dst, const ClearAttachmentPayload& src) {
         dst.mask |= src.mask;
         if ((src.mask & GL_COLOR_BUFFER_BIT) != 0) {
+            // The whole colour story travels together (same rule as
+            // VkRenderPassManager::QueueRenderbufferClear): a glClearBufferiv/uiv
+            // payload carries its value in colorInt/colorUint and its branch selector
+            // in colorEncoding - dropping them here would leave the pending clear
+            // reading as an all-zero float one.
             dst.color = src.color;
+            dst.colorEncoding = src.colorEncoding;
+            dst.colorInt = src.colorInt;
+            dst.colorUint = src.colorUint;
         }
         if ((src.mask & GL_DEPTH_BUFFER_BIT) != 0) {
             dst.depth = src.depth;

MobileGL/MG_Backend/DirectVulkan/Renderer/VkRenderPassManager.cpp

@@ -123,7 +123,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         }
 
         if (!attachment.IsComplete()) {
-            MGLOG_W("GetOrCreateRenderPass: draw buffer slot %u (%s) on FBO %u has an incomplete texture attachment; using VK_ATTACHMENT_UNUSED",
+            MGLOG_W_ONCE("GetOrCreateRenderPass: draw buffer slot %u (%s) on FBO %u has an incomplete texture attachment; using VK_ATTACHMENT_UNUSED",
                     drawBufferIndex,
                     MG_Util::ConvertFramebufferAttachmentTypeToString(attachmentType).c_str(),
                     fbo.GetExternalIndex());
@@ -132,7 +132,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
         auto* texture = attachment.GetTexture().get();
         if (texture == nullptr) {
-            MGLOG_W("GetOrCreateRenderPass: draw buffer slot %u (%s) on FBO %u resolved to a null texture; using VK_ATTACHMENT_UNUSED",
+            MGLOG_W_ONCE("GetOrCreateRenderPass: draw buffer slot %u (%s) on FBO %u resolved to a null texture; using VK_ATTACHMENT_UNUSED",
                     drawBufferIndex,
                     MG_Util::ConvertFramebufferAttachmentTypeToString(attachmentType).c_str(),
                     fbo.GetExternalIndex());
@@ -311,7 +311,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
         VkSampleCountFlagBits sampleCount = VK_SAMPLE_COUNT_1_BIT;
         if (!TryResolveSampleCountFlagBits(renderbuffer->GetSamples(), sampleCount)) {
-            MGLOG_E("GetOrCreateRenderbufferResource: unsupported renderbuffer sample count %d for renderbuffer %u",
+            MGLOG_E_ONCE("GetOrCreateRenderbufferResource: unsupported renderbuffer sample count %d for renderbuffer %u",
                     renderbuffer->GetSamples(),
                     renderbuffer->GetExternalIndex());
             return nullptr;
@@ -457,7 +457,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             m_physicalDevice, format, imageInfo.imageType, imageInfo.tiling, imageInfo.usage, imageInfo.flags,
             &imageFormatProperties);
         if (imageFormatResult != VK_SUCCESS || (imageFormatProperties.sampleCounts & sampleCount) == 0) {
-            MGLOG_E("GetOrCreateRenderbufferResource: unsupported renderbuffer format=%d samples=%d for renderbuffer %u",
+            MGLOG_E_ONCE("GetOrCreateRenderbufferResource: unsupported renderbuffer format=%d samples=%d for renderbuffer %u",
                     static_cast<Int>(format),
                     static_cast<Int>(sampleCount),
                     renderbuffer->GetExternalIndex());
@@ -831,6 +831,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // recreated since (texture + renderbuffer image epochs), and no pending clear (which alters
         // load ops). Any of these differing forces the full recompute below. Portable to VK 1.1.
         if (activeRenderPass != nullptr && m_rpFastValid && m_rpFastFbo == &fbo &&
+            m_rpFastFboLifetimeId == fbo.GetLifetimeId() &&
             m_rpFastFboVersion == fbo.GetObjectVersion() && m_rpFastSwapchainIndex == swapchainImageIndex &&
             m_rpFastTexEpoch == m_textureManager.GetTextureImageEpoch() &&
             m_rpFastRbEpoch == m_renderbufferImageEpoch &&
@@ -855,6 +856,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             // epochs AFTER ComputeHash: its attachment SyncTexture can create an image (bump the epoch).
             m_rpFastValid = true;
             m_rpFastFbo = &fbo;
+            m_rpFastFboLifetimeId = fbo.GetLifetimeId();
             m_rpFastFboVersion = fbo.GetObjectVersion();
             m_rpFastSwapchainIndex = swapchainImageIndex;
             m_rpFastTexEpoch = m_textureManager.GetTextureImageEpoch();
@@ -929,7 +931,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                     const auto& renderbuffer = rbAtt.GetRenderbuffer();
                     auto* rbResource = GetOrCreateRenderbufferResource(renderbuffer);
                     if (rbResource == nullptr || (rbResource->aspect & VK_IMAGE_ASPECT_COLOR_BIT) == 0) {
-                        MGLOG_E("GetOrCreateRenderPass: draw buffer slot %u on FBO %u has an unsupported color "
+                        MGLOG_E_ONCE("GetOrCreateRenderPass: draw buffer slot %u on FBO %u has an unsupported color "
                                 "renderbuffer %u; using VK_ATTACHMENT_UNUSED",
                                 i, fbo.GetExternalIndex(), renderbuffer->GetExternalIndex());
                         continue;
@@ -1105,7 +1107,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                     adoptRenderPassSampleCount(attachmentSampleCount, "color", texture->GetExternalIndex());
 
                     if (!hasClear && trackedColorLayout == VK_IMAGE_LAYOUT_UNDEFINED) {
-                        MGLOG_W("GetOrCreateRenderPass: color attachment textureId=%d starts with undefined layout and no clear; "
+                        MGLOG_W_ONCE("GetOrCreateRenderPass: color attachment textureId=%d starts with undefined layout and no clear; "
                                 "using LOAD_OP_DONT_CARE",
                                 texture->GetExternalIndex());
                         desc.loadOp = VK_ATTACHMENT_LOAD_OP_DONT_CARE;
@@ -1161,7 +1163,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             isUsableDepthStencilAttachment(depthAtt) && isUsableDepthStencilAttachment(stencilAtt) &&
             !sameDepthStencilAttachmentObject(depthAtt, stencilAtt);
         if (hasDistinctDepthAndStencilAttachments) {
-            MGLOG_E("GetOrCreateRenderPass: separate depth/stencil attachments are not supported yet; using the depth attachment and ignoring the standalone stencil attachment for framebuffer %u",
+            MGLOG_E_ONCE("GetOrCreateRenderPass: separate depth/stencil attachments are not supported yet; using the depth attachment and ignoring the standalone stencil attachment for framebuffer %u",
                     fbo.GetExternalIndex());
         }
         if (selectedDepthStencilAttachment != nullptr) {
@@ -1223,7 +1225,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             depthAttachmentDescription.finalLayout = VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL;
             depthAttachmentDescription.initialLayout = loadInfo.initialLayout;
             if (trackedDepthLayout == VK_IMAGE_LAYOUT_UNDEFINED && (!clearDepth || !clearStencil)) {
-                MGLOG_W("GetOrCreateRenderPass: depth/stencil attachment id=%d starts with undefined layout "
+                MGLOG_W_ONCE("GetOrCreateRenderPass: depth/stencil attachment id=%d starts with undefined layout "
                         "and partial/no clear; using DONT_CARE for uncleared aspects",
                         depthAttachmentId);
             }
@@ -1507,7 +1509,23 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             ClearAttachmentPayload clearPayload{};
             SharedPtr<MG_State::GLState::ITextureObject> liveTexture;
             if (pending.hasInlinePayload) {
+                // The inline payload was snapshotted when the entry was CREATED, but the
+                // clear VALUE is not part of the entry's hash - a cache hit with a newer
+                // glClear would replay the creation-time value and drop the new one (the
+                // texture path below is immune because it re-reads the live payload).
+                // Same defense as ClearAttachmentsOnActiveRenderPass: prefer the live
+                // pending clear, fall back to the snapshot only when none is queued.
+                if (s_renderPassManager != nullptr &&
+                    s_renderPassManager->GetPendingRenderbufferClear(pending.renderbuffer, clearPayload)) {
+                    if ((clearPayload.mask & GL_COLOR_BUFFER_BIT) != 0 && pending.renderbuffer != nullptr &&
+                        MG_Util::GetBaseInternalFormatComponentCount(pending.renderbuffer->GetInternalFormat()) ==
+                            3) {
+                        // RGB renderbuffers are backed by an RGBA image; the missing alpha reads as 1.
+                        ForceOpaqueClearAlpha(clearPayload);
+                    }
+                } else {
                     clearPayload = pending.inlinePayload;
+                }
             } else {
                 if (pending.key.texture == nullptr ||
                     !s_clearManager->GetPendingClear(pending.key, clearPayload, liveTexture)) {

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,
@@ -253,6 +289,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // or a pending clear. Portable to Vulkan 1.1 (no dynamic_rendering / imageless FB needed).
         Bool m_rpFastValid = false;
         const MG_State::GLState::FramebufferObject* m_rpFastFbo = nullptr;
+        // The FBO's never-reused lifetime id joins the raw pointer + Uint16 version:
+        // a deleted FBO reallocated at the same address whose fresh setup performed
+        // the same number of version bumps would otherwise compare equal (both count
+        // from 0), serving the dead framebuffer's pass to the new object.
+        Uint64 m_rpFastFboLifetimeId = 0;
         Uint16 m_rpFastFboVersion = 0;
         Uint32 m_rpFastSwapchainIndex = 0;
         Uint64 m_rpFastTexEpoch = 0;
@@ -315,26 +356,30 @@ 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
+        // materializes the source's pending clear, which looks that same resource up again. Growing
-        // operator[] runs its load-factor check before find_key and reallocates the whole bucket array
+        // an open-addressed table relocates every element, so the cached pointer went on to name
-        // when occupancy crosses it, so even a plain lookup relocates every element; erase only
+        // freed storage still holding the pre-clear VK_IMAGE_LAYOUT_UNDEFINED; BlitFramebuffer bailed
-        // tombstones and never decrements the occupancy, so the doubling keeps firing. After a
+        // out at "source image layout is undefined", silently dropping the blit -
-        // relocation the cached pointer names freed storage still holding the pre-clear
+        // renderbuffers_storage_multisample read back zero instead of the clear colour on exactly the
-        // VK_IMAGE_LAYOUT_UNDEFINED, and BlitFramebuffer bails out at "source image layout is
+        // iterations that grew the table.
-        // 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
+        // ordering rule. m_textureResources is node-based for the same reason.
-        // 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.
+        // 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/VkTextureManager.cpp

@@ -300,8 +300,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         Bool ok = VkTextureManager::TransitionImageLayout(
             commandBuffer, newResource.image, newResource.layout, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL,
             VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT, VK_PIPELINE_STAGE_TRANSFER_BIT,
-            0, VK_ACCESS_TRANSFER_WRITE_BIT, newResource.aspect, 0, newResource.mipLevels,
+            0, VK_ACCESS_TRANSFER_WRITE_BIT, newResource.aspect, 0, newResource.mipLevels);
-            newResource.arrayLayers);
         MOBILEGL_ASSERT(ok, "PreserveTextureContentsOnRecreate: failed to prepare destination image");
 
         VkImageLayout srcTrackedLayout = oldResource.layout;
@@ -311,8 +310,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         ok = VkTextureManager::TransitionImageLayout(
             commandBuffer, oldResource.image, srcTrackedLayout, VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL,
             srcStageMask, VK_PIPELINE_STAGE_TRANSFER_BIT,
-            srcAccessMask, VK_ACCESS_TRANSFER_READ_BIT, oldResource.aspect, 0, preservedMipLevels,
+            srcAccessMask, VK_ACCESS_TRANSFER_READ_BIT, oldResource.aspect, 0, preservedMipLevels);
-            oldResource.arrayLayers);
         MOBILEGL_ASSERT(ok, "PreserveTextureContentsOnRecreate: failed to prepare source image");
 
         Vector<VkImageCopy> copyRegions;
@@ -344,8 +342,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         ok = VkTextureManager::TransitionImageLayout(
             commandBuffer, newResource.image, newResource.layout, oldResource.layout,
             VK_PIPELINE_STAGE_TRANSFER_BIT, dstStageMask,
-            VK_ACCESS_TRANSFER_WRITE_BIT, dstAccessMask, newResource.aspect, 0, newResource.mipLevels,
+            VK_ACCESS_TRANSFER_WRITE_BIT, dstAccessMask, newResource.aspect, 0, newResource.mipLevels);
-            newResource.arrayLayers);
         MOBILEGL_ASSERT(ok, "PreserveTextureContentsOnRecreate: failed to restore destination layout");
 
         VK_VERIFY(vkEndCommandBuffer(commandBuffer), "vkEndCommandBuffer(texture preserve)");
@@ -950,7 +947,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
         const TextureFormatInfo formatInfo = ResolveTextureFormatInfo(texture.GetFormat());
         const VkComponentMapping sampledComponents = ResolveSampledViewComponents(texture, formatInfo);
-        const VkImageAspectFlags sampledAspect = ResolveSampledImageViewAspectMask(resource->aspect);
+        const VkImageAspectFlags sampledAspect =
+            ResolveSampledImageViewAspectMask(resource->aspect, texture.GetDepthStencilTextureMode());
         perMipSampledView = CreateImageView(resource->image, resource->format, sampledAspect, resource->viewType,
                                             mipLevel, 1, 0, resource->arrayLayers, &sampledComponents);
         if (perMipSampledView == VK_NULL_HANDLE) {
@@ -974,13 +972,13 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             return resource->sampledView;
         }
         if (!AreSampledImageViewFormatsCompatible(resource->format, format)) {
-            MGLOG_E("%s: incompatible sampled image view format=%d for textureId=%d imageFormat=%d",
+            MGLOG_E_ONCE("%s: incompatible sampled image view format=%d for textureId=%d imageFormat=%d",
                     __func__, static_cast<Int>(format), texture.GetExternalIndex(),
                     static_cast<Int>(resource->format));
             return VK_NULL_HANDLE;
         }
         if ((resource->imageCreateFlags & VK_IMAGE_CREATE_MUTABLE_FORMAT_BIT) == 0) {
-            MGLOG_E("%s: textureId=%d needs mutable image format=%d for sampled view format=%d",
+            MGLOG_E_ONCE("%s: textureId=%d needs mutable image format=%d for sampled view format=%d",
                     __func__, texture.GetExternalIndex(), static_cast<Int>(resource->format),
                     static_cast<Int>(format));
             return VK_NULL_HANDLE;
@@ -1000,7 +998,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         VkFormatProperties formatProperties{};
         vkGetPhysicalDeviceFormatProperties(m_physicalDevice, format, &formatProperties);
         if ((formatProperties.optimalTilingFeatures & VK_FORMAT_FEATURE_SAMPLED_IMAGE_BIT) == 0) {
-            MGLOG_E("%s: sampled image view format=%d lacks VK_FORMAT_FEATURE_SAMPLED_IMAGE_BIT "
+            MGLOG_E_ONCE("%s: sampled image view format=%d lacks VK_FORMAT_FEATURE_SAMPLED_IMAGE_BIT "
                     "for textureId=%d (available=0x%x)",
                     __func__, static_cast<Int>(format), texture.GetExternalIndex(),
                     static_cast<Uint32>(formatProperties.optimalTilingFeatures));
@@ -1014,7 +1012,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             resource->sampledBaseMipLevel, resource->sampledLevelCount, 0, resource->arrayLayers,
             &sampledComponents, VK_IMAGE_USAGE_SAMPLED_BIT);
         if (view == VK_NULL_HANDLE) {
-            MGLOG_E("%s: failed to create sampled image view textureId=%d imageFormat=%d viewFormat=%d",
+            MGLOG_E_ONCE("%s: failed to create sampled image view textureId=%d imageFormat=%d viewFormat=%d",
                     __func__, texture.GetExternalIndex(), static_cast<Int>(resource->format),
                     static_cast<Int>(format));
             return VK_NULL_HANDLE;
@@ -1042,14 +1040,14 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             format = resource->format;
         }
         if (!AreStorageImageViewFormatsCompatible(resource->format, format)) {
-            MGLOG_E("%s: incompatible storage image view format=%d for textureId=%d imageFormat=%d",
+            MGLOG_E_ONCE("%s: incompatible storage image view format=%d for textureId=%d imageFormat=%d",
                     __func__, static_cast<Int>(format), texture.GetExternalIndex(),
                     static_cast<Int>(resource->format));
             return VK_NULL_HANDLE;
         }
         if (format != resource->format &&
             (resource->imageCreateFlags & VK_IMAGE_CREATE_MUTABLE_FORMAT_BIT) == 0) {
-            MGLOG_E("%s: textureId=%d needs mutable image format=%d for storage view format=%d",
+            MGLOG_E_ONCE("%s: textureId=%d needs mutable image format=%d for storage view format=%d",
                     __func__, texture.GetExternalIndex(), static_cast<Int>(resource->format),
                     static_cast<Int>(format));
             return VK_NULL_HANDLE;
@@ -1069,7 +1067,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 viewType = VK_IMAGE_VIEW_TYPE_2D;
                 break;
             case VK_IMAGE_VIEW_TYPE_3D:
-                MGLOG_E("%s: non-layered 3D storage views are unsupported for textureId=%d",
+                MGLOG_E_ONCE("%s: non-layered 3D storage views are unsupported for textureId=%d",
                         __func__, texture.GetExternalIndex());
                 return VK_NULL_HANDLE;
             default:
@@ -1078,7 +1076,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
             if (viewType != resource->viewType) {
                 if (layer < 0 || static_cast<Uint32>(layer) >= resource->arrayLayers) {
-                    MGLOG_E("%s: storage image layer=%d is out of range for textureId=%d arrayLayers=%u",
+                    MGLOG_E_ONCE("%s: storage image layer=%d is out of range for textureId=%d arrayLayers=%u",
                             __func__, layer, texture.GetExternalIndex(), resource->arrayLayers);
                     return VK_NULL_HANDLE;
                 }
@@ -1113,7 +1111,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         VkFormatProperties formatProperties{};
         vkGetPhysicalDeviceFormatProperties(m_physicalDevice, format, &formatProperties);
         if ((formatProperties.optimalTilingFeatures & requiredFormatFeatures) != requiredFormatFeatures) {
-            MGLOG_E("%s: storage image view format=%d lacks required features=0x%x for textureId=%d "
+            MGLOG_E_ONCE("%s: storage image view format=%d lacks required features=0x%x for textureId=%d "
                     "(available=0x%x)",
                     __func__, static_cast<Int>(format), static_cast<Uint32>(requiredFormatFeatures),
                     texture.GetExternalIndex(), static_cast<Uint32>(formatProperties.optimalTilingFeatures));
@@ -1124,7 +1122,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                                                  mipLevel, 1, baseArrayLayer, layerCount, nullptr,
                                                  VK_IMAGE_USAGE_STORAGE_BIT);
         if (view == VK_NULL_HANDLE) {
-            MGLOG_E("%s: failed to create storage image view for textureId=%d mip=%u imageFormat=%d viewFormat=%d",
+            MGLOG_E_ONCE("%s: failed to create storage image view for textureId=%d mip=%u imageFormat=%d viewFormat=%d",
                     __func__, texture.GetExternalIndex(), mipLevel, static_cast<Int>(resource->format),
                     static_cast<Int>(format));
             return VK_NULL_HANDLE;
@@ -1190,7 +1188,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 const Bool lowerTransitioned = TransitionImageLayout(
                     commandBuffer, resource.image, lowerMipLayout, newLayout,
                     srcStageMask, dstStageMask, srcAccessMask, dstAccessMask,
-                    resource.aspect, 0, writtenMipLevel, resource.arrayLayers);
+                    resource.aspect, 0, writtenMipLevel);
                 MOBILEGL_ASSERT(lowerTransitioned,
                                 "UpdateTrackedImageLayoutAfterAttachmentWrite: failed to transition lower mip levels for textureId=%d",
                                 texture->GetExternalIndex());
@@ -1202,8 +1200,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 const Bool upperTransitioned = TransitionImageLayout(
                     commandBuffer, resource.image, upperMipLayout, newLayout,
                     srcStageMask, dstStageMask, srcAccessMask, dstAccessMask,
-                    resource.aspect, upperBaseMipLevel, resource.mipLevels - upperBaseMipLevel,
+                    resource.aspect, upperBaseMipLevel, resource.mipLevels - upperBaseMipLevel);
-                    resource.arrayLayers);
                 MOBILEGL_ASSERT(upperTransitioned,
                                 "UpdateTrackedImageLayoutAfterAttachmentWrite: failed to transition upper mip levels for textureId=%d",
                                 texture->GetExternalIndex());
@@ -1222,7 +1219,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             return true;
         }
         if (resource->layout == VK_IMAGE_LAYOUT_UNDEFINED) {
-            MGLOG_W("TransitionTextureForSampling: textureId=%d is still in VK_IMAGE_LAYOUT_UNDEFINED before sampling",
+            MGLOG_W_ONCE("TransitionTextureForSampling: textureId=%d is still in VK_IMAGE_LAYOUT_UNDEFINED before sampling",
                     texture.GetExternalIndex());
         }
 
@@ -1256,8 +1253,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
         const Bool ok = TransitionImageLayout(commandBuffer, resource->image, resource->layout, targetLayout, srcStageMask,
                                               s_sampledReadStages, srcAccessMask,
-                                              VK_ACCESS_SHADER_READ_BIT, resource->aspect, 0, resource->mipLevels,
+                                              VK_ACCESS_SHADER_READ_BIT, resource->aspect, 0, resource->mipLevels);
-                                              resource->arrayLayers);
         MOBILEGL_ASSERT(ok, "TransitionTextureForSampling: transition failed for textureId=%d", texture.GetExternalIndex());
         // Pre-pass stream bookkeeping: a command referencing the image was recorded.
         StampResourceRecordingUse(*resource);
@@ -1287,7 +1283,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                                               VK_IMAGE_LAYOUT_GENERAL, srcStageMask,
                                               VK_PIPELINE_STAGE_ALL_COMMANDS_BIT, srcAccessMask,
                                               VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT,
-                                              resource->aspect, 0, resource->mipLevels, resource->arrayLayers);
+                                              resource->aspect, 0, resource->mipLevels);
         MOBILEGL_ASSERT(ok, "TransitionTextureForStorageImage: transition failed for textureId=%d",
                         texture.GetExternalIndex());
         // Pre-pass stream bookkeeping: a command referencing the image was recorded.
@@ -1295,6 +1291,158 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         return ok;
     }
 
+    Bool VkTextureManager::SnapshotTextureForSampling(VkCommandBuffer commandBuffer,
+                                                      MG_State::GLState::ITextureObject& texture,
+                                                      SamplerNumericDomain numericDomain,
+                                                      VkPipelineStageFlags consumerShaderStageMask,
+                                                      SampledTextureSnapshot& outSnapshot) {
+        outSnapshot = {};
+        TextureResource* source = SyncTextureAndGetDescriptor(texture);
+        if (source == nullptr || source->image == VK_NULL_HANDLE || source->sampleCount != VK_SAMPLE_COUNT_1_BIT ||
+            source->sampledLevelCount == 0) {
+            return false;
+        }
+
+        const VkFormat sampledFormat = ResolveSampledImageViewFormat(source->format, numericDomain);
+        if (sampledFormat == VK_FORMAT_UNDEFINED ||
+            !AreSampledImageViewFormatsCompatible(source->format, sampledFormat)) {
+            MGLOG_E_ONCE("SnapshotTextureForSampling: textureId=%d cannot create sampled view format=%d from image format=%d",
+                         texture.GetExternalIndex(), static_cast<Int>(sampledFormat), static_cast<Int>(source->format));
+            return false;
+        }
+        if (sampledFormat != source->format &&
+            (source->imageCreateFlags & VK_IMAGE_CREATE_MUTABLE_FORMAT_BIT) == 0) {
+            MGLOG_E_ONCE("SnapshotTextureForSampling: textureId=%d needs unavailable mutable image format=%d for sampled view=%d",
+                         texture.GetExternalIndex(), static_cast<Int>(source->format), static_cast<Int>(sampledFormat));
+            return false;
+        }
+
+        VkImageType imageType = VK_IMAGE_TYPE_2D;
+        switch (source->viewType) {
+        case VK_IMAGE_VIEW_TYPE_1D:
+        case VK_IMAGE_VIEW_TYPE_1D_ARRAY:
+            imageType = VK_IMAGE_TYPE_1D;
+            break;
+        case VK_IMAGE_VIEW_TYPE_3D:
+            imageType = VK_IMAGE_TYPE_3D;
+            break;
+        default:
+            break;
+        }
+
+        TextureResource snapshot{};
+        VkImageCreateInfo imageInfo{};
+        imageInfo.sType = VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO;
+        imageInfo.flags = source->imageCreateFlags;
+        imageInfo.imageType = imageType;
+        imageInfo.extent = {source->extent.width, source->extent.height, source->depth};
+        imageInfo.mipLevels = source->mipLevels;
+        imageInfo.arrayLayers = source->arrayLayers;
+        imageInfo.format = source->format;
+        imageInfo.tiling = VK_IMAGE_TILING_OPTIMAL;
+        imageInfo.initialLayout = VK_IMAGE_LAYOUT_UNDEFINED;
+        imageInfo.usage = VK_IMAGE_USAGE_TRANSFER_DST_BIT | VK_IMAGE_USAGE_SAMPLED_BIT;
+        imageInfo.samples = VK_SAMPLE_COUNT_1_BIT;
+        imageInfo.sharingMode = VK_SHARING_MODE_EXCLUSIVE;
+
+        // Keep the temporary's view-format list just as narrow as the source's sampler use. This
+        // has no storage-image usage, so unlike an app image binding the exact list is knowable.
+        Vector<VkFormat> viewFormats;
+        VkImageFormatListCreateInfo formatListInfo{};
+        if (m_imageFormatListSupported && (imageInfo.flags & VK_IMAGE_CREATE_MUTABLE_FORMAT_BIT) != 0) {
+            viewFormats.push_back(source->format);
+            if (sampledFormat != source->format) {
+                viewFormats.push_back(sampledFormat);
+            }
+            formatListInfo.sType = VK_STRUCTURE_TYPE_IMAGE_FORMAT_LIST_CREATE_INFO;
+            formatListInfo.viewFormatCount = static_cast<Uint32>(viewFormats.size());
+            formatListInfo.pViewFormats = viewFormats.data();
+            imageInfo.pNext = &formatListInfo;
+        }
+
+        VmaAllocationCreateInfo allocationInfo{};
+        allocationInfo.usage = VMA_MEMORY_USAGE_AUTO_PREFER_DEVICE;
+        allocationInfo.requiredFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT;
+        const VkResult createResult =
+            vmaCreateImage(m_allocator, &imageInfo, &allocationInfo, &snapshot.image, &snapshot.allocation, nullptr);
+        if (createResult != VK_SUCCESS) {
+            MGLOG_E_ONCE("SnapshotTextureForSampling: vmaCreateImage failed result=%d textureId=%d", createResult,
+                         texture.GetExternalIndex());
+            return false;
+        }
+
+        snapshot.extent = source->extent;
+        snapshot.depth = source->depth;
+        snapshot.arrayLayers = source->arrayLayers;
+        snapshot.mipLevels = source->mipLevels;
+        snapshot.sampledBaseMipLevel = source->sampledBaseMipLevel;
+        snapshot.sampledLevelCount = source->sampledLevelCount;
+        snapshot.format = source->format;
+        snapshot.aspect = source->aspect;
+        snapshot.viewType = source->viewType;
+        snapshot.sampleCount = VK_SAMPLE_COUNT_1_BIT;
+        snapshot.imageCreateFlags = imageInfo.flags;
+        snapshot.usageFlags = imageInfo.usage;
+
+        const TextureFormatInfo formatInfo = ResolveTextureFormatInfo(texture.GetFormat());
+        const VkComponentMapping sampledComponents = ResolveSampledViewComponents(texture, formatInfo);
+        const VkImageAspectFlags sampledAspect =
+            ResolveSampledImageViewAspectMask(snapshot.aspect, texture.GetDepthStencilTextureMode());
+        snapshot.sampledView = CreateImageView(snapshot.image, sampledFormat, sampledAspect, snapshot.viewType,
+                                               snapshot.sampledBaseMipLevel, snapshot.sampledLevelCount, 0,
+                                               snapshot.arrayLayers, &sampledComponents);
+        if (snapshot.sampledView == VK_NULL_HANDLE) {
+            MGLOG_E_ONCE("SnapshotTextureForSampling: failed to create sampled view textureId=%d", texture.GetExternalIndex());
+            return false;
+        }
+
+        VkPipelineStageFlags sourceStageMask = VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT;
+        VkAccessFlags sourceAccessMask = 0;
+        const VkImageLayout sourceLayout = source->layout;
+        GetImageTransitionSourceState(sourceLayout, sourceStageMask, sourceAccessMask);
+        if (!TransitionImageLayout(commandBuffer, source->image, source->layout, VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL,
+                                   sourceStageMask, VK_PIPELINE_STAGE_TRANSFER_BIT, sourceAccessMask,
+                                   VK_ACCESS_TRANSFER_READ_BIT, source->aspect, 0, source->mipLevels) ||
+            !TransitionImageLayout(commandBuffer, snapshot.image, snapshot.layout, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL,
+                                   VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT, VK_PIPELINE_STAGE_TRANSFER_BIT, 0,
+                                   VK_ACCESS_TRANSFER_WRITE_BIT, snapshot.aspect, snapshot.sampledBaseMipLevel,
+                                   snapshot.sampledLevelCount)) {
+            return false;
+        }
+
+        Vector<VkImageCopy> copyRegions;
+        copyRegions.reserve(snapshot.sampledLevelCount);
+        for (Uint32 level = snapshot.sampledBaseMipLevel;
+             level < snapshot.sampledBaseMipLevel + snapshot.sampledLevelCount; ++level) {
+            VkImageCopy copy{};
+            copy.srcSubresource = {source->aspect, level, 0, source->arrayLayers};
+            copy.dstSubresource = {snapshot.aspect, level, 0, snapshot.arrayLayers};
+            copy.extent = {std::max(source->extent.width >> level, 1u),
+                           std::max(source->extent.height >> level, 1u),
+                           std::max(source->depth >> level, 1u)};
+            copyRegions.push_back(copy);
+        }
+        vkCmdCopyImage(commandBuffer, source->image, VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL, snapshot.image,
+                       VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, static_cast<Uint32>(copyRegions.size()), copyRegions.data());
+
+        if (!TransitionImageLayout(commandBuffer, snapshot.image, snapshot.layout,
+                                   ResolveSampledReadOnlyLayout(snapshot.aspect), VK_PIPELINE_STAGE_TRANSFER_BIT,
+                                   consumerShaderStageMask, VK_ACCESS_TRANSFER_WRITE_BIT,
+                                   VK_ACCESS_SHADER_READ_BIT, snapshot.aspect, snapshot.sampledBaseMipLevel,
+                                   snapshot.sampledLevelCount) ||
+            !TransitionImageLayout(commandBuffer, source->image, source->layout, sourceLayout,
+                                   VK_PIPELINE_STAGE_TRANSFER_BIT, consumerShaderStageMask,
+                                   VK_ACCESS_TRANSFER_READ_BIT, VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT,
+                                   source->aspect, 0, source->mipLevels)) {
+            return false;
+        }
+
+        StampResourceRecordingUse(*source);
+        outSnapshot = {.imageView = snapshot.sampledView, .layout = snapshot.layout};
+        DeferResourceRelease(Move(snapshot));
+        return true;
+    }
+
     void VkTextureManager::MarkStorageImageTexture(MG_State::GLState::ITextureObject& texture) {
         m_storageImageTextures.insert(MakeTextureIdentity(&texture));
     }
@@ -1354,8 +1502,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                                                  VkImageLayout& trackedLayout, VkImageLayout newLayout,
                                                  VkPipelineStageFlags srcStageMask, VkPipelineStageFlags dstStageMask,
                                                  VkAccessFlags srcAccessMask, VkAccessFlags dstAccessMask,
-                                                 VkImageAspectFlags aspectMask, Uint32 baseMipLevel, Uint32 levelCount,
+                                                 VkImageAspectFlags aspectMask, Uint32 baseMipLevel,
-                                                 Uint32 layerCount) {
+                                                 Uint32 levelCount) {
         MOBILEGL_ASSERT(image != VK_NULL_HANDLE, "TransitionImageLayout: m_image == VK_NULL_HANDLE");
         MOBILEGL_ASSERT(!((dstAccessMask & VK_ACCESS_TRANSFER_READ_BIT) != 0 &&
                           (dstStageMask & VK_PIPELINE_STAGE_TRANSFER_BIT) == 0),
@@ -1380,7 +1528,13 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         barrier.subresourceRange.baseMipLevel = baseMipLevel;
         barrier.subresourceRange.levelCount = levelCount;
         barrier.subresourceRange.baseArrayLayer = 0;
-        barrier.subresourceRange.layerCount = layerCount;
+        // Every layer, always - see the declaration for why layout tracking leaves no other
+        // correct answer. VK_REMAINING_ARRAY_LAYERS rather than the image's own `arrayLayers`
+        // because those are not the same number for a 3D image: MobileGL creates 3D images
+        // 2D_ARRAY_COMPATIBLE and their arrayLayers is 1, which today Vulkan reads as "all depth
+        // slices" but will read as "depth slice 0" once VK_KHR_maintenance9 is enabled. The
+        // validation layer warns about that literal 1 by name.
+        barrier.subresourceRange.layerCount = VK_REMAINING_ARRAY_LAYERS;
         vkCmdPipelineBarrier(commandBuffer, srcStageMask, dstStageMask, 0, 0, nullptr, 0, nullptr, 1, &barrier);
 
         trackedLayout = newLayout;
@@ -1573,7 +1727,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             // targets this manager has no Vulkan image shape for yet (cube map arrays above all).
             // Declining the sync leaves the texture unbacked - wrong, but recoverable - where an
             // assertion would take the whole process down instead.
-            MGLOG_W("SyncTextureResource: unsupported uploadTarget=%s textureTarget=%s textureId=%d size=(%d,%d,%d) "
+            MGLOG_W_ONCE("SyncTextureResource: unsupported uploadTarget=%s textureTarget=%s textureId=%d size=(%d,%d,%d) "
                     "mipLevels=%u vkViewType=%d",
                     MG_Util::ConvertTextureUploadTargetToString(uploadTarget).c_str(),
                     MG_Util::ConvertTextureTargetToString(texture.GetTarget()).c_str(), texture.GetExternalIndex(),
@@ -1802,7 +1956,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 // Losing reinterpreted views only degrades the formatless-image feature for
                 // this texture; failing creation would lose the texture entirely, so retry
                 // as a plain immutable-format image.
-                MGLOG_W("%s: mutable image format=%d is unsupported for textureId=%d; creating "
+                MGLOG_W_ONCE("%s: mutable image format=%d is unsupported for textureId=%d; creating "
                         "without VK_IMAGE_CREATE_MUTABLE_FORMAT_BIT (format reinterpretation "
                         "will be unavailable for it)",
                         __func__, static_cast<Int>(format), texture.GetExternalIndex());
@@ -1820,7 +1974,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 // Losing 2D-array compatibility only costs per-slice framebuffer attachment for this
                 // format; failing creation would lose the texture entirely. Remembered so later syncs
                 // neither reprobe nor flag-mismatch against this image and recreate it.
-                MGLOG_W("%s: VK_IMAGE_CREATE_2D_ARRAY_COMPATIBLE_BIT is unsupported for format=%d "
+                MGLOG_W_ONCE("%s: VK_IMAGE_CREATE_2D_ARRAY_COMPATIBLE_BIT is unsupported for format=%d "
                         "textureId=%d; creating without it (per-slice framebuffer attachment will be "
                         "unavailable for it)",
                         __func__, static_cast<Int>(format), texture.GetExternalIndex());
@@ -1839,6 +1993,13 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                         texture.GetExternalIndex(),
                         MG_Util::ConvertTextureUploadTargetToString(uploadTarget).c_str(),
                         static_cast<Int>(format), static_cast<Uint32>(imageInfo.usage));
+                // The preserved image was written by GPU work that may still be in flight
+                // (preserve requires layout != UNDEFINED); park it on the deferred ring
+                // like every other destruction path instead of letting the unique_ptr
+                // destroy it synchronously under the GPU.
+                if (preservedResource) {
+                    DeferResourceRelease(Move(*preservedResource));
+                }
                 return false;
             }
         }
@@ -1852,13 +2013,21 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         const VkResult createImageResult =
             vmaCreateImage(m_allocator, &imageInfo, &allocationInfo, &resource.image, &resource.allocation, nullptr);
         if (createImageResult != VK_SUCCESS) {
-            MGLOG_F("SyncTextureResource: vmaCreateImage failed (%d) textureId=%d extent=%ux%u depth=%u layers=%u "
+            // E_ONCE, not F: the comment above says it - this is a soft failure the caller
+            // recovers from, and it re-fires on every sync of every texture the driver refuses.
+            MGLOG_E_ONCE("SyncTextureResource: vmaCreateImage failed (%d) textureId=%d extent=%ux%u depth=%u layers=%u "
                     "mips=%u samples=%d format=%d",
                     createImageResult, texture.GetExternalIndex(), imageInfo.extent.width, imageInfo.extent.height,
                     imageInfo.extent.depth, imageInfo.arrayLayers, imageInfo.mipLevels,
                     static_cast<Int>(imageInfo.samples), static_cast<Int>(imageInfo.format));
             resource.image = VK_NULL_HANDLE;
             resource.allocation = nullptr;
+            // Same as the probe failure above: the preserved live image must go through
+            // the deferred ring, never a synchronous destructor while frames that
+            // reference it are still in flight.
+            if (preservedResource) {
+                DeferResourceRelease(Move(*preservedResource));
+            }
             return false;
         }
         ++m_textureImageEpoch; // a new attachment image invalidates cached render passes
@@ -2238,7 +2407,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         if (resource.fullView == VK_NULL_HANDLE) {
             return false;
         }
-        const VkImageAspectFlags sampledAspect = ResolveSampledImageViewAspectMask(resource.aspect);
+        const VkImageAspectFlags sampledAspect =
+            ResolveSampledImageViewAspectMask(resource.aspect, texture.GetDepthStencilTextureMode());
         resource.sampledView = CreateImageView(resource.image, resource.format, sampledAspect, resource.viewType,
                                                baseMipLevel, levelCount, 0, resource.arrayLayers, &sampledComponents);
         if (resource.sampledView == VK_NULL_HANDLE) {
@@ -2424,7 +2594,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             const Bool srcIsD24S8 = outResource.format == VK_FORMAT_D24_UNORM_S8_UINT;
             const Bool srcIsD32FS8 = outResource.format == VK_FORMAT_D32_SFLOAT_S8_UINT;
             if (!srcIsD24S8 && !srcIsD32FS8) {
-                MGLOG_E("UploadDirtyMipLevels: unsupported combined depth-stencil format %d for textureId=%d",
+                MGLOG_E_ONCE("UploadDirtyMipLevels: unsupported combined depth-stencil format %d for textureId=%d",
                         static_cast<Int>(outResource.format), mipmapTexture.GetExternalIndex());
                 for (const auto& item : uploadItems) {
                     mipmapTexture.MarkStorageDirty(item.target, item.level, false);
@@ -2601,7 +2771,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                                         VK_PIPELINE_STAGE_TRANSFER_BIT,
                                         uploadSrcAccessMask,
                                         VK_ACCESS_TRANSFER_WRITE_BIT,
-                                        aspectMask, 0, outResource.mipLevels, outResource.arrayLayers);
+                                        aspectMask, 0, outResource.mipLevels);
         MOBILEGL_ASSERT(ok, "TransitionImageLayout to VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL failed");
 
         // Array textures keep their GL "depth" in VkImage array layers, so the
@@ -2705,7 +2875,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                                    s_sampledReadStages,
                                    VK_ACCESS_TRANSFER_WRITE_BIT,
                                    VK_ACCESS_SHADER_READ_BIT,
-                                   aspectMask, 0, outResource.mipLevels, outResource.arrayLayers);
+                                   aspectMask, 0, outResource.mipLevels);
         MOBILEGL_ASSERT(ok, "TransitionImageLayout to sampled read-only layout failed");
         outResource.layout = finalLayout;
 
@@ -2860,10 +3030,19 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         }
     }
 
-    VkImageAspectFlags VkTextureManager::ResolveSampledImageViewAspectMask(VkImageAspectFlags imageAspect) {
+    VkImageAspectFlags VkTextureManager::ResolveSampledImageViewAspectMask(VkImageAspectFlags imageAspect,
+                                                                           GLenum depthStencilTextureMode) {
         if ((imageAspect & VK_IMAGE_ASPECT_COLOR_BIT) != 0) {
             return VK_IMAGE_ASPECT_COLOR_BIT;
         }
+        // A sampled view of a combined depth/stencil image may name exactly one aspect
+        // (VUID-VkDescriptorImageInfo-imageView-01976), and GL_DEPTH_STENCIL_TEXTURE_MODE is
+        // what picks it - the whole content of GL_ARB_stencil_texturing. Depth stays the
+        // default, so nothing that never sets the mode changes shape. The texture's params
+        // version moves with the mode, which is what makes the cached views be rebuilt.
+        if (depthStencilTextureMode == GL_STENCIL_INDEX && (imageAspect & VK_IMAGE_ASPECT_STENCIL_BIT) != 0) {
+            return VK_IMAGE_ASPECT_STENCIL_BIT;
+        }
         if ((imageAspect & VK_IMAGE_ASPECT_DEPTH_BIT) != 0) {
             return VK_IMAGE_ASPECT_DEPTH_BIT;
         }

MobileGL/MG_Backend/DirectVulkan/Renderer/VkTextureManager.h

@@ -310,6 +310,11 @@ public:
         static inline VmaAllocator s_allocator = VK_NULL_HANDLE;
     };
 
+    struct SampledTextureSnapshot {
+        VkImageView imageView = VK_NULL_HANDLE;
+        VkImageLayout layout = VK_IMAGE_LAYOUT_UNDEFINED;
+    };
+
     Bool Initialize(const InitInfo& initInfo);
     void Shutdown();
     void BeginFrame(Uint32 frameIndex);
@@ -343,6 +348,13 @@ public:
                                                       VkImageLayout newLayout);
     Bool TransitionTextureForSampling(VkCommandBuffer commandBuffer, MG_State::GLState::ITextureObject& texture);
     Bool TransitionTextureForStorageImage(VkCommandBuffer commandBuffer, MG_State::GLState::ITextureObject& texture);
+    // Copies the complete sampler-visible mip range into a transient sampled image. The source is
+    // restored to its prior layout, so image-store descriptors continue to name the original image.
+    // The transient ownership is tied to the current frame slot and is safe through its submission.
+    Bool SnapshotTextureForSampling(VkCommandBuffer commandBuffer, MG_State::GLState::ITextureObject& texture,
+                                    SamplerNumericDomain numericDomain,
+                                    VkPipelineStageFlags consumerShaderStageMask,
+                                    SampledTextureSnapshot& outSnapshot);
 
     // Recording-generation bookkeeping for the pre-pass command stream. The
     // generation advances every time the frame command buffer (re)begins
@@ -379,17 +391,33 @@ public:
     // true - a false positive merely ends the render pass, a false negative would skip a barrier.
     Bool NeedsStorageImagePreparation(MG_State::GLState::ITextureObject& texture) const;
 
-    static VkImageAspectFlags ResolveSampledImageViewAspectMask(VkImageAspectFlags imageAspect);
+    // `depthStencilTextureMode` is the texture's GL_DEPTH_STENCIL_TEXTURE_MODE; it only decides
+    // anything for an image that carries both aspects. Defaulted so the call sites that have no
+    // texture in hand keep the depth-aspect answer they have always given.
+    static VkImageAspectFlags ResolveSampledImageViewAspectMask(VkImageAspectFlags imageAspect,
+                                                                GLenum depthStencilTextureMode = GL_DEPTH_COMPONENT);
     static VkFormat ResolveSampledImageViewFormat(VkFormat imageFormat, SamplerNumericDomain numericDomain);
     static Bool AreSampledImageViewFormatsCompatible(VkFormat imageFormat, VkFormat viewFormat);
     static Bool AreStorageImageViewFormatsCompatible(VkFormat imageFormat, VkFormat viewFormat);
 
+    // Moves `image` to `newLayout` and writes the new layout back through `trackedLayout`.
+    //
+    // The barrier covers EVERY array layer of the image, and there is deliberately no layer
+    // parameter to say otherwise: layout here is tracked per IMAGE (one `TextureResource::layout`,
+    // or one caller-owned variable), so a barrier narrower than the image would leave the layers it
+    // skipped in the old layout while the tracker claims they moved. Every transfer against a
+    // framebuffer attachment above layer 0 - glReadPixels, glBlitFramebuffer, glCopyTexSubImage,
+    // glCopyImageSubData - then ran its copy on a layer no barrier had transitioned.
+    //
+    // The mip range IS a parameter, because mip levels really are transitioned piecewise (see
+    // UpdateTrackedImageLayoutAfterAttachmentWrite and the mipmap generation loops): those callers
+    // move the complement of the level they wrote so the whole image converges on one layout again.
+    // Nothing does, or can, do that per layer.
     static Bool TransitionImageLayout(VkCommandBuffer commandBuffer, VkImage image, VkImageLayout& trackedLayout,
                                VkImageLayout newLayout, VkPipelineStageFlags srcStageMask,
                                VkPipelineStageFlags dstStageMask, VkAccessFlags srcAccessMask,
                                VkAccessFlags dstAccessMask, VkImageAspectFlags aspectMask,
-                               Uint32 baseMipLevel = 0, Uint32 levelCount = 1,
+                               Uint32 baseMipLevel = 0, Uint32 levelCount = 1);
-                               Uint32 layerCount = 1);
 
     SizeT CollectGarbage();
 

MobileGL/MG_Backend/DirectVulkan/Renderer/VkTimerQueryManager.cpp

@@ -15,7 +15,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         MOBILEGL_ASSERT(initInfo.device != VK_NULL_HANDLE, "VkTimerQueryManager::Initialize requires valid VkDevice");
         MOBILEGL_ASSERT(initInfo.frameCount > 0, "VkTimerQueryManager::Initialize requires non-zero frame count");
         if (initInfo.timestampValidBits == 0 || initInfo.timestampPeriodNs <= 0.0f || initInfo.slotsPerPool == 0) {
-            MGLOG_W("VkTimerQueryManager: timestamps unsupported (validBits=%u, period=%f, slots=%u)",
+            MGLOG_W_ONCE("VkTimerQueryManager: timestamps unsupported (validBits=%u, period=%f, slots=%u)",
                     initInfo.timestampValidBits, initInfo.timestampPeriodNs, initInfo.slotsPerPool);
             return false;
         }
@@ -35,7 +35,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         for (auto& poolState : m_pools) {
             const VkResult result = vkCreateQueryPool(m_device, &poolInfo, nullptr, &poolState.pool);
             if (result != VK_SUCCESS) {
-                MGLOG_E("VkTimerQueryManager: vkCreateQueryPool failed with %s", VkResultToString(result));
+                MGLOG_E_ONCE("VkTimerQueryManager: vkCreateQueryPool failed with %s", VkResultToString(result));
                 Shutdown();
                 return false;
             }
@@ -90,7 +90,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         auto& poolState = m_pools[frameIndex];
         if (poolState.cursor >= m_slotsPerPool) {
             if (!poolState.exhaustionWarned) {
-                MGLOG_W("VkTimerQueryManager: frame %u timestamp pool exhausted (%u slots); further timer queries "
+                MGLOG_W_ONCE("VkTimerQueryManager: frame %u timestamp pool exhausted (%u slots); further timer queries "
                         "this frame fall back to the frontend path",
                         frameIndex, m_slotsPerPool);
                 poolState.exhaustionWarned = true;
@@ -120,7 +120,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             m_device, m_pools[record.poolIndex].pool, record.slot, 1, sizeof(resultWithAvailability),
             resultWithAvailability, sizeof(Uint64), VK_QUERY_RESULT_64_BIT | VK_QUERY_RESULT_WITH_AVAILABILITY_BIT);
         if (result != VK_SUCCESS && result != VK_NOT_READY) {
-            MGLOG_E("VkTimerQueryManager: vkGetQueryPoolResults failed with %s", VkResultToString(result));
+            MGLOG_E_ONCE("VkTimerQueryManager: vkGetQueryPoolResults failed with %s", VkResultToString(result));
             return false;
         }
         if (resultWithAvailability[1] == 0) {

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,
@@ -224,6 +229,18 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                                     GLint dstY, GLint width, GLint height, VkImageLayout srcRestoreLayout,
                                     VkImageLayout dstRestoreLayout, Bool stencilAspect);
         static SizeT GetReadbackTexelSize(VkFormat sourceFormat);
+        // Map a GL bottom-left-origin rectangle into the display-oriented swapchain image.
+        // Quarter-turn surface transforms swap the copy extent's axes.
+        static Bool MapDefaultFramebufferReadbackRect(GLint x, GLint y, GLsizei width, GLsizei height,
+                                                      VkExtent2D imageExtent,
+                                                      VkSurfaceTransformFlagBitsKHR preTransform,
+                                                      VkOffset2D* imageOffset, VkExtent2D* imageCopyExtent);
+        // Reorder a tightly packed block copied with MapDefaultFramebufferReadbackRect back into
+        // GL row order. The input block has swapped dimensions for 90/270 degree transforms.
+        static Bool RemapDefaultFramebufferReadback(const Uint8* rawPixels, Uint32 logicalWidth,
+                                                    Uint32 logicalHeight,
+                                                    VkSurfaceTransformFlagBitsKHR preTransform,
+                                                    SizeT texelSize, Uint8* outPixels);
         static Bool ConvertReadbackPixels(const Uint8* sourcePixels, VkFormat sourceFormat,
                                           GLsizei width, GLsizei height, GLenum destinationFormat,
                                           GLenum destinationType, SizeT destinationRowStride,
@@ -293,6 +310,12 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // The samplerAnisotropy device feature was granted, so GL_TEXTURE_MAX_ANISOTROPY_EXT is
         // honored rather than accepted-and-ignored.
         Bool IsSamplerAnisotropySupported() const { return m_samplerAnisotropyFeatureEnabled; }
+        // ARB_base_instance extends indirect command records with a non-zero firstInstance and
+        // requires gl_InstanceID to remain zero-based. Vulkan needs both features to honor that
+        // complete contract: one legalizes the command word, the other enables the shader rebase.
+        Bool IsNonZeroIndirectBaseInstanceSupported() const {
+            return m_drawIndirectFirstInstanceFeatureEnabled && m_shaderDrawParametersFeatureEnabled;
+        }
         // Ensures the frame command buffer is recording (same lazy pattern as
         // SetupDraw) and writes a bottom-of-pipe timestamp into the current
         // frame's pool. Null when unsupported or the pool is exhausted.
@@ -363,6 +386,31 @@ 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 +493,13 @@ 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
+        // Whether the loader exposes VK_EXT_headless_surface, detected once in
-        // VK_EXT_headless_surface at all. Detected once in CreateInstance() from the
+        // CreateInstance() from the enumerated instance extensions. On desktop an
-        // enumerated instance extensions; when false, CreateSurface() falls back to a
+        // offscreen surface REQUIRES it: false is a clean, loud bring-up failure, never
-        // hidden Xlib window instead of vkCreateHeadlessSurfaceEXT.
+        // 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
@@ -508,7 +554,20 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         Bool m_samplerAnisotropyFeatureEnabled = false;
         Bool m_shaderDrawParametersExtensionEnabled = false;
         Bool m_shaderDrawParametersFeatureEnabled = false;
+        // Native subgroup topology, queried at device creation for the compute-module
+        // subgroup repairs (SubgroupSupportPolicy.h) and the REQUIRE_FULL_SUBGROUPS
+        // stage flag; 0 / false when the device has no usable compute subgroups or
+        // MOBILEGL_DISABLE_SUBGROUP forced them off.
+        Uint32 m_nativeSubgroupSize = 0;
+        Bool m_nativeSubgroupSupported = false;
+        Bool m_computeFullSubgroupsFeatureEnabled = false;
+        // VkPhysicalDeviceSubgroupSizeControlProperties::maxComputeWorkgroupSubgroups;
+        // 0 when the extension (and therefore the full-subgroups flag) is unavailable.
+        Uint32 m_maxComputeWorkgroupSubgroups = 0;
         Bool m_unformattedFloatStorageImagesEnabled = false;
+        // Set only after descriptor-indexing feature AND property queries prove that
+        // update-after-bind is legal for every descriptor category this renderer emits.
+        ProgramFactory::UpdateAfterBindLimits m_updateAfterBindLimits{};
         // fillModeNonSolid gates VK_POLYGON_MODE_LINE/_POINT (glPolygonMode); independentBlend gates
         // per-draw-buffer color write masks (glColorMaski). Both are cached at device creation and
         // drive a runtime fallback when the device lacks them.
@@ -519,6 +578,12 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // needs no feature). Both cached at device creation and drive a hard-fail-at-draw when absent.
         Bool m_dualSrcBlendFeatureEnabled = false;
         Bool m_primitiveTopologyListRestartFeatureEnabled = false;
+        // multiViewport gates rasterizing into more than one of ARB_viewport_array's 16 viewports
+        // (gl_ViewportIndex). m_maxRasterizableViewports is min(MAX_VIEWPORTS, device limit), or 1
+        // when the feature is off, and is the viewportCount a gl_ViewportIndex-writing pipeline
+        // declares - it is NOT what GL_MAX_VIEWPORTS reports, which is the frontend state width.
+        Bool m_multiViewportFeatureEnabled = false;
+        Uint32 m_maxRasterizableViewports = 1;
         // Union of shader stages sampled-read barriers may name; built at device creation
         // because geometry/tessellation stage bits are invalid in a barrier when their
         // feature is off (VUID-vkCmdPipelineBarrier-srcStageMask-04090/-04091), and
@@ -742,9 +807,26 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         Uint32 m_lastLodProgramVersion = 0;
         Uint64 m_lastLodBindGeneration = 0;
         Uint64 m_lastLodParamsSum = 0;
+        // Sampling-resolution generation at probe time. The probe reads the effective
+        // sampler's filters/aniso/LOD range, whose setters bump only this counter -
+        // the params-version sum above never moves for them.
+        Uint64 m_lastLodSamplingGeneration = 0;
         ProgramFactory::CompileOptionFlags m_lastLodBaseFlags = {};
         ProgramFactory::CompileOptionFlags m_lastLodResultFlags = {};
 
+        // Does the current program's vertex stage declare the BaseVertex builtin? A property
+        // of the program's SPIR-V, so (lifetime id, backend-state version) is the whole key.
+        //
+        // Memoized rather than re-asked because asking means resolving the UN-zeroed program
+        // variant, and a program that only ever draws non-indexed would then compile a variant
+        // no draw uses AND re-stamp its use every draw, so the idle sweep could never retire
+        // it. With the memo the answer is known before the first lookup and only the variant
+        // the draw actually needs is resolved.
+        Bool m_lastBaseVertexQueryValid = false;
+        Uint64 m_lastBaseVertexProgramLifetimeId = 0;
+        Uint32 m_lastBaseVertexProgramVersion = 0;
+        Bool m_lastBaseVertexReads = false;
+
         // Snapshot behind TrySetupDrawFastPath. Values only: the program and
         // render-pass caches are open-addressing maps whose entries move on
         // insert, so no pointers into them are cached; the pipeline handle is
@@ -766,6 +848,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             Uint64 vaoLifetimeId = 0;
             Uint32 vaoConfigVersion = 0;
             const void* drawFbo = nullptr;
+            // Never-reused lifetime id beside the raw pointer + Uint16 version: a
+            // deleted FBO recycled at the same address with the same fresh version
+            // count would otherwise compare equal (same ABA as the render-pass
+            // manager's fast-path memo).
+            Uint64 drawFboLifetimeId = 0;
             Uint16 fboVersion = 0;
             Bool drawFboIsDefault = false;
             Uint renderStateVersion = 0;
@@ -789,6 +876,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             // re-resolve just the pipeline against the active pass; a change that
             // flips it must fall back to the full path's pass selection.
             Bool drawUsesDepthStencil = false;
+            // The snapshotting draw's pipeline viewportCount. A pure function of the PROGRAM
+            // (writesViewportIndexBuiltin) and of a device feature fixed at renderer init, both
+            // of which the programLifetimeId/programVersion guards above already pin - carried
+            // here so the fast path does not re-fetch the program object to re-derive it.
+            Uint32 viewportCount = 1;
             IntVec2 renderPassExtent = {0, 0};
             // colorAttachmentCount of the snapshotting draw's render pass: the
             // pipeline-state hash input, so the fast path can refresh that hash and
@@ -856,6 +948,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // already sampleable.
         Vector<VkTextureManager::TextureResource*> m_sampledResourcesScratch;
         Vector<MG_State::GLState::ITextureObject*> m_storageImageTexturesScratch;
+        Vector<UniformManager::SamplerImageFeedbackBinding> m_samplerImageFeedbackScratch;
+        Vector<UniformManager::SamplerBindingOverride> m_samplerImageBindingOverridesScratch;
         Vector<VkBuffer> m_vertexBuffersScratch;
         Vector<VkDeviceSize> m_vertexOffsetsScratch;
         Vector<VkVertexInputAttributeDescription> m_patchedAttributesScratch;
@@ -982,6 +1076,14 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             VkBuffer indexVkBuffer = VK_NULL_HANDLE;
             VkDeviceSize indexSliceOffset = 0;
             Uint64 indexFrameSerial = 0;
+            // The EBO carried a host map when the slice was recorded - the mirror of
+            // anyBufferMapped on the vertex half. A shadow-backed (non-adopted)
+            // persistent map mutates its shadow with no API call and no epoch bump, so
+            // the one-compare rescue must decline and re-run the acquire, whose
+            // SyncPersistentMappedRange is the push-down. A map taken AFTER the record
+            // is already covered: AcquirePersistentMap bumps the slice epoch for the
+            // request itself, adopted or declined.
+            Bool indexBufferMapped = false;
 
             // Bound per draw (first bindingCount elements).
             VkBuffer vkBuffers[kMaxBindings] = {};
@@ -1075,11 +1177,34 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             FrameContext::FrameData& frame,
             const MG_State::GLState::ProgramObject& program,
             const ProgramFactory::VkProgramObject& programObj);
+        // Vulkan forbids a sampled descriptor and writable storage descriptor from naming the
+        // same image subresource in one shader operation. Snapshot only the sampler side; the
+        // storage descriptor continues to name the application texture.
+        Bool PrepareSamplerImageFeedbackSnapshots(
+            FrameContext::FrameData& frame,
+            const MG_State::GLState::ProgramObject& program,
+            const ProgramFactory::VkProgramObject& programObj,
+            VkPipelineStageFlags consumerShaderStageMask);
 
         // The per-draw dynamic-state tail (viewport, scissor, blend constants, depth
         // bias, line width, stencil), gated behind one render-state-parameters-version
         // compare per command buffer - see the gate fields in DynamicStateShadow.
-        void ApplyDynamicDrawStateTail(FrameContext::FrameData& frame, const IntVec2& extent, Bool isDefaultFbo);
+        // viewportCount is the bound pipeline's declared viewport count: 1 for every program that
+        // does not write gl_ViewportIndex (the memoized fast path), otherwise the renderer's
+        // rasterizable viewport count, which takes the unmemoized array path.
+        void ApplyDynamicDrawStateTail(FrameContext::FrameData& frame, const IntVec2& extent, Bool isDefaultFbo,
+                                       Uint32 viewportCount = 1);
+        void ApplyMultiViewportDynamicState(VkCommandBuffer commandBuffer, Uint32 viewportCount, const IntVec2& extent,
+                                            VkSurfaceTransformFlagBitsKHR preTransform, Bool isDefaultFbo);
+        VkRect2D ComputeGLScissorRect(Uint32 index, const IntVec2& extent,
+                                      VkSurfaceTransformFlagBitsKHR preTransform, Bool isDefaultFbo) const;
+        // How many viewports a draw with this program rasterizes into: 1 unless the program
+        // assigns gl_ViewportIndex AND the device enabled multiViewport. Both the pipeline's
+        // baked viewportCount and the dynamic arrays come from this one answer, so they cannot
+        // disagree.
+        Uint32 ResolveDrawViewportCount(Bool programWritesViewportIndex) const {
+            return programWritesViewportIndex && m_multiViewportFeatureEnabled ? m_maxRasterizableViewports : 1u;
+        }
 
         Bool UploadAndBindVertexBuffers(VkCommandBuffer commandBuffer, const MG_State::GLState::VertexArrayObject& vao,
                                         const ProgramFactory::VkProgramObject& programObj,
@@ -1120,6 +1245,18 @@ 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_Backend/DirectVulkan/SubgroupSupportPolicy.h

@@ -0,0 +1,63 @@
+// MobileGL - MobileGL/MG_Backend/DirectVulkan/SubgroupSupportPolicy.h
+// Copyright (c) 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 <Config.h>
+#include <Includes.h>
+
+namespace MobileGL::MG_Backend::DirectVulkan {
+    // The single decision point for how DirectVulkan implements GL_KHR_shader_subgroup,
+    // shared by capability advertisement (BackendObject) and module lowering
+    // (VulkanRenderer / ProgramFactory) so the two can never disagree.
+    //
+    // Native subgroups are the implementation whenever the device has them, whatever
+    // their width - subgroup operations execute on the hardware paths they were made
+    // for. Module-level repairs keep the GL contract intact around them:
+    //   - FixIterationRPSubgroupScratchPass patches the one known pack bug: iterationRP's
+    //     prefixSumCache[32], under-declared for sub-16-lane devices (8-lane lavapipe);
+    //   - FixIterationRPBarrierPass repairs Program 203's race between two reductions
+    //     reusing that scratch, when explicitly enabled;
+    //   - DeriveNumSubgroupsPass replaces the one builtin drivers get wrong
+    //     (gl_NumSubgroups) with the value the rest of the topology implies.
+    // The 32-lane shared-memory emulation (EmulateSubgroupsPass) is a LAST RESORT for
+    // devices with no subgroup support at all, and only when the user opts in with
+    // MOBILEGL_MAGMA_EMULATE_SUBGROUP=1; it never replaces available native operations.
+
+    inline constexpr Uint32 kEmulatedSubgroupSize = 32u;
+    inline constexpr Uint32 kEmulatedSubgroupStages = GL_COMPUTE_SHADER_BIT;
+    inline constexpr Uint32 kEmulatedSubgroupFeatures =
+        GL_SUBGROUP_FEATURE_BASIC_BIT_KHR | GL_SUBGROUP_FEATURE_VOTE_BIT_KHR |
+        GL_SUBGROUP_FEATURE_ARITHMETIC_BIT_KHR | GL_SUBGROUP_FEATURE_BALLOT_BIT_KHR |
+        GL_SUBGROUP_FEATURE_SHUFFLE_BIT_KHR | GL_SUBGROUP_FEATURE_SHUFFLE_RELATIVE_BIT_KHR |
+        GL_SUBGROUP_FEATURE_CLUSTERED_BIT_KHR | GL_SUBGROUP_FEATURE_QUAD_BIT_KHR;
+
+    inline Bool ShouldEmulateSubgroups(const Bool nativeSubgroupSupported) {
+        return MG_Config::Features.MagmaEmulateSubgroup && !nativeSubgroupSupported &&
+               !MG_Config::Features.DisableSubgroup;
+    }
+
+    inline Bool ShouldFixIterationRPSubgroupScratch() {
+        // Auto is ON: the patch is fingerprint-gated to iterationRP's reduction and
+        // grows one under-declared array; every other module passes through untouched.
+        return MG_Config::Features.FixIterationRPSubgroupScratch !=
+               MG_Config::QuirkOverride::ForceOff;
+    }
+
+    inline Bool ShouldFixIterationRPBarrier() {
+        return MG_Config::Features.IterationRPFixBarrier;
+    }
+
+    inline Bool ShouldDeriveNumSubgroups() {
+        // Auto is ON: gl_NumSubgroups must agree with the gl_SubgroupID range for the GL
+        // contract to hold, and the derived ceil() value is the one the renderer can pin
+        // with REQUIRE_FULL_SUBGROUPS - the driver builtin is the value with no
+        // cross-driver guarantee (Adreno returns 1 for an 8-subgroup dispatch).
+        return MG_Config::Features.DeriveNumSubgroups != MG_Config::QuirkOverride::ForceOff;
+    }
+} // namespace MobileGL::MG_Backend::DirectVulkan

MobileGL/MG_Backend/DirectVulkan/VkIncludes.h

@@ -74,6 +74,18 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 // The context line (__VA_ARGS__ = its own format string + args) must be a SEPARATE log
 // call: appending its format to the base format while its arguments precede the base
 // arguments makes every conversion read the wrong slot (a %s pulling an int crashes).
+//
+// MGLOG_F and deliberately NOT latched. VK_VERIFY is the invariant-check macro: a Vulkan call
+// MobileGL believes it has already made legal came back non-success, which is a
+// should-never-happen state, not an expected failure mode a user hits. Those fast-fail loudly
+// and keep saying so - the log-quietness rules that latch W/E cover expected failures (driver
+// capability gaps, app misuse), not broken internal invariants. MOBILEGL_ASSERT below traps in
+// a DEBUG build; MGLOG_F is what makes the same condition visible in an INFO test run, where
+// the assert is compiled out by contract.
+//
+// A soft, recoverable failure must therefore NOT be routed through VK_VERIFY. Check the
+// VkResult directly and report it with MGLOG_E_ONCE - see VkTextureManager::SyncTextureResource,
+// where a driver legitimately refuses an image the format pre-check accepted.
 #define VK_VERIFY(expr, ...)                                                                                           \
     do {                                                                                                               \
         VkResult _vk_verify_result = (expr);                                                                           \

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/EGLImpl/EGLImpl.cpp

@@ -21,7 +21,7 @@ namespace MobileGL::MG_Impl::EGLImpl {
 
         EGLStateContext* GetState() {
             if (!MG_State::pEGLContext) {
-                MGLOG_E("pEGLContext is null. MG_State may not be initialized.");
+                MGLOG_E_ONCE("pEGLContext is null. MG_State may not be initialized.");
             }
             return MG_State::pEGLContext.get();
         }
@@ -146,7 +146,7 @@ namespace MobileGL::MG_Impl::EGLImpl {
 
         auto* backendObject = GetBackendObject(state);
         if (!backendObject) {
-            MGLOG_E("activeBackendObject not initialized!");
+            MGLOG_E_ONCE("activeBackendObject not initialized!");
             state->DestroySurface(dpy, surface);
             return EGL_NO_SURFACE;
         }
@@ -172,11 +172,11 @@ namespace MobileGL::MG_Impl::EGLImpl {
 
         auto* backendObject = GetBackendObject(state);
         if (!backendObject) {
-            MGLOG_E("activeBackendObject not initialized!");
+            MGLOG_E_ONCE("activeBackendObject not initialized!");
             return EGL_FALSE;
         }
         if (!backendObject->SwapEGLBuffers(dpy, draw)) {
-            MGLOG_E("eglSwapBuffers failed on thread=%s dpy=%p draw=%p", CurrentThreadIdString().c_str(), dpy, draw);
+            MGLOG_E_ONCE("eglSwapBuffers failed on thread=%s dpy=%p draw=%p", CurrentThreadIdString().c_str(), dpy, draw);
             state->SetError(EGL_BAD_SURFACE);
             return EGL_FALSE;
         }
@@ -211,7 +211,7 @@ namespace MobileGL::MG_Impl::EGLImpl {
 
         auto* backendObject = GetBackendObject(state);
         if (!backendObject) {
-            MGLOG_E("activeBackendObject not initialized!");
+            MGLOG_E_ONCE("activeBackendObject not initialized!");
             return EGL_FALSE;
         }
         if (!backendObject->InitializeEGLDisplay(dpy, major, minor)) {
@@ -265,7 +265,7 @@ namespace MobileGL::MG_Impl::EGLImpl {
         if (releaseCurrentRequest) {
             if (auto* backendObject = MG_Backend::pActiveBackendObject.get()) {
                 if (!backendObject->MakeEGLCurrent(dpy, draw, read, ctx)) {
-                    MGLOG_E("eglMakeCurrent release failed in backend thread=%s", threadId.c_str());
+                    MGLOG_E_ONCE("eglMakeCurrent release failed in backend thread=%s", threadId.c_str());
                     state->MakeCurrent(oldDisplay, oldDraw, oldRead, oldContext);
                     state->SetError(EGL_BAD_ACCESS);
                     return EGL_FALSE;
@@ -277,12 +277,12 @@ namespace MobileGL::MG_Impl::EGLImpl {
 
         auto* backendObject = GetBackendObject(state);
         if (!backendObject) {
-            MGLOG_E("activeBackendObject not initialized!");
+            MGLOG_E_ONCE("activeBackendObject not initialized!");
             state->MakeCurrent(oldDisplay, oldDraw, oldRead, oldContext);
             return EGL_FALSE;
         }
         if (!backendObject->MakeEGLCurrent(dpy, draw, read, ctx)) {
-            MGLOG_E("eglMakeCurrent backend attach failed thread=%s dpy=%p draw=%p read=%p ctx=%p", threadId.c_str(),
+            MGLOG_E_ONCE("eglMakeCurrent backend attach failed thread=%s dpy=%p draw=%p read=%p ctx=%p", threadId.c_str(),
                     dpy, draw, read, ctx);
             state->SetError(EGL_BAD_ACCESS);
             state->MakeCurrent(oldDisplay, oldDraw, oldRead, oldContext);
@@ -703,7 +703,7 @@ namespace MobileGL::MG_Impl::EGLImpl {
 
         auto* backendObject = GetBackendObject(state);
         if (!backendObject) {
-            MGLOG_E("activeBackendObject not initialized!");
+            MGLOG_E_ONCE("activeBackendObject not initialized!");
             state->DestroySurface(dpy, surface);
             return EGL_NO_SURFACE;
         }
@@ -726,7 +726,7 @@ namespace MobileGL::MG_Impl::EGLImpl {
         }
         auto* backendObject = GetBackendObject(state);
         if (!backendObject) {
-            MGLOG_E("activeBackendObject not initialized!");
+            MGLOG_E_ONCE("activeBackendObject not initialized!");
             return EGL_FALSE;
         }
         width = std::max<EGLint>(width, 1);
@@ -764,7 +764,7 @@ namespace MobileGL::MG_Impl::EGLImpl {
         MGLOG_D("eglGetProcAddress(%s)", name);
         void* proc = MG_Impl::GetProcAddress(name);
         if (!proc) {
-            MGLOG_W("Failed to get function: %s", name);
+            MGLOG_D("Failed to get function: %s", name);
             return nullptr;
         }
         return (__eglMustCastToProperFunctionPointerType)proc;

MobileGL/MG_Impl/GLImpl/Buffer/GL_Buffer.cpp

@@ -18,6 +18,7 @@
 #include <MG_Util/Converters/GLToStr/GLEnumConverter.h>
 #include <MG_Util/Converters/GLToMG/BufferEnumConverter.h>
 #include <MG_Util/Converters/MGToGL/BufferEnumConverter.h>
+#include <MG_Util/Texture/PixelStoreProcessor.h>
 
 namespace MobileGL::MG_Impl::GLImpl {
     namespace {
@@ -31,6 +32,8 @@ namespace MobileGL::MG_Impl::GLImpl {
             NamedBufferData,
             NamedBufferSubData,
             CopyNamedBufferSubData,
+            ClearBufferData,
+            ClearBufferSubData,
             ClearNamedBufferData,
             ClearNamedBufferSubData,
             MapBufferRange,
@@ -65,6 +68,10 @@ namespace MobileGL::MG_Impl::GLImpl {
                 return "NamedBufferSubData";
             case BufferOp::CopyNamedBufferSubData:
                 return "CopyNamedBufferSubData";
+            case BufferOp::ClearBufferData:
+                return "ClearBufferData";
+            case BufferOp::ClearBufferSubData:
+                return "ClearBufferSubData";
             case BufferOp::ClearNamedBufferData:
                 return "ClearNamedBufferData";
             case BufferOp::ClearNamedBufferSubData:
@@ -143,16 +150,6 @@ namespace MobileGL::MG_Impl::GLImpl {
                 return 0;
             }
 
-            // The pattern is replicated verbatim, which is only the whole story while the client
-            // layout already matches the internal format - the case every entry point in practice
-            // uses, and the only one the conversion machinery here can express. Say so rather than
-            // quietly writing a differently-sized pattern.
-            const SizeT sourceSize = MG_Util::GetInputBytesPerPixel(inputFormat, pixelType);
-            if (sourceSize != elementSize) {
-                MGLOG_W("%s: clear pattern is %zu bytes but internalformat 0x%X stores %zu; "
-                        "converting between them is not implemented",
-                        GetBufferOpName(op), sourceSize, internalformat, elementSize);
-            }
             return elementSize;
         }
 
@@ -194,27 +191,59 @@ namespace MobileGL::MG_Impl::GLImpl {
             return true;
         }
 
-        void ClearNamedBufferRange_State(GLuint buffer, GLenum internalformat, GLintptr offset, GLsizeiptr size,
+        Bool BuildClearPattern(GLenum internalformat, GLenum format, GLenum type, const void* data,
+                               SizeT patternSize, BufferOp op, Vector<Uint8>& pattern) {
+            const TextureInternalFormat internal = MG_Util::ConvertGLEnumToTextureInternalFormat(internalformat);
+            const TextureInputFormat inputFormat = MG_Util::ConvertGLEnumToTextureInputFormat(format);
+            const TexturePixelDataType inputType = MG_Util::ConvertGLEnumToTexturePixelDataType(type);
+
+            Vector<Uint8> zeroInput;
+            const void* inputPixel = data;
+            if (inputPixel == nullptr) {
+                const SizeT inputSize = MG_Util::GetInputBytesPerPixel(inputFormat, inputType);
+                if (inputSize == 0) {
+                    MG_State::pGLContext->RecordError(
+                        ErrorCode::InvalidValue,
+                        MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", GetBufferOpName(op),
+                                                     "format and type do not describe a source pixel."));
+                    return false;
+                }
+                zeroInput.resize(inputSize);
+                inputPixel = zeroInput.data();
+            }
+
+            if (!MG_Util::PixelStoreProcessor::ConvertOnePixelToInternal(
+                    internal, inputFormat, inputType, inputPixel, pattern)) {
+                MG_State::pGLContext->RecordError(
+                    ErrorCode::InvalidValue,
+                    MakeUnique<GenericErrorInfo>(
+                        "MG_Impl/GLImpl", GetBufferOpName(op),
+                        std::format("Cannot convert one ({}, {}) pixel into internalformat 0x{:X}.",
+                                    MG_Util::ConvertGLEnumToString(format), MG_Util::ConvertGLEnumToString(type),
+                                    internalformat)));
+                return false;
+            }
+
+            if (data == nullptr) {
+                // GL defines a null clear value as all zero bits in the destination store, while
+                // retaining the format/type validation above.
+                pattern.assign(patternSize, 0);
+            }
+            return true;
+        }
+
+        void ClearBufferRange_State(const SharedPtr<MG_State::GLState::BufferObject>& bufferObject,
+                                    GLenum internalformat, GLintptr offset, GLsizeiptr size,
                                     GLenum format, GLenum type, const void* data, BufferOp op) {
             const SizeT patternSize = GetClearPatternSize(internalformat, format, type, op);
             if (patternSize == 0) return;
-
-            auto bufferObject = GetNamedBufferObject(buffer, op);
-            if (!bufferObject) return;
             if (!ValidateBufferClearRange(bufferObject, offset, size, patternSize, op)) return;
             if (size == 0) return;
 
-            Vector<Uint8> clearData(static_cast<SizeT>(size));
+            Vector<Uint8> pattern;
-            if (data) {
+            if (!BuildClearPattern(internalformat, format, type, data, patternSize, op, pattern)) return;
-                const auto* pattern = static_cast<const Uint8*>(data);
+            bufferObject->FillSubData({pattern.data(), pattern.size()}, static_cast<SizeT>(offset),
-                for (SizeT at = 0; at < clearData.size(); at += patternSize) {
+                                      static_cast<SizeT>(size));
-                    Memcpy(clearData.data() + at, pattern, patternSize);
-                }
-            } else {
-                Memset(clearData.data(), 0, clearData.size());
-            }
-
-            bufferObject->UploadSubData({clearData.data(), clearData.size()}, static_cast<SizeT>(offset));
         }
 
         auto& GetBufferBindingSlot(BufferTarget target) {
@@ -1197,16 +1226,33 @@ namespace MobileGL::MG_Impl::GLImpl {
                                         static_cast<SizeT>(writeOffset), static_cast<SizeT>(size));
     }
 
+    void ClearBufferData_State(GLenum target, GLenum internalformat, GLenum format, GLenum type, const void* data) {
+        auto bufferObject = GetBoundBufferObject(target, BufferOp::ClearBufferData);
+        if (!bufferObject) return;
+        ClearBufferRange_State(bufferObject, internalformat, 0, static_cast<GLsizeiptr>(bufferObject->GetSize()), format,
+                               type, data, BufferOp::ClearBufferData);
+    }
+
+    void ClearBufferSubData_State(GLenum target, GLenum internalformat, GLintptr offset, GLsizeiptr size,
+                                  GLenum format, GLenum type, const void* data) {
+        auto bufferObject = GetBoundBufferObject(target, BufferOp::ClearBufferSubData);
+        if (!bufferObject) return;
+        ClearBufferRange_State(bufferObject, internalformat, offset, size, format, type, data,
+                               BufferOp::ClearBufferSubData);
+    }
+
     void ClearNamedBufferData_State(GLuint buffer, GLenum internalformat, GLenum format, GLenum type, const void* data) {
         auto bufferObject = GetNamedBufferObject(buffer, BufferOp::ClearNamedBufferData);
         if (!bufferObject) return;
-        ClearNamedBufferRange_State(buffer, internalformat, 0, static_cast<GLsizeiptr>(bufferObject->GetSize()), format,
+        ClearBufferRange_State(bufferObject, internalformat, 0, static_cast<GLsizeiptr>(bufferObject->GetSize()), format,
                                type, data, BufferOp::ClearNamedBufferData);
     }
 
     void ClearNamedBufferSubData_State(GLuint buffer, GLenum internalformat, GLintptr offset, GLsizeiptr size,
                                        GLenum format, GLenum type, const void* data) {
-        ClearNamedBufferRange_State(buffer, internalformat, offset, size, format, type, data,
+        auto bufferObject = GetNamedBufferObject(buffer, BufferOp::ClearNamedBufferSubData);
+        if (!bufferObject) return;
+        ClearBufferRange_State(bufferObject, internalformat, offset, size, format, type, data,
                                BufferOp::ClearNamedBufferSubData);
     }
 
@@ -1491,8 +1537,8 @@ 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) {
+                                                 const char* funcName, Bool hasBuffer = true) {
-        if (size <= 0) {
+        if (hasBuffer && size <= 0) {
             MG_State::pGLContext->RecordError(
                 ErrorCode::InvalidValue,
                 MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", funcName,
@@ -1527,16 +1573,27 @@ namespace MobileGL::MG_Impl::GLImpl {
                 return false;
             }
         }
-        // A transform feedback capture binding is addressed in 32-bit components, so BOTH the
+        // GL 4.6 core 6.1.1 constrains the OFFSET to a multiple of four for both
-        // offset and the size must be multiples of 4.
+        // TRANSFORM_FEEDBACK_BUFFER and ATOMIC_COUNTER_BUFFER (the atomic-counter one has no
-        if (target == GL_TRANSFORM_FEEDBACK_BUFFER && ((offset % 4) != 0 || (size % 4) != 0)) {
+        // 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 "
+                    std::format("size ({}) must be a multiple of 4 for GL_TRANSFORM_FEEDBACK_BUFFER.", size)));
-                                "GL_TRANSFORM_FEEDBACK_BUFFER.",
-                                offset, size)));
             return false;
         }
         return true;
@@ -1548,7 +1605,12 @@ 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,
@@ -1646,6 +1708,15 @@ namespace MobileGL::MG_Impl::GLImpl {
         CopyNamedBufferSubData_State(readBuffer, writeBuffer, readOffset, writeOffset, size);
     }
 
+    void ClearBufferData(GLenum target, GLenum internalformat, GLenum format, GLenum type, const void* data) {
+        ClearBufferData_State(target, internalformat, format, type, data);
+    }
+
+    void ClearBufferSubData(GLenum target, GLenum internalformat, GLintptr offset, GLsizeiptr size, GLenum format,
+                            GLenum type, const void* data) {
+        ClearBufferSubData_State(target, internalformat, offset, size, format, type, data);
+    }
+
     void ClearNamedBufferData(GLuint buffer, GLenum internalformat, GLenum format, GLenum type, const void* data) {
         ClearNamedBufferData_State(buffer, internalformat, format, type, data);
     }
@@ -1732,10 +1803,30 @@ 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 +1840,7 @@ 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/Buffer/GL_Buffer.h

@@ -27,6 +27,9 @@ namespace MobileGL::MG_Impl::GLImpl {
     void NamedBufferSubData(GLuint buffer, GLintptr offset, GLsizeiptr size, const void* data);
     void CopyNamedBufferSubData(GLuint readBuffer, GLuint writeBuffer, GLintptr readOffset, GLintptr writeOffset,
                                 GLsizeiptr size);
+    void ClearBufferData(GLenum target, GLenum internalformat, GLenum format, GLenum type, const void* data);
+    void ClearBufferSubData(GLenum target, GLenum internalformat, GLintptr offset, GLsizeiptr size, GLenum format,
+                            GLenum type, const void* data);
     void ClearNamedBufferData(GLuint buffer, GLenum internalformat, GLenum format, GLenum type, const void* data);
     void ClearNamedBufferSubData(GLuint buffer, GLenum internalformat, GLintptr offset, GLsizeiptr size, GLenum format,
                                  GLenum type, const void* data);

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) {
+    static Bool ValidateProgramForExecution(const SharedPtr<MG_State::GLState::ProgramObject>& currentProgram,
-        const auto& currentProgram = MG_State::pGLContext->GetProgramForDraw();
+                                            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;
+        // Argument and binding validation runs FIRST. Both are properties of the call and of GL
-        if (!dispatchComputeIndirect) {
+        // state, so a context whose backend cannot dispatch at all must still report the
-            MG_State::pGLContext->RecordError(
+        // argument error the spec names rather than masking every one of them with
-                ErrorCode::InvalidOperation,
+        // "unsupported" - which is what put GL_INVALID_OPERATION where
-                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", __func__,
+        // KHR-GL43.compute_shader.api-indirect expects GL_INVALID_VALUE.
-                                             "Backend does not support indirect compute dispatch."));
+        //
-            return;
-        }
-        if (!ValidateCurrentProgramForCompute(__func__)) return;
         // 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(
@@ -1152,7 +1259,15 @@ namespace MobileGL::MG_Impl::GLImpl {
                 MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", functionName, "instancecount must be non-negative."));
             return;
         }
-        if (!MG_State::pGLContext->ValidateTransformFeedbackName(id)) {
+        // "id is not the name of a transform feedback object" has to mean the same thing here
+        // as it does to glIsTransformFeedback, and the two predicates are not interchangeable:
+        // a name glGenTransformFeedbacks handed out is only reserved until it is first bound,
+        // and only the bind turns it into an object (GL 4.6 core 13.2.1). ValidateTransformFeedbackName
+        // answers the reservation question - the right one for glBindTransformFeedback, which is
+        // what turns a reserved name into an object - so using it here let a generated-but-unbound
+        // name through to the completed-span check below and raised INVALID_OPERATION where the
+        // spec asks for INVALID_VALUE. Name 0 is the default object and always drawable.
+        if (id != 0 && !MG_State::pGLContext->IsTransformFeedbackObject(id)) {
             MG_State::pGLContext->RecordError(
                 ErrorCode::InvalidValue,
                 MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", functionName,

MobileGL/MG_Impl/GLImpl/Exporting/Definitions.cpp

@@ -25,12 +25,12 @@
 #define DECLARE_GL_FUNCTION_STUB_HEAD(type, name, ...) MOBILEGL_GL_API type gl##name(__VA_ARGS__) {
 
 #define DECLARE_GL_FUNCTION_STUB_END(type, name, ...)                                                                  \
-    MGLOG_W("Stub function: %s(...)", __FUNCTION__);                                                                   \
+    MGLOG_W_ONCE("Stub function: %s(...)", __FUNCTION__);                                                                   \
     return (type)1;                                                                                                    \
     }
 
 #define DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(type, name, ...)                                                        \
-    MGLOG_W("Stub function: %s(...)", __FUNCTION__);                                                                   \
+    MGLOG_W_ONCE("Stub function: %s(...)", __FUNCTION__);                                                                   \
     }
 
 #define DECLARE_GL_FUNCTION_HEAD(type, name, ...) MOBILEGL_GL_API type gl##name(__VA_ARGS__) {
@@ -969,14 +969,14 @@ DECLARE_GL_FUNCTION_STUB_HEAD(void, VertexAttribL3dv, GLuint index, const GLdoub
 DECLARE_GL_FUNCTION_STUB_HEAD(void, VertexAttribL4dv, GLuint index, const GLdouble* v) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, VertexAttribL4dv, index, v)
 DECLARE_GL_FUNCTION_STUB_HEAD(void, VertexAttribLPointer, GLuint index, GLint size, GLenum type, GLsizei stride, const void* pointer) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, VertexAttribLPointer, index, size, type, stride, pointer)
 DECLARE_GL_FUNCTION_STUB_HEAD(void, GetVertexAttribLdv, GLuint index, GLenum pname, GLdouble* params) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, GetVertexAttribLdv, index, pname, params)
-DECLARE_GL_FUNCTION_STUB_HEAD(void, ViewportArrayv, GLuint first, GLsizei count, const GLfloat* v) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, ViewportArrayv, first, count, v)
+DECLARE_GL_FUNCTION_HEAD(void, ViewportArrayv, GLuint first, GLsizei count, const GLfloat* v) DECLARE_GL_FUNCTION_END_NO_RETURN(void, ViewportArrayv, first, count, v)
-DECLARE_GL_FUNCTION_STUB_HEAD(void, ViewportIndexedf, GLuint index, GLfloat x, GLfloat y, GLfloat w, GLfloat h) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, ViewportIndexedf, index, x, y, w, h)
+DECLARE_GL_FUNCTION_HEAD(void, ViewportIndexedf, GLuint index, GLfloat x, GLfloat y, GLfloat w, GLfloat h) DECLARE_GL_FUNCTION_END_NO_RETURN(void, ViewportIndexedf, index, x, y, w, h)
-DECLARE_GL_FUNCTION_STUB_HEAD(void, ViewportIndexedfv, GLuint index, const GLfloat* v) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, ViewportIndexedfv, index, v)
+DECLARE_GL_FUNCTION_HEAD(void, ViewportIndexedfv, GLuint index, const GLfloat* v) DECLARE_GL_FUNCTION_END_NO_RETURN(void, ViewportIndexedfv, index, v)
-DECLARE_GL_FUNCTION_STUB_HEAD(void, ScissorArrayv, GLuint first, GLsizei count, const GLint* v) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, ScissorArrayv, first, count, v)
+DECLARE_GL_FUNCTION_HEAD(void, ScissorArrayv, GLuint first, GLsizei count, const GLint* v) DECLARE_GL_FUNCTION_END_NO_RETURN(void, ScissorArrayv, first, count, v)
-DECLARE_GL_FUNCTION_STUB_HEAD(void, ScissorIndexed, GLuint index, GLint left, GLint bottom, GLsizei width, GLsizei height) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, ScissorIndexed, index, left, bottom, width, height)
+DECLARE_GL_FUNCTION_HEAD(void, ScissorIndexed, GLuint index, GLint left, GLint bottom, GLsizei width, GLsizei height) DECLARE_GL_FUNCTION_END_NO_RETURN(void, ScissorIndexed, index, left, bottom, width, height)
-DECLARE_GL_FUNCTION_STUB_HEAD(void, ScissorIndexedv, GLuint index, const GLint* v) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, ScissorIndexedv, index, v)
+DECLARE_GL_FUNCTION_HEAD(void, ScissorIndexedv, GLuint index, const GLint* v) DECLARE_GL_FUNCTION_END_NO_RETURN(void, ScissorIndexedv, index, v)
-DECLARE_GL_FUNCTION_STUB_HEAD(void, DepthRangeArrayv, GLuint first, GLsizei count, const GLdouble* v) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, DepthRangeArrayv, first, count, v)
+DECLARE_GL_FUNCTION_HEAD(void, DepthRangeArrayv, GLuint first, GLsizei count, const GLdouble* v) DECLARE_GL_FUNCTION_END_NO_RETURN(void, DepthRangeArrayv, first, count, v)
-DECLARE_GL_FUNCTION_STUB_HEAD(void, DepthRangeIndexed, GLuint index, GLdouble n, GLdouble f) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, DepthRangeIndexed, index, n, f)
+DECLARE_GL_FUNCTION_HEAD(void, DepthRangeIndexed, GLuint index, GLdouble n, GLdouble f) DECLARE_GL_FUNCTION_END_NO_RETURN(void, DepthRangeIndexed, index, n, f)
 DECLARE_GL_FUNCTION_HEAD(void, GetFloati_v, GLenum target, GLuint index, GLfloat* data) DECLARE_GL_FUNCTION_END_NO_RETURN(void, GetFloati_v, target, index, data)
 DECLARE_GL_FUNCTION_HEAD(void, GetDoublei_v, GLenum target, GLuint index, GLdouble* data) DECLARE_GL_FUNCTION_END_NO_RETURN(void, GetDoublei_v, target, index, data)
 DECLARE_GL_FUNCTION_HEAD(void, DrawArraysInstancedBaseInstance, GLenum mode, GLint first, GLsizei count, GLsizei instancecount, GLuint baseinstance) DECLARE_GL_FUNCTION_END_NO_RETURN(void, DrawArraysInstancedBaseInstance, mode, first, count, instancecount, baseinstance)
@@ -985,8 +985,8 @@ DECLARE_GL_FUNCTION_HEAD(void, DrawElementsInstancedBaseVertexBaseInstance, GLen
 DECLARE_GL_FUNCTION_STUB_HEAD(void, GetActiveAtomicCounterBufferiv, GLuint program, GLuint bufferIndex, GLenum pname, GLint* params) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, GetActiveAtomicCounterBufferiv, program, bufferIndex, pname, params)
 DECLARE_GL_FUNCTION_HEAD(void, DrawTransformFeedbackInstanced, GLenum mode, GLuint id, GLsizei instancecount) DECLARE_GL_FUNCTION_END_NO_RETURN(void, DrawTransformFeedbackInstanced, mode, id, instancecount)
 DECLARE_GL_FUNCTION_HEAD(void, DrawTransformFeedbackStreamInstanced, GLenum mode, GLuint id, GLuint stream, GLsizei instancecount) DECLARE_GL_FUNCTION_END_NO_RETURN(void, DrawTransformFeedbackStreamInstanced, mode, id, stream, instancecount)
-DECLARE_GL_FUNCTION_STUB_HEAD(void, ClearBufferData, GLenum target, GLenum internalformat, GLenum format, GLenum type, const void* data) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, ClearBufferData, target, internalformat, format, type, data)
+DECLARE_GL_FUNCTION_HEAD(void, ClearBufferData, GLenum target, GLenum internalformat, GLenum format, GLenum type, const void* data) DECLARE_GL_FUNCTION_END_NO_RETURN(void, ClearBufferData, target, internalformat, format, type, data)
-DECLARE_GL_FUNCTION_STUB_HEAD(void, ClearBufferSubData, GLenum target, GLenum internalformat, GLintptr offset, GLsizeiptr size, GLenum format, GLenum type, const void* data) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, ClearBufferSubData, target, internalformat, offset, size, format, type, data)
+DECLARE_GL_FUNCTION_HEAD(void, ClearBufferSubData, GLenum target, GLenum internalformat, GLintptr offset, GLsizeiptr size, GLenum format, GLenum type, const void* data) DECLARE_GL_FUNCTION_END_NO_RETURN(void, ClearBufferSubData, target, internalformat, offset, size, format, type, data)
 DECLARE_GL_FUNCTION_STUB_HEAD(void, GetInternalformati64v, GLenum target, GLenum internalformat, GLenum pname, GLsizei count, GLint64* params) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, GetInternalformati64v, target, internalformat, pname, count, params)
 DECLARE_GL_FUNCTION_STUB_HEAD(void, InvalidateTexSubImage, GLuint texture, GLint level, GLint xoffset, GLint yoffset, GLint zoffset, GLsizei width, GLsizei height, GLsizei depth) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, InvalidateTexSubImage, texture, level, xoffset, yoffset, zoffset, width, height, depth)
 DECLARE_GL_FUNCTION_STUB_HEAD(void, InvalidateTexImage, GLuint texture, GLint level) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, InvalidateTexImage, texture, level)
@@ -1003,9 +1003,9 @@ DECLARE_GL_FUNCTION_HEAD(void, ClearTexImage, GLuint texture, GLint level, GLenu
 DECLARE_GL_FUNCTION_HEAD(void, ClearTexSubImage, GLuint texture, GLint level, GLint xoffset, GLint yoffset, GLint zoffset, GLsizei width, GLsizei height, GLsizei depth, GLenum format, GLenum type, const void* data) DECLARE_GL_FUNCTION_END_NO_RETURN(void, ClearTexSubImage, texture, level, xoffset, yoffset, zoffset, width, height, depth, format, type, data)
 DECLARE_GL_FUNCTION_HEAD(void, BindBuffersBase, GLenum target, GLuint first, GLsizei count, const GLuint* buffers) DECLARE_GL_FUNCTION_END_NO_RETURN(void, BindBuffersBase, target, first, count, buffers)
 DECLARE_GL_FUNCTION_HEAD(void, BindBuffersRange, GLenum target, GLuint first, GLsizei count, const GLuint* buffers, const GLintptr* offsets, const GLsizeiptr* sizes) DECLARE_GL_FUNCTION_END_NO_RETURN(void, BindBuffersRange, target, first, count, buffers, offsets, sizes)
-DECLARE_GL_FUNCTION_STUB_HEAD(void, BindTextures, GLuint first, GLsizei count, const GLuint* textures) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, BindTextures, first, count, textures)
+DECLARE_GL_FUNCTION_HEAD(void, BindTextures, GLuint first, GLsizei count, const GLuint* textures) DECLARE_GL_FUNCTION_END_NO_RETURN(void, BindTextures, first, count, textures)
 DECLARE_GL_FUNCTION_HEAD(void, BindSamplers, GLuint first, GLsizei count, const GLuint* samplers) DECLARE_GL_FUNCTION_END_NO_RETURN(void, BindSamplers, first, count, samplers)
-DECLARE_GL_FUNCTION_STUB_HEAD(void, BindImageTextures, GLuint first, GLsizei count, const GLuint* textures) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, BindImageTextures, first, count, textures)
+DECLARE_GL_FUNCTION_HEAD(void, BindImageTextures, GLuint first, GLsizei count, const GLuint* textures) DECLARE_GL_FUNCTION_END_NO_RETURN(void, BindImageTextures, first, count, textures)
 DECLARE_GL_FUNCTION_HEAD(void, BindVertexBuffers, GLuint first, GLsizei count, const GLuint* buffers, const GLintptr* offsets, const GLsizei* strides) DECLARE_GL_FUNCTION_END_NO_RETURN(void, BindVertexBuffers, first, count, buffers, offsets, strides)
 DECLARE_GL_FUNCTION_STUB_HEAD(void, ClipControl, GLenum origin, GLenum depth) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, ClipControl, origin, depth)
 DECLARE_GL_FUNCTION_HEAD(void, CreateTransformFeedbacks, GLsizei n, GLuint* ids) DECLARE_GL_FUNCTION_END_NO_RETURN(void, CreateTransformFeedbacks, n, ids)
@@ -1061,7 +1061,7 @@ DECLARE_GL_FUNCTION_HEAD(void, TextureSubImage1D, GLuint texture, GLint level, G
 DECLARE_GL_FUNCTION_HEAD(void, TextureSubImage2D, GLuint texture, GLint level, GLint xoffset, GLint yoffset, GLsizei width, GLsizei height, GLenum format, GLenum type, const void* pixels) DECLARE_GL_FUNCTION_END_NO_RETURN(void, TextureSubImage2D, texture, level, xoffset, yoffset, width, height, format, type, pixels)
 DECLARE_GL_FUNCTION_HEAD(void, TextureSubImage3D, GLuint texture, GLint level, GLint xoffset, GLint yoffset, GLint zoffset, GLsizei width, GLsizei height, GLsizei depth, GLenum format, GLenum type, const void* pixels) DECLARE_GL_FUNCTION_END_NO_RETURN(void, TextureSubImage3D, texture, level, xoffset, yoffset, zoffset, width, height, depth, format, type, pixels)
 DECLARE_GL_FUNCTION_STUB_HEAD(void, CompressedTextureSubImage1D, GLuint texture, GLint level, GLint xoffset, GLsizei width, GLenum format, GLsizei imageSize, const void* data) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, CompressedTextureSubImage1D, texture, level, xoffset, width, format, imageSize, data)
-DECLARE_GL_FUNCTION_STUB_HEAD(void, CompressedTextureSubImage2D, GLuint texture, GLint level, GLint xoffset, GLint yoffset, GLsizei width, GLsizei height, GLenum format, GLsizei imageSize, const void* data) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, CompressedTextureSubImage2D, texture, level, xoffset, yoffset, width, height, format, imageSize, data)
+DECLARE_GL_FUNCTION_HEAD(void, CompressedTextureSubImage2D, GLuint texture, GLint level, GLint xoffset, GLint yoffset, GLsizei width, GLsizei height, GLenum format, GLsizei imageSize, const void* data) DECLARE_GL_FUNCTION_END_NO_RETURN(void, CompressedTextureSubImage2D, texture, level, xoffset, yoffset, width, height, format, imageSize, data)
 DECLARE_GL_FUNCTION_STUB_HEAD(void, CompressedTextureSubImage3D, GLuint texture, GLint level, GLint xoffset, GLint yoffset, GLint zoffset, GLsizei width, GLsizei height, GLsizei depth, GLenum format, GLsizei imageSize, const void* data) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, CompressedTextureSubImage3D, texture, level, xoffset, yoffset, zoffset, width, height, depth, format, imageSize, data)
 DECLARE_GL_FUNCTION_HEAD(void, CopyTextureSubImage1D, GLuint texture, GLint level, GLint xoffset, GLint x, GLint y, GLsizei width) DECLARE_GL_FUNCTION_END_NO_RETURN(void, CopyTextureSubImage1D, texture, level, xoffset, x, y, width)
 DECLARE_GL_FUNCTION_HEAD(void, CopyTextureSubImage2D, GLuint texture, GLint level, GLint xoffset, GLint yoffset, GLint x, GLint y, GLsizei width, GLsizei height) DECLARE_GL_FUNCTION_END_NO_RETURN(void, CopyTextureSubImage2D, texture, level, xoffset, yoffset, x, y, width, height)
@@ -1849,7 +1849,7 @@ DECLARE_GL_FUNCTION_STUB_HEAD(void, CompressedTextureImage3DEXT, GLuint texture,
 DECLARE_GL_FUNCTION_STUB_HEAD(void, CompressedTextureImage2DEXT, GLuint texture, GLenum target, GLint level, GLenum internalformat, GLsizei width, GLsizei height, GLint border, GLsizei imageSize, const void* bits) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, CompressedTextureImage2DEXT, texture, target, level, internalformat, width, height, border, imageSize, bits)
 DECLARE_GL_FUNCTION_STUB_HEAD(void, CompressedTextureImage1DEXT, GLuint texture, GLenum target, GLint level, GLenum internalformat, GLsizei width, GLint border, GLsizei imageSize, const void* bits) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, CompressedTextureImage1DEXT, texture, target, level, internalformat, width, border, imageSize, bits)
 DECLARE_GL_FUNCTION_STUB_HEAD(void, CompressedTextureSubImage3DEXT, GLuint texture, GLenum target, GLint level, GLint xoffset, GLint yoffset, GLint zoffset, GLsizei width, GLsizei height, GLsizei depth, GLenum format, GLsizei imageSize, const void* bits) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, CompressedTextureSubImage3DEXT, texture, target, level, xoffset, yoffset, zoffset, width, height, depth, format, imageSize, bits)
-DECLARE_GL_FUNCTION_STUB_HEAD(void, CompressedTextureSubImage2DEXT, GLuint texture, GLenum target, GLint level, GLint xoffset, GLint yoffset, GLsizei width, GLsizei height, GLenum format, GLsizei imageSize, const void* bits) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, CompressedTextureSubImage2DEXT, texture, target, level, xoffset, yoffset, width, height, format, imageSize, bits)
+DECLARE_GL_FUNCTION_HEAD(void, CompressedTextureSubImage2DEXT, GLuint texture, GLenum target, GLint level, GLint xoffset, GLint yoffset, GLsizei width, GLsizei height, GLenum format, GLsizei imageSize, const void* bits) DECLARE_GL_FUNCTION_END_NO_RETURN(void, CompressedTextureSubImage2D, texture, level, xoffset, yoffset, width, height, format, imageSize, bits)
 DECLARE_GL_FUNCTION_STUB_HEAD(void, CompressedTextureSubImage1DEXT, GLuint texture, GLenum target, GLint level, GLint xoffset, GLsizei width, GLenum format, GLsizei imageSize, const void* bits) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, CompressedTextureSubImage1DEXT, texture, target, level, xoffset, width, format, imageSize, bits)
 DECLARE_GL_FUNCTION_STUB_HEAD(void, GetCompressedTextureImageEXT, GLuint texture, GLenum target, GLint lod, void* img) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, GetCompressedTextureImageEXT, texture, target, lod, img)
 DECLARE_GL_FUNCTION_STUB_HEAD(void, CompressedMultiTexImage3DEXT, GLenum texunit, GLenum target, GLint level, GLenum internalformat, GLsizei width, GLsizei height, GLsizei depth, GLint border, GLsizei imageSize, const void* bits) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, CompressedMultiTexImage3DEXT, texunit, target, level, internalformat, width, height, depth, border, imageSize, bits)
@@ -2585,7 +2585,7 @@ DECLARE_GL_FUNCTION_STUB_HEAD(void, BindTransformFeedbackNV, GLenum target, GLui
 DECLARE_GL_FUNCTION_STUB_HEAD(void, DeleteTransformFeedbacksNV, GLsizei n, const GLuint* ids) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, DeleteTransformFeedbacksNV, n, ids)
 DECLARE_GL_FUNCTION_STUB_HEAD(void, GenTransformFeedbacksNV, GLsizei n, GLuint* ids) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, GenTransformFeedbacksNV, n, ids)
 MOBILEGL_GL_API GLboolean glIsTransformFeedbackNV(GLuint id) {
-    MGLOG_W("Stub function: %s(...)", __FUNCTION__);
+    MGLOG_W_ONCE("Stub function: %s(...)", __FUNCTION__);
     return GL_FALSE;
 }
 DECLARE_GL_FUNCTION_STUB_HEAD(void, PauseTransformFeedbackNV, void) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, PauseTransformFeedbackNV, )
@@ -3181,5 +3181,5 @@ MOBILEGL_GL_API void glVertexAttribDivisorARB(GLuint index, GLuint divisor) {
 }
 
 MOBILEGL_GL_API void glWindowRectanglesEXT(GLenum mode, GLsizei count, const GLint* box) {
-    MGLOG_W("Stub function: %s(...)", __FUNCTION__);
+    MGLOG_W_ONCE("Stub function: %s(...)", __FUNCTION__);
 }

MobileGL/MG_Impl/GLImpl/Framebuffer/GL_Framebuffer.cpp

@@ -547,7 +547,7 @@ namespace MobileGL::MG_Impl::GLImpl {
                                       GLint dstX1, GLint dstY1, GLbitfield mask, GLenum filter) {
         auto blitNamedFramebuffer = MG_Backend::gBackendFunctionsTable.GL.BlitNamedFramebuffer;
         if (!blitNamedFramebuffer) {
-            MGLOG_E("glBlitNamedFramebuffer skipped: backend does not implement explicit framebuffer blit.");
+            MGLOG_E_ONCE("glBlitNamedFramebuffer skipped: backend does not implement explicit framebuffer blit.");
             return;
         }
         blitNamedFramebuffer(readFramebuffer, drawFramebuffer, srcX0, srcY0, srcX1, srcY1, dstX0, dstY0, dstX1,
@@ -558,7 +558,7 @@ namespace MobileGL::MG_Impl::GLImpl {
                                          GLenum buffer, GLint drawbuffer, const GLfloat* value) {
         auto clearNamedFramebufferfv = MG_Backend::gBackendFunctionsTable.GL.ClearNamedFramebufferfv;
         if (!clearNamedFramebufferfv) {
-            MGLOG_E("glClearNamedFramebufferfv skipped: backend does not implement explicit framebuffer clear.");
+            MGLOG_E_ONCE("glClearNamedFramebufferfv skipped: backend does not implement explicit framebuffer clear.");
             return;
         }
         clearNamedFramebufferfv(framebuffer, buffer, drawbuffer, value);
@@ -568,7 +568,7 @@ namespace MobileGL::MG_Impl::GLImpl {
                                          GLenum buffer, GLint drawbuffer, GLfloat depth, GLint stencil) {
         auto clearNamedFramebufferfi = MG_Backend::gBackendFunctionsTable.GL.ClearNamedFramebufferfi;
         if (!clearNamedFramebufferfi) {
-            MGLOG_E("glClearNamedFramebufferfi skipped: backend does not implement explicit framebuffer clear.");
+            MGLOG_E_ONCE("glClearNamedFramebufferfi skipped: backend does not implement explicit framebuffer clear.");
             return;
         }
         clearNamedFramebufferfi(framebuffer, buffer, drawbuffer, depth, stencil);
@@ -578,7 +578,7 @@ namespace MobileGL::MG_Impl::GLImpl {
                                          GLenum buffer, GLint drawbuffer, const GLint* value) {
         auto clearNamedFramebufferiv = MG_Backend::gBackendFunctionsTable.GL.ClearNamedFramebufferiv;
         if (!clearNamedFramebufferiv) {
-            MGLOG_E("glClearNamedFramebufferiv skipped: backend does not implement explicit framebuffer clear.");
+            MGLOG_E_ONCE("glClearNamedFramebufferiv skipped: backend does not implement explicit framebuffer clear.");
             return;
         }
         clearNamedFramebufferiv(framebuffer, buffer, drawbuffer, value);
@@ -588,7 +588,7 @@ namespace MobileGL::MG_Impl::GLImpl {
                                           GLenum buffer, GLint drawbuffer, const GLuint* value) {
         auto clearNamedFramebufferuiv = MG_Backend::gBackendFunctionsTable.GL.ClearNamedFramebufferuiv;
         if (!clearNamedFramebufferuiv) {
-            MGLOG_E("glClearNamedFramebufferuiv skipped: backend does not implement explicit framebuffer clear.");
+            MGLOG_E_ONCE("glClearNamedFramebufferuiv skipped: backend does not implement explicit framebuffer clear.");
             return;
         }
         clearNamedFramebufferuiv(framebuffer, buffer, drawbuffer, value);
@@ -620,6 +620,34 @@ namespace MobileGL::MG_Impl::GLImpl {
         return std::max(MG_Backend::pActiveBackendObject->GetDynamicParameters().MaxSamples, 1);
     }
 
+    // GL_MAX_SAMPLES is the ceiling over all formats; an integer format has its own, lower
+    // one (GL_MAX_INTEGER_SAMPLES) and GL 4.6 core 9.2.4 makes exceeding it INVALID_OPERATION.
+    // The multisample TEXTURE path already resolves the limit per format
+    // (GL_Texture.cpp, GetMaxTextureSamplesForFormat); renderbuffers only ever compared
+    // against GL_MAX_SAMPLES, so on a driver where the two differ - Adreno reports
+    // GL_MAX_SAMPLES 4 and GL_MAX_INTEGER_SAMPLES 1 - an integer renderbuffer accepted a
+    // sample count the format cannot deliver, and said GL_NO_ERROR about it.
+    Int GetMaxRenderbufferSamplesForFormat_State(TextureInternalFormat format) {
+        if (MG_Backend::pActiveBackendObject == nullptr) {
+            return std::numeric_limits<Int>::max();
+        }
+        const auto& dynamicParameters = MG_Backend::pActiveBackendObject->GetDynamicParameters();
+
+        GLenum normalizedInternalFormat = MG_Util::ConvertTextureInternalFormatToGLEnum(format);
+        GLenum normalizedFormat = GL_RGBA;
+        GLenum normalizedType = GL_UNSIGNED_BYTE;
+        MG_Util::TextureFormatProcessor::NormalizePixelFormat(normalizedInternalFormat,
+                                                              PixelFormatNormalizeOptionBit::None,
+                                                              &normalizedInternalFormat, &normalizedFormat,
+                                                              &normalizedType);
+        const Bool isIntegerFormat = normalizedFormat == GL_RED_INTEGER || normalizedFormat == GL_RG_INTEGER ||
+                                     normalizedFormat == GL_RGB_INTEGER || normalizedFormat == GL_RGBA_INTEGER;
+        if (!isIntegerFormat) {
+            return GetMaxRenderbufferSamples_State();
+        }
+        return std::max(dynamicParameters.MaxIntegerSamples, 1);
+    }
+
     Bool ValidateRenderbufferStorageSize_State(GLsizei width, GLsizei height, const char* caller) {
         if (width < 0 || height < 0) {
             MG_State::pGLContext->RecordError(
@@ -641,7 +669,7 @@ namespace MobileGL::MG_Impl::GLImpl {
         return true;
     }
 
-    Bool ValidateRenderbufferStorageSamples_State(GLsizei samples, const char* caller) {
+    Bool ValidateRenderbufferStorageSamples_State(GLsizei samples, TextureInternalFormat format, const char* caller) {
         if (samples < 0) {
             MG_State::pGLContext->RecordError(
                 ErrorCode::InvalidValue,
@@ -649,9 +677,10 @@ namespace MobileGL::MG_Impl::GLImpl {
             return false;
         }
 
-        const Int maxSamples = GetMaxRenderbufferSamples_State();
+        // TODO: Resolve the remaining per-internalformat renderbuffer sample limits once
+        // glGetInternalformativ is backed; integer formats are handled below.
+        const Int maxSamples = GetMaxRenderbufferSamplesForFormat_State(format);
         if (samples > maxSamples) {
-            // TODO: Use per-internalformat renderbuffer sample limits once glGetInternalformativ is backed.
             // GL 4.6 core 9.2.4 makes asking for more samples than the format supports
             // INVALID_OPERATION, not INVALID_VALUE - the count is well formed, this format just
             // cannot deliver it. Only a negative count is INVALID_VALUE.
@@ -659,7 +688,7 @@ namespace MobileGL::MG_Impl::GLImpl {
                 ErrorCode::InvalidOperation,
                 MakeUnique<GenericErrorInfo>(
                     "MG_Impl/GLImpl", caller,
-                    std::format("Sample count {} exceeds GL_MAX_SAMPLES ({}).", samples, maxSamples)));
+                    std::format("Sample count {} exceeds this format's sample limit ({}).", samples, maxSamples)));
             return false;
         }
         return true;
@@ -684,7 +713,7 @@ namespace MobileGL::MG_Impl::GLImpl {
         TextureInternalFormat format = MG_Util::ConvertGLEnumToTextureInternalFormat(internalformat);
         if (!TextureImpl::ValidateTextureInternalFormat(format)) return;
 
-        if (!ValidateRenderbufferStorageSamples_State(samples, kCaller)) return;
+        if (!ValidateRenderbufferStorageSamples_State(samples, format, kCaller)) return;
         if (!ValidateRenderbufferStorageSize_State(width, height, kCaller)) return;
 
         renderbufferObject->AllocateStorage({width, height});
@@ -931,7 +960,8 @@ namespace MobileGL::MG_Impl::GLImpl {
 
         TextureInternalFormat format = MG_Util::ConvertGLEnumToTextureInternalFormat(internalformat);
         if (!TextureImpl::ValidateTextureInternalFormat(format)) return;
-        if (!ValidateRenderbufferStorageSamples_State(samples, "NamedRenderbufferStorageMultisample_State")) return;
+        if (!ValidateRenderbufferStorageSamples_State(samples, format, "NamedRenderbufferStorageMultisample_State"))
+            return;
         if (!ValidateRenderbufferStorageSize_State(width, height, "NamedRenderbufferStorageMultisample_State")) return;
 
         renderbufferObject->AllocateStorage({width, height});

MobileGL/MG_Impl/GLImpl/Getter/GL_Getter.cpp

@@ -16,6 +16,7 @@
 #include <MG_State/GLState/ErrorState/ErrorInfo.h>
 #include <MG_Util/Converters/GLToStr/GLEnumConverter.h>
 #include <MG_Util/Converters/GLToMG/BufferEnumConverter.h>
+#include <MG_Util/Converters/GLToMG/RenderStateEnumConverter.h>
 #include <MG_Util/Converters/MGToGL/FramebufferEnumConverter.h>
 #include <MG_Util/Converters/MGToGL/ErrorCodeConverter.h>
 #include <MG_Util/Converters/MGToGL/TextureEnumConverter.h>
@@ -27,6 +28,11 @@
 #include <MG_Backend/BackendObjects.h>
 
 namespace MobileGL::MG_Impl::GLImpl {
+    // Declared rather than #included from GL_RenderState.h on purpose: that header also declares
+    // a free function named BlendEquation, which would hide the ::MobileGL::BlendEquation enum
+    // this file's blend-state queries name unqualified.
+    GLboolean IsEnabledi(GLenum target, GLuint index);
+
     namespace {
         enum class IndexedBufferQueryKind {
             Binding,
@@ -51,6 +57,15 @@ namespace MobileGL::MG_Impl::GLImpl {
         constexpr GLint kFrontendMaxTessControlAtomicCounters = 0;
         constexpr GLint kFrontendMaxTessEvaluationAtomicCounters = 0;
         constexpr GLint kFrontendMaxVertexAtomicCounters = 0;
+        // Zero counters means zero buffers to hold them. These have to be ANSWERED rather than
+        // left to the default INVALID_ENUM: a well-behaved application queries the limit exactly
+        // to find out that the stage cannot do this, and an error instead both leaves its output
+        // untouched (so it reads uninitialised memory and may conclude the opposite) and leaves a
+        // GL error pending that surfaces at whatever unrelated call checks next.
+        constexpr GLint kFrontendMaxGeometryAtomicCounterBuffers = 0;
+        constexpr GLint kFrontendMaxTessControlAtomicCounterBuffers = 0;
+        constexpr GLint kFrontendMaxTessEvaluationAtomicCounterBuffers = 0;
+        constexpr GLint kFrontendMaxVertexAtomicCounterBuffers = 0;
         // One atomic counter is a uint, and a buffer never has to hold more counters than the
         // combined limit the frontend advertises. GL 4.6 table 23.63 floors this at 32 bytes.
         constexpr GLint kFrontendMaxAtomicCounterBufferSize =
@@ -174,6 +189,30 @@ namespace MobileGL::MG_Impl::GLImpl {
             return frontendCount;
         }
 
+        // A per-stage or combined BLOCK count is an amount of indexed binding points an
+        // application will occupy, and GL 4.6 table 23.64 orders the two accordingly:
+        // MAX_UNIFORM_BUFFER_BINDINGS >= MAX_COMBINED_UNIFORM_BLOCKS >= every per-stage count,
+        // and the same for the shader-storage family. The two families are answered from
+        // unrelated places here - frontend constants, backend dynamic parameters, and a few
+        // hard-coded TODOs - so nothing kept them ordered, and a backend that reports Vulkan
+        // descriptor-indexing counts advertised 256 compute uniform blocks over 36 binding
+        // points. KHR-GL44.multi_bind.dispatch_bind_buffers_base reads the block count and binds
+        // that many buffers in ONE glBindBuffersBase, which is then INVALID_OPERATION before it
+        // binds anything. Clamping is the only direction available: the binding count is the
+        // capacity of the state layer's indexed-binding array, not a number we may inflate.
+        GLint ClampBlockCountToBindingPoints(GLint blockCount, BufferTarget bufferTarget) {
+            const GLint bindingPoints = static_cast<GLint>(GetIndexedBufferQueryPointCount(bufferTarget));
+            return std::min(std::max(blockCount, 0), bindingPoints);
+        }
+
+        GLint ClampUniformBlockCount(GLint blockCount) {
+            return ClampBlockCountToBindingPoints(blockCount, BufferTarget::Uniform);
+        }
+
+        GLint ClampStorageBlockCount(GLint blockCount) {
+            return ClampBlockCountToBindingPoints(blockCount, BufferTarget::ShaderStorage);
+        }
+
         bool TryDecodeDrawBufferQuery(GLenum pname, SizeT& drawBufferIndex) {
             if (pname == GL_DRAW_BUFFER) {
                 drawBufferIndex = 0;
@@ -306,26 +345,70 @@ namespace MobileGL::MG_Impl::GLImpl {
             return sampler ? static_cast<GLint>(sampler->GetExternalIndex()) : 0;
         }
 
-        // The ARB_viewport_array indexed rectangles. MobileGL keeps exactly one viewport, one
+        // The ARB_viewport_array indexed rectangles. Each of these is genuinely per-viewport
-        // scissor box and one depth range, so every in-range index answers with that single
+        // frontend state (RenderStateParameters::Viewports / ScissorBoxes / DepthRanges), so the
-        // value - but it has to come from the frontend state the non-indexed getters read.
+        // indexed getters must read the indexed storage - the generic path at the bottom of
-        // The generic path at the bottom of GetIntegeri_v is a raw backend passthrough that
+        // GetIntegeri_v is a raw backend passthrough that has no case for them and returned
-        // has no case for these, so routing them through it returned zeros.
+        // zeros, and routing them to the NON-indexed getter (what this used to do) answered every
+        // index with viewport 0's value, which is what
+        // KHR-GL43.viewport_array.{viewport,scissor,depth_range}_api caught.
         Bool IsIndexedViewportQuery(GLenum target) {
             return target == GL_VIEWPORT || target == GL_SCISSOR_BOX || target == GL_DEPTH_RANGE;
         }
 
-        // ARB_viewport_array: `index` selects a viewport and MAX_VIEWPORTS bounds it.
+        // Component count of an indexed viewport-array query, so every width of getter writes the
+        // caller's whole buffer instead of just element 0 (GL 4.6 core 22.1).
+        GLsizei IndexedViewportQueryComponents(GLenum target) {
+            return target == GL_DEPTH_RANGE ? 2 : 4;
+        }
+
+        // ARB_viewport_array: `index` selects a viewport and MAX_VIEWPORTS bounds it. The bound is
+        // the frontend's own state width, which is also exactly what GL_MAX_VIEWPORTS reports -
+        // taking it from the backend caps instead would let a device limit of 1 (a Vulkan device
+        // without the multiViewport feature) make index 1 illegal even though the state exists.
         Bool ValidateViewportQueryIndex(GLuint index, const char* caller) {
-            GLint maxViewports = 0;
+            if (index < RenderStateParameters::MAX_VIEWPORTS) return true;
-            GetIntegerv(GL_MAX_VIEWPORTS, &maxViewports);
-            if (index < static_cast<GLuint>(std::max(maxViewports, 1))) return true;
             MG_State::pGLContext->RecordError(
                 ErrorCode::InvalidValue,
                 MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", caller, "Viewport index is out of range."));
             return false;
         }
 
+        // The indexed viewport/scissor/depth-range state as floats, which is the widest lossless
+        // shape MobileGL stores (the viewport really is float state; the scissor box is integral
+        // and well inside float's exact range, and every depth range is in [0, 1]). Every indexed
+        // getter width funnels through this so they can never disagree with each other.
+        void ReadIndexedViewportStateFloat(GLenum target, GLuint index, GLfloat* out) {
+            switch (target) {
+            case GL_VIEWPORT: {
+                const FloatVec4& viewport = MG_State::pGLContext->GetViewportIndexed(index);
+                out[0] = viewport.x();
+                out[1] = viewport.y();
+                out[2] = viewport.z();
+                out[3] = viewport.w();
+                return;
+            }
+            case GL_SCISSOR_BOX: {
+                const IntVec4& box = MG_State::pGLContext->GetScissorBoxIndexed(index);
+                out[0] = static_cast<GLfloat>(box.x());
+                out[1] = static_cast<GLfloat>(box.y());
+                out[2] = static_cast<GLfloat>(box.z());
+                out[3] = static_cast<GLfloat>(box.w());
+                return;
+            }
+            case GL_DEPTH_RANGE: {
+                const FloatVec2& range = MG_State::pGLContext->GetDepthRangeIndexed(index);
+                out[0] = range.x();
+                out[1] = range.y();
+                return;
+            }
+            default:
+                MOBILEGL_ASSERT(false, "ReadIndexedViewportStateFloat: unexpected target 0x%x",
+                                static_cast<Uint32>(target));
+                return;
+            }
+        }
+
         void CopyIntsToBooleans(const GLint* src, SizeT count, GLboolean* dst) {
             for (SizeT i = 0; i < count; ++i) {
                 dst[i] = src[i] ? GL_TRUE : GL_FALSE;
@@ -350,7 +433,7 @@ namespace MobileGL::MG_Impl::GLImpl {
 
         MGLOG_D("glGetString, name: %s", MG_Util::ConvertGLEnumToString(name).c_str());
         if (!activeBackendObject) {
-            MGLOG_E("activeBackendObject is not initialized!");
+            MGLOG_E_ONCE("activeBackendObject is not initialized!");
             return (GLubyte*)"Unknown";
         }
 
@@ -409,7 +492,7 @@ namespace MobileGL::MG_Impl::GLImpl {
 
         const auto& activeBackendObject = MG_Backend::pActiveBackendObject;
         if (!activeBackendObject) {
-            MGLOG_E("activeBackendObject is not initialized!");
+            MGLOG_E_ONCE("activeBackendObject is not initialized!");
             return (GLubyte*)"Unknown";
         }
         const auto& rendererInfo = activeBackendObject->GetRendererInfo();
@@ -596,6 +679,17 @@ namespace MobileGL::MG_Impl::GLImpl {
             params[1] = dynamicParameters.ViewportBoundsRangeMax;
             return;
         }
+        // Viewport 0's rectangle, verbatim. Falling through to the integer width below would
+        // round the fractional rectangle a glViewportIndexedf(0, ...) is allowed to set, and
+        // glGetFloatv(GL_VIEWPORT) is a lossless query of float state.
+        case GL_VIEWPORT: {
+            const FloatVec4& viewport = MG_State::pGLContext->GetViewportIndexed(0);
+            params[0] = viewport.x();
+            params[1] = viewport.y();
+            params[2] = viewport.z();
+            params[3] = viewport.w();
+            return;
+        }
         case GL_MIN_FRAGMENT_INTERPOLATION_OFFSET:
         case GL_MAX_FRAGMENT_INTERPOLATION_OFFSET:
         case GL_FRAGMENT_INTERPOLATION_OFFSET_BITS: {
@@ -722,10 +816,14 @@ 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());
+                *data = static_cast<GLint>(range.end - range.start);
-                const auto end = std::min(range.end, bufferObject->GetSize());
-                *data = static_cast<GLint>(end - start);
                 return;
             }
             default:
@@ -755,15 +853,32 @@ namespace MobileGL::MG_Impl::GLImpl {
             return;
         }
 
+        // GL 4.6 core 22.1: an indexed query answers EVERY indexed state, and GL_SCISSOR_TEST is
+        // indexed by viewport just like GL_BLEND is by draw buffer. Without this the integer
+        // width fell through to the backend passthrough and answered GL_INVALID_ENUM, which is
+        // the sticky error KHR-GL43.viewport_array.queries trips over at its next error check.
+        if (MG_Util::ConvertGLEnumToCapabilityInput(target) != CapabilityInput::Unknown) {
+            *data = IsEnabledi(target, index);
+            return;
+        }
+
         switch (target) {
         // ARB_viewport_array queries the indexed rectangles through glGetIntegeri_v as well
-        // (gl4cMultiBindTests and the viewport_array group both do). The frontend keeps one
+        // (gl4cMultiBindTests and the viewport_array group both do).
-        // viewport and one scissor box, so every in-range index reports that one.
         case GL_VIEWPORT:
         case GL_SCISSOR_BOX:
+        case GL_DEPTH_RANGE: {
             if (!ValidateViewportQueryIndex(index, __func__)) return;
-            GetIntegerv(target, data);
+            GLfloat values[4] = {};
+            ReadIndexedViewportStateFloat(target, index, values);
+            const GLsizei components = IndexedViewportQueryComponents(target);
+            for (GLsizei i = 0; i < components; ++i) {
+                // Round, not truncate: glGetIntegerv on floating-point state rounds to nearest
+                // (GL 4.6 core 22.2), so a 255.875-wide viewport reads back as 256 and not 255.
+                data[i] = static_cast<GLint>(std::lround(values[i]));
+            }
             return;
+        }
         // The vertex buffer binding points of the vertex array object that is bound. Indexed by
         // binding point, not by attribute (GL 4.6 core 10.3.1).
         case GL_VERTEX_BINDING_BUFFER:
@@ -890,7 +1005,10 @@ namespace MobileGL::MG_Impl::GLImpl {
         }
         if (IsIndexedViewportQuery(target)) {
             if (!ValidateViewportQueryIndex(index, __func__)) return;
-            GetFloatv(target, data);
+            // Verbatim, NOT via the integer width: the viewport is float state and
+            // KHR-GL43.viewport_array.viewport_api compares the read-back with ==, so a
+            // glViewportIndexedf(i, 0.125f, ...) has to come back as 0.125f exactly.
+            ReadIndexedViewportStateFloat(target, index, data);
             return;
         }
         GLint ints[4] = {};
@@ -907,7 +1025,12 @@ namespace MobileGL::MG_Impl::GLImpl {
         }
         if (IsIndexedViewportQuery(target)) {
             if (!ValidateViewportQueryIndex(index, __func__)) return;
-            GetDoublev(target, data);
+            GLfloat values[4] = {};
+            ReadIndexedViewportStateFloat(target, index, values);
+            const GLsizei components = IndexedViewportQueryComponents(target);
+            for (GLsizei i = 0; i < components; ++i) {
+                data[i] = static_cast<GLdouble>(values[i]);
+            }
             return;
         }
         GLint ints[4] = {};
@@ -951,9 +1074,8 @@ namespace MobileGL::MG_Impl::GLImpl {
                     *data = 0;
                     return;
                 }
-                const auto start = std::min(range.start, bufferObject->GetSize());
+                // Verbatim, unclamped - see the GetIntegeri_v arm.
-                const auto end = std::min(range.end, bufferObject->GetSize());
+                *data = static_cast<GLint64>(range.end - range.start);
-                *data = static_cast<GLint64>(end - start);
                 return;
             }
             default:
@@ -961,15 +1083,35 @@ namespace MobileGL::MG_Impl::GLImpl {
             }
         }
 
-        auto getInteger64i = MG_Backend::gBackendFunctionsTable.GL.GetInteger64i_v;
+        // The one indexed pname whose value genuinely needs 64 bits: a vertex buffer binding
-        if (!getInteger64i) {
+        // offset is an intptr, so taking the 32-bit route below would truncate it.
-            *data = 0;
+        if (target == GL_VERTEX_BINDING_OFFSET) {
+            if (index >= VertexArrayImpl::GetMaxVertexAttribBindings()) {
                 MG_State::pGLContext->RecordError(
-                ErrorCode::InvalidOperation,
+                    ErrorCode::InvalidValue,
-                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", __func__, "Backend does not support indexed integer queries."));
+                    MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", __func__,
+                                                 "Vertex buffer binding index is out of range."));
                 return;
             }
-        getInteger64i(target, index, data);
+            const auto& vao = MG_State::pGLContext->GetBoundVertexArray();
+            *data = vao ? static_cast<GLint64>(vao->GetBindingPoint(index).Offset) : 0;
+            return;
+        }
+
+        // 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);
+        // The viewport-array rectangles are the only multi-component indexed state here; every
+        // other pname is scalar, so widening element 0 alone would silently truncate them.
+        const GLsizei components = IsIndexedViewportQuery(target) ? IndexedViewportQueryComponents(target) : 1;
+        for (GLsizei i = 0; i < components; ++i) {
+            data[i] = static_cast<GLint64>(values[i]);
+        }
     }
 
     void GetInteger64v(GLenum pname, GLint64* params) {
@@ -1379,7 +1521,7 @@ namespace MobileGL::MG_Impl::GLImpl {
             *params = kFrontendMaxCombinedAtomicCounters;
             return;
         case GL_MAX_COMBINED_UNIFORM_BLOCKS:
-            *params = kFrontendMaxCombinedUniformBlocks;
+            *params = ClampUniformBlockCount(kFrontendMaxCombinedUniformBlocks);
             return;
         case GL_MAX_DUAL_SOURCE_DRAW_BUFFERS:
             *params = 1; // TODO
@@ -1394,7 +1536,7 @@ namespace MobileGL::MG_Impl::GLImpl {
             *params = kFrontendMaxFragmentAtomicCounters;
             return;
         case GL_MAX_FRAGMENT_SHADER_STORAGE_BLOCKS:
-            *params = 16; // TODO
+            *params = ClampStorageBlockCount(16); // TODO
             return;
         case GL_MAX_FRAGMENT_INPUT_COMPONENTS:
             *params = kFrontendMaxFragmentInputComponents;
@@ -1411,13 +1553,16 @@ namespace MobileGL::MG_Impl::GLImpl {
             *params = kFrontendMaxFragmentUniformVectors;
             return;
         case GL_MAX_FRAGMENT_UNIFORM_BLOCKS:
-            *params = kFrontendMaxFragmentUniformBlocks;
+            *params = ClampUniformBlockCount(kFrontendMaxFragmentUniformBlocks);
             return;
         case GL_MAX_GEOMETRY_ATOMIC_COUNTERS:
             *params = kFrontendMaxGeometryAtomicCounters;
             return;
+        case GL_MAX_GEOMETRY_ATOMIC_COUNTER_BUFFERS:
+            *params = kFrontendMaxGeometryAtomicCounterBuffers;
+            return;
         case GL_MAX_GEOMETRY_SHADER_STORAGE_BLOCKS:
-            *params = 16; // TODO
+            *params = ClampStorageBlockCount(16); // TODO
             return;
         case GL_MAX_GEOMETRY_INPUT_COMPONENTS:
             *params = kFrontendMaxGeometryInputComponents;
@@ -1440,7 +1585,7 @@ namespace MobileGL::MG_Impl::GLImpl {
             *params = kFrontendMaxGeometryTotalOutputComponents;
             return;
         case GL_MAX_GEOMETRY_UNIFORM_BLOCKS:
-            *params = kFrontendMaxGeometryUniformBlocks;
+            *params = ClampUniformBlockCount(kFrontendMaxGeometryUniformBlocks);
             return;
         case GL_MAX_GEOMETRY_UNIFORM_COMPONENTS:
             *params = kFrontendMaxGeometryUniformComponents;
@@ -1472,9 +1617,15 @@ namespace MobileGL::MG_Impl::GLImpl {
         case GL_MAX_TESS_CONTROL_ATOMIC_COUNTERS:
             *params = kFrontendMaxTessControlAtomicCounters;
             return;
+        case GL_MAX_TESS_CONTROL_ATOMIC_COUNTER_BUFFERS:
+            *params = kFrontendMaxTessControlAtomicCounterBuffers;
+            return;
         case GL_MAX_TESS_EVALUATION_ATOMIC_COUNTERS:
             *params = kFrontendMaxTessEvaluationAtomicCounters;
             return;
+        case GL_MAX_TESS_EVALUATION_ATOMIC_COUNTER_BUFFERS:
+            *params = kFrontendMaxTessEvaluationAtomicCounterBuffers;
+            return;
         case GL_MAX_TESS_CONTROL_IMAGE_UNIFORMS:
             *params = 0;
             return;
@@ -1482,10 +1633,10 @@ namespace MobileGL::MG_Impl::GLImpl {
             *params = 0;
             return;
         case GL_MAX_TESS_CONTROL_SHADER_STORAGE_BLOCKS:
-            *params = 16; // TODO
+            *params = ClampStorageBlockCount(16); // TODO
             return;
         case GL_MAX_TESS_EVALUATION_SHADER_STORAGE_BLOCKS:
-            *params = 16; // TODO
+            *params = ClampStorageBlockCount(16); // TODO
             return;
         case GL_MAX_TEXTURE_LOD_BIAS:
             *params = 15; // TODO
@@ -1502,13 +1653,16 @@ namespace MobileGL::MG_Impl::GLImpl {
         case GL_MAX_VERTEX_ATOMIC_COUNTERS:
             *params = kFrontendMaxVertexAtomicCounters;
             return;
+        case GL_MAX_VERTEX_ATOMIC_COUNTER_BUFFERS:
+            *params = kFrontendMaxVertexAtomicCounterBuffers;
+            return;
         case GL_MAX_VERTEX_IMAGE_UNIFORMS:
             *params = MG_Backend::pActiveBackendObject
                           ? MG_Backend::pActiveBackendObject->GetDynamicParameters().MaxVertexImageUniforms
                           : MG_Backend::DynamicBackendParameters{}.MaxVertexImageUniforms;
             return;
         case GL_MAX_VERTEX_SHADER_STORAGE_BLOCKS:
-            *params = 16; // TODO
+            *params = ClampStorageBlockCount(16); // TODO
             return;
         case GL_MAX_VERTEX_UNIFORM_COMPONENTS:
             *params = kFrontendMaxVertexUniformComponents;
@@ -1520,7 +1674,7 @@ namespace MobileGL::MG_Impl::GLImpl {
             *params = kFrontendMaxVertexOutputComponents;
             return;
         case GL_MAX_VERTEX_UNIFORM_BLOCKS:
-            *params = kFrontendMaxVertexUniformBlocks;
+            *params = ClampUniformBlockCount(kFrontendMaxVertexUniformBlocks);
             return;
         case GL_NUM_COMPRESSED_TEXTURE_FORMATS:
             *params = 0; // compressed texture upload entrypoints are still unimplemented
@@ -1891,7 +2045,7 @@ namespace MobileGL::MG_Impl::GLImpl {
 
         const auto& activeBackendObject = MG_Backend::pActiveBackendObject;
         if (!activeBackendObject) {
-            MGLOG_E("activeBackendObject is not initialized!");
+            MGLOG_E_ONCE("activeBackendObject is not initialized!");
             return;
         }
         const auto& rendererInfo = activeBackendObject->GetRendererInfo();
@@ -1920,13 +2074,13 @@ namespace MobileGL::MG_Impl::GLImpl {
             *params = dynamicParameters.SubgroupQuadOperationsInAllStages ? GL_TRUE : GL_FALSE;
             break;
         case GL_MAX_COMPUTE_SHADER_STORAGE_BLOCKS:
-            *params = dynamicParameters.MaxComputeShaderStorageBlocks;
+            *params = ClampStorageBlockCount(dynamicParameters.MaxComputeShaderStorageBlocks);
             break;
         case GL_MAX_COMBINED_SHADER_STORAGE_BLOCKS:
-            *params = dynamicParameters.MaxCombinedShaderStorageBlocks;
+            *params = ClampStorageBlockCount(dynamicParameters.MaxCombinedShaderStorageBlocks);
             break;
         case GL_MAX_COMPUTE_UNIFORM_BLOCKS:
-            *params = dynamicParameters.MaxComputeUniformBlocks;
+            *params = ClampUniformBlockCount(dynamicParameters.MaxComputeUniformBlocks);
             break;
         case GL_MAX_COMPUTE_TEXTURE_IMAGE_UNITS:
             *params = dynamicParameters.MaxComputeTextureImageUnits;
@@ -2056,7 +2210,15 @@ namespace MobileGL::MG_Impl::GLImpl {
             *params = static_cast<GLint>(GetIndexedBufferQueryPointCount(BufferTarget::AtomicCounter));
             break;
         case GL_MAX_ATOMIC_COUNTER_BUFFER_SIZE:
-            *params = kFrontendMaxAtomicCounterBufferSize;
+            // The conformance suite splits this evenly across every advertised binding point and
+            // binds all of them in one glBindBuffersRange
+            // (KHR-GL44.multi_bind.functional_bind_buffers_range), so the pair has to divide:
+            // 32 bytes over 36 binding points is a zero-sized range, which BindBufferRange
+            // rejects with INVALID_VALUE before it binds anything. Floor the advertised size at
+            // one counter per binding point.
+            *params = std::max<GLint>(
+                kFrontendMaxAtomicCounterBufferSize,
+                static_cast<GLint>(GetIndexedBufferQueryPointCount(BufferTarget::AtomicCounter) * sizeof(GLuint)));
             break;
         case GL_MAX_TEXTURE_BUFFER_SIZE:
             *params = dynamicParameters.MaxTextureBufferSize;
@@ -2121,7 +2283,15 @@ namespace MobileGL::MG_Impl::GLImpl {
             params[1] = dynamicParameters.MaxViewportHeight;
             break;
         case GL_MAX_VIEWPORTS:
-            *params = dynamicParameters.MaxViewports;
+            // The frontend's own state width, not the backend's device limit. GL 4.3 core
+            // requires MAX_VIEWPORTS >= 16 and every indexed viewport entry point validates
+            // against RenderStateParameters::MAX_VIEWPORTS, so reporting anything else would
+            // either advertise viewports the state cannot hold or reject indices it can. A
+            // Vulkan device without the multiViewport feature reports maxViewports == 1, which
+            // limits what can be RASTERIZED to more than one rectangle (see the multiViewport
+            // gate in VulkanRenderer), not what the GL state can hold; caps.MaxViewports keeps
+            // carrying that device number for exactly that decision.
+            *params = static_cast<GLint>(RenderStateParameters::MAX_VIEWPORTS);
             break;
         case GL_MINOR_VERSION:
             *params = rendererInfo.RendererGLInfo.TargetGLVersion.Minor;
@@ -2177,7 +2347,7 @@ namespace MobileGL::MG_Impl::GLImpl {
             *params = static_cast<GLint>(std::lround(dynamicParameters.MaxTextureMaxAnisotropy));
             break;
         default:
-            MGLOG_E("glGetIntegerv: Invalid enum %s (0x%X)", MG_Util::ConvertGLEnumToString(pname).c_str(), pname);
+            MGLOG_D("glGetIntegerv: Invalid enum %s (0x%X)", MG_Util::ConvertGLEnumToString(pname).c_str(), pname);
             MG_State::pGLContext->RecordError(ErrorCode::InvalidEnum,
                                               MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", "GetIntegerv",
                                                                            std::format("Invalid enum: 0x{:X}", pname)));

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());
     }
@@ -815,7 +850,12 @@ namespace MobileGL::MG_Impl::GLImpl {
     // vector per column - while the value glGetUniform* must return is tightly packed
     // columns * rows floats. Only mat4 is the same either way; every other shape needs the
     // padding undone, and the readback has to undo exactly what UniformMatrixfv_Object put
-    // there. Returns false when `ttype` is not a float matrix (nothing to unpack).
+    // there. Returns false when there is nothing here to unpack.
+    //
+    // A DOUBLE matrix is declined not because it is laid out differently - it is not, the
+    // demotion makes a dmat4 a mat4 in the shader and a mat4-shaped slot here - but because it
+    // is ROUTED differently: the caller's component-by-component EbtDouble branch has to widen
+    // each float back to the queried type, and it undoes the same padding itself.
     Bool TryGatherFloatMatrixColumns(const glslang::TType* ttype, const char* pBase, void* params) {
         if (ttype == nullptr || !ttype->isMatrix() || ttype->getBasicType() == glslang::EbtDouble) return false;
         const Int columns = ttype->getMatrixCols();
@@ -828,12 +868,11 @@ 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 MG_State::GLState::ProgramObject::UniformStorageSpanInBytes(ttype, tightSize);
-            return static_cast<SizeT>(ttype->getMatrixCols()) * 4 * sizeof(GLfloat);
-        }
-        return tightSize;
     }
 
     void GetUniform_State(GLuint program, GLint location, void* params) {
@@ -869,13 +908,19 @@ namespace MobileGL::MG_Impl::GLImpl {
             const SizeT span = UniformStorageSpanInBytes(ttype, size);
             if (pUBO == nullptr || offset == MG_State::GLState::ProgramObject::kInvalidUniformOffset ||
                 offset + span > programObject->GetUBOSize()) {
-                MGLOG_E("%s: uniform at program %u location %d has no backing storage; returning nothing", __func__,
+                MGLOG_E_ONCE("%s: uniform at program %u location %d has no backing storage; returning nothing", __func__,
                         program, location);
                 return;
             }
 
             if (!TryGatherFloatMatrixColumns(ttype, pUBO + offset, params)) {
-                Memcpy(params, pUBO + offset, size);
+                // Never more than the uniform actually occupies. `size` is the GL type size,
+                // which for a `double` uniform is twice its storage - every 64-bit float is
+                // narrowed before the module reaches a backend, so the slot holds floats. The
+                // typed entry points (glGetUniformdv and friends) go through
+                // GetUniformScalar_State, which converts component by component; this raw
+                // copy has no type to convert with, so it is bounded rather than converted.
+                Memcpy(params, pUBO + offset, std::min<SizeT>(size, span));
             }
         }
         // TODO: handle 1i variant as texture unit
@@ -917,7 +962,7 @@ namespace MobileGL::MG_Impl::GLImpl {
         const SizeT span = UniformStorageSpanInBytes(ttype, size);
         if (pUBO == nullptr || offset == MG_State::GLState::ProgramObject::kInvalidUniformOffset ||
             offset + span > programObject->GetUBOSize()) {
-            MGLOG_E("%s: uniform at program %u location %d has no backing storage; returning nothing", __func__,
+            MGLOG_E_ONCE("%s: uniform at program %u location %d has no backing storage; returning nothing", __func__,
                     program, location);
             return;
         }
@@ -926,22 +971,27 @@ namespace MobileGL::MG_Impl::GLImpl {
             if (TryGatherFloatMatrixColumns(ttype, pUBO + offset, params)) return;
         }
 
-        // A double-precision uniform is the one case where the stored component type can
+        // A double-precision uniform is the one case where the stored component type differs
-        // differ from the queried one for a non-opaque uniform, and the difference is not
+        // from the DECLARED one for a non-opaque uniform: the shader's 64-bit floats are
-        // just a reinterpretation: it is twice as wide, so a raw copy would overrun the
+        // narrowed to 32 bits before the module reaches a backend
-        // caller's buffer as well as return nonsense. Read component by component and let
+        // (ShaderTranspiler::DemoteFloat64Pass), so what is in the global UBO is a float per
-        // GL's conversion rules (7.6: round to nearest for the integer queries) apply.
+        // component, laid out exactly like the float-typed twin of this uniform - std140
+        // 16-byte column stride for a matrix included. Reading it as a GLdouble would return
+        // two components reinterpreted as one. Read component by component and let GL's
+        // conversion rules (7.6: round to nearest for the integer queries) apply; the value
+        // widens back to the queried type, having lost precision at the glUniform*d that
+        // stored it and not here.
         if (ttype->getBasicType() == glslang::EbtDouble) {
             const Int columns = ttype->isMatrix() ? ttype->getMatrixCols() : 1;
             const Int rows = ttype->isMatrix() ? ttype->getMatrixRows()
                                                : (ttype->isVector() ? ttype->getVectorSize() : 1);
-            // The slot the linker handed out is exactly `columns` columns wide, so it also
+            // std140 gives every matrix column its own 16-byte slot; a non-matrix is one
-            // states the column stride - which for a double matrix is not a float's 16 bytes.
+            // tightly packed run and never reaches the stride at all.
-            const SizeT columnStride = columns > 0 ? size / static_cast<SizeT>(columns) : size;
+            const SizeT columnStride = 4 * sizeof(GLfloat);
             for (Int column = 0; column < columns; ++column) {
                 for (Int row = 0; row < rows; ++row) {
-                    GLdouble component = 0.0;
+                    GLfloat component = 0.0f;
-                    Memcpy(&component, pUBO + offset + column * columnStride + row * sizeof(GLdouble),
+                    Memcpy(&component, pUBO + offset + column * columnStride + row * sizeof(GLfloat),
                            sizeof(component));
                     if constexpr (std::is_integral_v<T>) {
                         // Rounded to the nearest integer and clamped into the queried type's
@@ -1007,21 +1057,20 @@ namespace MobileGL::MG_Impl::GLImpl {
             return;
         }
 
-        static Bool allowVSOnlyPrograms;
+        // Read fresh every link, never latched in a static: the capability is
-        static Bool initialized = false;
+        // per-backend, and a latch would freeze it across a backend teardown +
-        if (!initialized) {
+        // re-initialization (the previous function-static memo here never even set
+        // its own initialized flag, so it re-read every call anyway - this makes
+        // the always-fresh behavior the stated one). A struct-field read per
+        // glLinkProgram costs nothing.
         const auto& activeBackendObject = MG_Backend::pActiveBackendObject;
         if (!activeBackendObject) {
-                MGLOG_E("activeBackendObject is not initialized!");
+            MGLOG_E_ONCE("activeBackendObject is not initialized!");
             return;
         }
-            const auto& rendererInfo = activeBackendObject->GetRendererInfo();
+        const Bool allowVSOnlyPrograms =
-            allowVSOnlyPrograms = (Int)rendererInfo.StaticBackendCapability.AllowVSOnlyPrograms;
+            activeBackendObject->GetRendererInfo().StaticBackendCapability.AllowVSOnlyPrograms;
-        }
-        const auto& activeBackendObject = MG_Backend::pActiveBackendObject;
-        if (activeBackendObject) {
         programObject->SetMaxFragmentOutputColorNumber(activeBackendObject->GetDynamicParameters().MaxDrawBuffers);
-        }
         programObject->Link(!allowVSOnlyPrograms);
     }
 
@@ -1085,23 +1134,45 @@ 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.
-                MGLOG_E("%s: uniform size mismatch at program %u location %u: expected at least %zu bytes, got %zu "
+                MGLOG_E_ONCE("%s: uniform size mismatch at program %u location %u: expected at least %zu bytes, got %zu "
                         "bytes; clamping",
                         __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
                 // storage. Log and drop the write instead of faulting.
-                MGLOG_E("%s: uniform at program %u location %u has no backing storage (ubo=%p offset=%u size=%zu "
+                MGLOG_E_ONCE("%s: uniform at program %u location %u has no backing storage (ubo=%p offset=%u size=%zu "
                         "uboSize=%zu); dropping write",
                         __func__, programObject.GetExternalIndex(), location, static_cast<void*>(pUBO), offset,
                         writeSize, uboSize);
@@ -1192,36 +1263,39 @@ namespace MobileGL::MG_Impl::GLImpl {
         }
     }
 
-    // glUniform*d / glUniformMatrix*dv. The vector forms need nothing beyond the shared
+    // glUniform*d / glUniformMatrix*dv. Neither needs a layout of its own any more: the
-    // upload template - it is already typed on the component - but a matrix does: the
+    // transpile chain narrows every 64-bit float in the shader to 32 bits
-    // column stride the linker used for a double matrix is not the 16 bytes a float one
+    // (ShaderTranspiler::DemoteFloat64Pass) and the global UBO is laid out by reflecting that
-    // gets. It is not guessed here; the slot the uniform was given is exactly `columns`
+    // demoted module, so a double uniform's storage IS a float uniform's - same offset, same
-    // columns wide, so dividing states the stride the rest of the pipeline agreed on.
+    // 4-byte components, same std140 column padding for matrices. Narrowing here, at the one
-    template <typename Program>
+    // place the 64-bit value enters, and then handing the bytes to the ordinary float upload
-    void UniformMatrixdv_Object(Program& programObject, GLint location, GLsizei count, GLboolean transpose,
+    // path is what keeps the two in step; a separate double-shaped layout here would write
-                                const GLdouble* value, Int columns, Int rows) {
+    // 8-byte components into 4-byte slots and silently address the wrong ones.
-        const SizeT slotSize = programObject.GetUniformSizesInBytes(location);
+    //
-        const SizeT columnStride = columns > 0 ? slotSize / static_cast<SizeT>(columns) : slotSize;
+    // The narrowing is the same static_cast the shader's own arithmetic now performs, so the
-        const SizeT componentCount = static_cast<SizeT>(columns) * static_cast<SizeT>(rows);
+    // value the shader reads is the value glUniform*d was given, at float precision.
-        Vector<GLdouble> column(static_cast<SizeT>(rows));
+    template <GLsizei ItemCount>
-        for (GLint matrix = 0; matrix < count; ++matrix) {
+    void UniformvNarrowed_State(GLint location, GLsizei count, const GLdouble* value) {
-            if (matrix > 0 && !programObject.UniformLocationsAliasSameUniform(location, location + matrix)) break;
+        if (value == nullptr || count <= 0) {
-            if (!programObject.IsValidUniformLocation(location + matrix)) {
+            // Same shape as the float entry points: the location validation still runs, and a
-                RecordInvalidUniformLocationError(__func__, location + matrix, "the current program object");
+            // null pointer is left to fault exactly where glUniform*fv would.
+            Uniformv_State<ItemCount>(location, count, reinterpret_cast<const GLfloat*>(value));
             return;
         }
-            const GLdouble* source = value + matrix * componentCount;
+        Vector<GLfloat> narrowed(static_cast<SizeT>(count) * ItemCount);
-            for (Int c = 0; c < columns; ++c) {
+        for (SizeT i = 0; i < narrowed.size(); ++i) narrowed[i] = static_cast<GLfloat>(value[i]);
-                for (Int r = 0; r < rows; ++r) {
+        Uniformv_State<ItemCount>(location, count, narrowed.data());
-                    column[r] = transpose == GL_TRUE ? source[r * columns + c] : source[c * rows + r];
-                }
-                Uniform_State<1>(programObject, location + matrix, column.data(), c * columnStride);
-                for (Int r = 1; r < rows; ++r) {
-                    Uniform_State<1>(programObject, location + matrix, column.data() + r,
-                                     c * columnStride + r * sizeof(GLdouble));
-                }
     }
+
+    template <GLsizei ItemCount>
+    void ProgramUniformvNarrowed_State(GLuint program, GLint location, GLsizei count, const GLdouble* value) {
+        if (value == nullptr || count <= 0) {
+            ProgramUniformv_State<ItemCount>(program, location, count, reinterpret_cast<const GLfloat*>(value));
+            return;
         }
+        Vector<GLfloat> narrowed(static_cast<SizeT>(count) * ItemCount);
+        for (SizeT i = 0; i < narrowed.size(); ++i) narrowed[i] = static_cast<GLfloat>(value[i]);
+        ProgramUniformv_State<ItemCount>(program, location, count, narrowed.data());
     }
 
     // glUniformMatrix*fv / glProgramUniformMatrix*fv, every shape (square and non-square).
@@ -1270,6 +1344,22 @@ namespace MobileGL::MG_Impl::GLImpl {
         }
     }
 
+    // glUniformMatrix*dv / glProgramUniformMatrix*dv. Narrowed to the float form and handed
+    // straight to it: after DemoteFloat64Pass a `dmat4` uniform is a `mat4` in the shader and a
+    // mat4-shaped slot in the global UBO, columns padded to a vec4 and all. Everything else
+    // about the call - transpose handling, the array-element walk, the opaque-uniform refusal -
+    // is then the one implementation both spellings share.
+    template <typename Program>
+    void UniformMatrixdv_Object(Program& programObject, GLint location, GLsizei count, GLboolean transpose,
+                                const GLdouble* value, Int columns, Int rows) {
+        if (value == nullptr || count <= 0) return;
+        const SizeT componentCount = static_cast<SizeT>(columns) * static_cast<SizeT>(rows);
+        Vector<GLfloat> narrowed(static_cast<SizeT>(count) * componentCount);
+        for (SizeT i = 0; i < narrowed.size(); ++i) narrowed[i] = static_cast<GLfloat>(value[i]);
+        UniformMatrixfv_Object(programObject, "glUniformMatrixdv", location, count, transpose, narrowed.data(),
+                               columns, rows, "the current program object");
+    }
+
     // Helper function to transpose a 2x2 matrix
     void TransposeMatrix2x2(const GLfloat* input, GLfloat* output) {
         // Input matrix is in column-major order (OpenGL default)
@@ -1716,7 +1806,7 @@ namespace MobileGL::MG_Impl::GLImpl {
             break;
         }
         default:
-            MGLOG_E("%s: unknown pname = %p %s", __func__, pname, MG_Util::ConvertGLEnumToString(pname).c_str());
+            MGLOG_D("%s: unknown pname = %p %s", __func__, pname, MG_Util::ConvertGLEnumToString(pname).c_str());
             MG_State::pGLContext->RecordError(
                 ErrorCode::InvalidEnum,
                 MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", __func__,
@@ -2033,71 +2123,71 @@ namespace MobileGL::MG_Impl::GLImpl {
     }
     void Uniform1d(GLint location, GLdouble v0) {
         const GLdouble v[] = {v0};
-        Uniformv_State<1>(location, 1, v);
+        UniformvNarrowed_State<1>(location, 1, v);
     }
 
     void Uniform1dv(GLint location, GLsizei count, const GLdouble* value) {
-        Uniformv_State<1>(location, count, value);
+        UniformvNarrowed_State<1>(location, count, value);
     }
 
     void ProgramUniform1d(GLuint program, GLint location, GLdouble v0) {
         const GLdouble v[] = {v0};
-        ProgramUniformv_State<1>(program, location, 1, v);
+        ProgramUniformvNarrowed_State<1>(program, location, 1, v);
     }
 
     void ProgramUniform1dv(GLuint program, GLint location, GLsizei count, const GLdouble* value) {
-        ProgramUniformv_State<1>(program, location, count, value);
+        ProgramUniformvNarrowed_State<1>(program, location, count, value);
     }
     void Uniform2d(GLint location, GLdouble v0, GLdouble v1) {
         const GLdouble v[] = {v0, v1};
-        Uniformv_State<2>(location, 1, v);
+        UniformvNarrowed_State<2>(location, 1, v);
     }
 
     void Uniform2dv(GLint location, GLsizei count, const GLdouble* value) {
-        Uniformv_State<2>(location, count, value);
+        UniformvNarrowed_State<2>(location, count, value);
     }
 
     void ProgramUniform2d(GLuint program, GLint location, GLdouble v0, GLdouble v1) {
         const GLdouble v[] = {v0, v1};
-        ProgramUniformv_State<2>(program, location, 1, v);
+        ProgramUniformvNarrowed_State<2>(program, location, 1, v);
     }
 
     void ProgramUniform2dv(GLuint program, GLint location, GLsizei count, const GLdouble* value) {
-        ProgramUniformv_State<2>(program, location, count, value);
+        ProgramUniformvNarrowed_State<2>(program, location, count, value);
     }
     void Uniform3d(GLint location, GLdouble v0, GLdouble v1, GLdouble v2) {
         const GLdouble v[] = {v0, v1, v2};
-        Uniformv_State<3>(location, 1, v);
+        UniformvNarrowed_State<3>(location, 1, v);
     }
 
     void Uniform3dv(GLint location, GLsizei count, const GLdouble* value) {
-        Uniformv_State<3>(location, count, value);
+        UniformvNarrowed_State<3>(location, count, value);
     }
 
     void ProgramUniform3d(GLuint program, GLint location, GLdouble v0, GLdouble v1, GLdouble v2) {
         const GLdouble v[] = {v0, v1, v2};
-        ProgramUniformv_State<3>(program, location, 1, v);
+        ProgramUniformvNarrowed_State<3>(program, location, 1, v);
     }
 
     void ProgramUniform3dv(GLuint program, GLint location, GLsizei count, const GLdouble* value) {
-        ProgramUniformv_State<3>(program, location, count, value);
+        ProgramUniformvNarrowed_State<3>(program, location, count, value);
     }
     void Uniform4d(GLint location, GLdouble v0, GLdouble v1, GLdouble v2, GLdouble v3) {
         const GLdouble v[] = {v0, v1, v2, v3};
-        Uniformv_State<4>(location, 1, v);
+        UniformvNarrowed_State<4>(location, 1, v);
     }
 
     void Uniform4dv(GLint location, GLsizei count, const GLdouble* value) {
-        Uniformv_State<4>(location, count, value);
+        UniformvNarrowed_State<4>(location, count, value);
     }
 
     void ProgramUniform4d(GLuint program, GLint location, GLdouble v0, GLdouble v1, GLdouble v2, GLdouble v3) {
         const GLdouble v[] = {v0, v1, v2, v3};
-        ProgramUniformv_State<4>(program, location, 1, v);
+        ProgramUniformvNarrowed_State<4>(program, location, 1, v);
     }
 
     void ProgramUniform4dv(GLuint program, GLint location, GLsizei count, const GLdouble* value) {
-        ProgramUniformv_State<4>(program, location, count, value);
+        ProgramUniformvNarrowed_State<4>(program, location, count, value);
     }
     void UniformMatrix2dv(GLint location, GLsizei count, GLboolean transpose, const GLdouble* value) {
         if (location == -1) return;
@@ -2658,6 +2748,15 @@ 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
+        // GL 4.6 core 7.4 asks only that the name came from GenProgramPipelines and has not been
-        // reserved-but-unmaterialised name is INVALID_OPERATION (GL 4.6 core 7.4).
+        // 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/Program/ProgramInterface.cpp

@@ -210,11 +210,66 @@ namespace MobileGL::MG_Impl::GLImpl::ProgramInterface {
 
         // ---- model construction --------------------------------------------------------
 
+        // GL_REFERENCED_BY_*_SHADER for an ARRAYED block instance, refined per element.
+        //
+        // glslang records a block reference by walking up to the base symbol and calling
+        // addBlockName with the whole ARRAY type, which ORs the referencing stage into every
+        // element at once - it has not resolved the subscript yet at that point. So reading
+        // "e[0].b" marks both TrickyBlock[0] and TrickyBlock[1] as referenced by the fragment
+        // stage (KHR-GL43.program_interface_query.uniform-block-types).
+        //
+        // The MEMBER masks are exact: EShReflectionAllBlockVariables enumerates every member of
+        // every element with the stage mask suppressed, and only the dereference chain actually
+        // walked turns a bit on - and that chain carries the subscript. So the union of a block
+        // instance's members is the reference set of that instance.
+        //
+        // Applied ONLY to arrayed instances, because for a scalar block glslang is already exact.
+        // Note the union is used even when it is empty: an array element nobody dereferenced has
+        // no member bits and is genuinely referenced by nobody, which is the whole point - falling
+        // back to the block's own mask there would restore the over-approximation.
+        Vector<Uint32> BuildBlockStagesFromMembers(const glslang::TProgram& reflection, Int blockCount) {
+            auto& mutableReflection = const_cast<glslang::TProgram&>(reflection);
+            Vector<Uint32> stagesByBlock(static_cast<SizeT>(blockCount < 0 ? 0 : blockCount), 0u);
+            const Int uniformCount = mutableReflection.getNumUniformVariables();
+            for (Int index = 0; index < uniformCount; ++index) {
+                const auto& uniform = mutableReflection.getUniform(index);
+                const Int owner = uniform.index;
+                if (owner < 0 || owner >= blockCount) continue;
+                stagesByBlock[static_cast<SizeT>(owner)] |= static_cast<Uint32>(uniform.stages);
+            }
+            return stagesByBlock;
+        }
+
+        // UNIFORM blocks only, and that scope is load-bearing rather than cautious. The member
+        // names glslang produces for a uniform block array carry the subscript
+        // ("TrickyBlock[0].b", via EShReflectionStrictArraySuffix), so each element's members are
+        // distinct entries and the bits land on the right one. A SHADER STORAGE block array does
+        // NOT get that treatment - its buffer variables reflect under one subscript-free spelling
+        // shared by every element - so a union over them credits element 0 and starves the rest.
+        // KHR-GL43.program_interface_query.ssb-types is the case that says so: it reads ss[0] and
+        // ss[1] and requires both to report the fragment stage, which only glslang's own
+        // (deliberately over-approximating) block mask gets right. Storage and atomic-counter
+        // blocks therefore keep that mask untouched.
+        Uint32 UniformBlockStages(const glslang::TObjectReflection& block, const Vector<Uint32>& stagesFromMembers,
+                                  Int tIndex) {
+            String arrayBase;
+            Uint element = 0;
+            Bool malformed = false;
+            if (!SplitTrailingSubscript(block.name, arrayBase, element, malformed) || malformed) {
+                return static_cast<Uint32>(block.stages);
+            }
+            if (tIndex < 0 || tIndex >= static_cast<Int>(stagesFromMembers.size())) {
+                return static_cast<Uint32>(block.stages);
+            }
+            return stagesFromMembers[static_cast<SizeT>(tIndex)];
+        }
+
         void BuildBlocks(ProgramObject& program, const glslang::TProgram& reflection, Model& model,
                          Vector<BlockKind>& blockKind, Vector<Int>& blockInterfaceIndex) {
             const Int blockCount = const_cast<glslang::TProgram&>(reflection).getNumUniformBlocks();
             blockKind.assign(blockCount, BlockKind::Uniform);
             blockInterfaceIndex.assign(blockCount, -1);
+            const Vector<Uint32> stagesFromMembers = BuildBlockStagesFromMembers(reflection, blockCount);
 
             for (Int tIndex = 0; tIndex < blockCount; ++tIndex) {
                 const auto& block = const_cast<glslang::TProgram&>(reflection).getUniformBlock(tIndex);
@@ -260,8 +315,8 @@ namespace MobileGL::MG_Impl::GLImpl::ProgramInterface {
                 resource.bufferDataSize = static_cast<GLint>(program.GetUBOSizeAt(glIndex));
                 const Int tIndex = program.TProgramBlockIndex(static_cast<Uint>(glIndex));
                 if (tIndex >= 0 && tIndex < blockCount) {
-                    resource.stages =
+                    resource.stages = UniformBlockStages(const_cast<glslang::TProgram&>(reflection).getUniformBlock(tIndex),
-                        static_cast<Uint32>(const_cast<glslang::TProgram&>(reflection).getUniformBlock(tIndex).stages);
+                                                      stagesFromMembers, tIndex);
                 }
                 model.uniformBlocks.push_back(Move(resource));
             }
@@ -352,6 +407,17 @@ namespace MobileGL::MG_Impl::GLImpl::ProgramInterface {
             }
         }
 
+        // A built-in interface block that a shader redeclares with fewer members keeps the
+        // omitted ones in its type when the redeclaration is ANONYMOUS - glslang hides them
+        // (basic type void) instead of erasing them, because the original shared declaration
+        // has to stay usable. Only the instance-named form erases. So a separable vertex
+        // program that redeclares `out gl_PerVertex { vec4 gl_Position; }` still carries
+        // gl_PointSize and gl_ClipDistance through the block-unwrapping reflection, and they
+        // are not part of its output interface.
+        Bool IsHiddenBlockMember(const glslang::TType* type) {
+            return type != nullptr && type->getBasicType() == glslang::EbtVoid;
+        }
+
         void BuildStageIO(ProgramObject& program, const glslang::TProgram& reflection, Model& model) {
             auto& mutableReflection = const_cast<glslang::TProgram&>(reflection);
 
@@ -359,6 +425,7 @@ namespace MobileGL::MG_Impl::GLImpl::ProgramInterface {
             for (Int index = 0; index < inputCount; ++index) {
                 const auto& refl = mutableReflection.getPipeInput(index);
                 const glslang::TType* type = refl.getType();
+                if (IsHiddenBlockMember(type)) continue;
                 Resource resource;
                 // The Vulkan-semantics parse reflects the vertex builtins under their SPIR-V
                 // names; GL enumerates the GL spellings.
@@ -373,18 +440,28 @@ namespace MobileGL::MG_Impl::GLImpl::ProgramInterface {
                 model.programInputs.push_back(Move(resource));
             }
 
+            // A color number, and therefore a color INDEX, exists only for a fragment stage's
+            // outputs. The output interface belongs to the program's last stage, so for a
+            // separable tessellation/geometry/vertex program these are varyings: asking the
+            // frag-data maps about them can still answer a location (a tess-control output
+            // carries its own layout(location=N)), and a location then manufactures a color
+            // index of 0 where GL requires -1
+            // (KHR-GL43.program_interface_query.separate-programs-tess-control).
+            const Bool lastStageIsFragment = mutableReflection.getIntermediate(EShLangFragment) != nullptr;
             const Int outputCount = mutableReflection.getNumPipeOutputs();
             for (Int index = 0; index < outputCount; ++index) {
                 const auto& refl = mutableReflection.getPipeOutput(index);
                 const glslang::TType* type = refl.getType();
+                if (IsHiddenBlockMember(type)) continue;
                 Resource resource;
                 resource.name = WithArraySuffix(refl.name, type);
                 resource.type = static_cast<GLenum>(refl.glDefineType);
                 resource.arraySize = ArraySizeOf(type, refl.size);
                 resource.location = MappedLocation(program.GetFragmentDataLocation(refl.name.c_str()));
-                if (resource.location < 0) {
+                if (resource.location < 0 || !lastStageIsFragment) {
-                    // A built-in output (gl_FragDepth, gl_SampleMask) and a non-fragment stage
+                    // A built-in output (gl_FragDepth, gl_SampleMask) has no location, and a
-                    // output both have no location, and therefore no color index either.
+                    // non-fragment stage's outputs have no color number at all - either way there
+                    // is no color index.
                     resource.locationIndex = -1;
                 } else {
                     resource.locationIndex = program.GetFragmentDataIndex(refl.name.c_str());

MobileGL/MG_Impl/GLImpl/Query/GL_Query.cpp

@@ -648,4 +648,39 @@ namespace MobileGL::MG_Impl::GLImpl {
         if (!ValidateQueryStreamIndex(__FUNCTION__, target, index)) return;
         GetQueryiv(target, pname, params);
     }
+
+    void DestroyAllQueryObjects() {
+        // Detach the registry under the lock, release outside it - same discipline
+        // (and the same accepted teardown race) as DestroyAllSyncObjects. Without
+        // this drain, every query the app left undeleted survived full library
+        // teardown in the process-global registry: the objects and their backend
+        // wrappers leaked across Destroy/Initialize cycles, stale ids kept
+        // answering IsQuery == GL_TRUE in the re-initialized library, and a later
+        // glDeleteQueries could hand the OLD backend's handle to a DIFFERENT
+        // backend's DeleteBackendQuery, which casts it to the wrong wrapper type.
+        UnorderedMap<GLuint, QueryObject*> orphans;
+        {
+            const std::lock_guard<std::mutex> lock(g_queryObjectsMutex);
+            orphans.swap(g_liveQueryObjects);
+            g_activeTimeElapsedQueryId = 0;
+            g_activePrimitivesWrittenQueryId = 0;
+            g_activePrimitivesGeneratedQueryId = 0;
+            g_activeSamplesPassedQueryId = 0;
+        }
+        if (orphans.empty()) {
+            return;
+        }
+        // Backend handles must be released by the backend that created them, so
+        // this runs while the function table is still populated. Both backends'
+        // DeleteBackendQuery are generation-guarded, so a handle whose renderer
+        // or ES context is already gone frees only the wrapper.
+        const auto deleteBackendQuery = MG_Backend::gBackendFunctionsTable.GL.DeleteBackendQuery;
+        for (const auto& [_, queryObject] : orphans) {
+            if (deleteBackendQuery && queryObject->backendHandle) {
+                deleteBackendQuery(queryObject->backendHandle);
+            }
+            delete queryObject;
+        }
+        MGLOG_D("DestroyAllQueryObjects: reclaimed %zu query object(s) the app left undeleted", orphans.size());
+    }
 } // namespace MobileGL::MG_Impl::GLImpl

MobileGL/MG_Impl/GLImpl/Query/GL_Query.h

@@ -29,4 +29,13 @@ namespace MobileGL::MG_Impl::GLImpl {
     void GetQueryBufferObjecti64v(GLuint id, GLuint buffer, GLenum pname, GLintptr offset);
     void GetQueryBufferObjectui64v(GLuint id, GLuint buffer, GLenum pname, GLintptr offset);
     void QueryCounter(GLuint id, GLenum target);
+    // Destroys every still-registered query object exactly as DeleteQueries would.
+    // GL requires queries to die with their context; called only from full library
+    // teardown (DestroyImpl), where no context survives on any thread, so the
+    // process-global registry can be drained wholesale. Must run while the backend
+    // function table is still populated: each backend handle has to be released by
+    // the backend that created it, never by a later re-initialized one (whose
+    // DeleteBackendQuery would cast the wrapper to the wrong backend's type).
+    // Same contract as DestroyAllSyncObjects.
+    void DestroyAllQueryObjects();
 } // namespace MobileGL::MG_Impl::GLImpl

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>
@@ -19,28 +20,118 @@ namespace MobileGL::MG_Impl::GLImpl {
         return std::clamp(static_cast<Float>(value), 0.0f, 1.0f);
     }
 
-    static Bool ValidateIndexedBlendCapability(GLenum target, GLuint index, const char* functionName) {
+    // GL 4.6 core 17.3.2 and 22.1 give exactly two indexed capabilities: GL_BLEND, indexed by
-        if (target != GL_BLEND) {
+    // draw buffer, and GL_SCISSOR_TEST, indexed by viewport. They have DIFFERENT bounds
+    // (MAX_DRAW_BUFFERS vs MAX_VIEWPORTS), so the limit is picked per target rather than shared.
+    static Bool ValidateIndexedCapability(GLenum target, GLuint index, const char* functionName) {
+        GLuint limit = 0;
+        const char* indexName = nullptr;
+        switch (target) {
+        case GL_BLEND:
+            limit = MG_State::GLState::FramebufferObject::MAX_DRAW_BUFFERS;
+            indexName = "Buffer";
+            break;
+        case GL_SCISSOR_TEST:
+            limit = RenderStateParameters::MAX_VIEWPORTS;
+            indexName = "Viewport";
+            break;
+        default:
             MG_State::pGLContext->RecordError(
                 ErrorCode::InvalidEnum,
                 MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", functionName,
-                                             "Only GL_BLEND is supported for indexed capability state."));
+                                             "Only GL_BLEND and GL_SCISSOR_TEST are supported for indexed "
+                                             "capability state."));
             return false;
         }
 
-        if (index >= MG_State::GLState::FramebufferObject::MAX_DRAW_BUFFERS) {
+        if (index >= limit) {
             MG_State::pGLContext->RecordError(
                 ErrorCode::InvalidValue,
-                MakeUnique<GenericErrorInfo>(
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", functionName,
-                    "MG_Impl/GLImpl", functionName,
+                                             String(indexName) + " index " + std::to_string(index) +
-                    "Buffer index " + std::to_string(index) + " is out of range. Max supported is " +
+                                                 " is out of range. Max supported is " + std::to_string(limit - 1) +
-                        std::to_string(MG_State::GLState::FramebufferObject::MAX_DRAW_BUFFERS - 1) + "."));
+                                                 "."));
             return false;
         }
 
         return true;
     }
 
+    // ------------------ ARB_viewport_array parameter validation ------------------
+    // All three families share the same two shapes, so they share the two checkers. GL 4.6 core
+    // 13.6.1/17.3.2: an out-of-range index is GL_INVALID_VALUE, and so is a negative width or
+    // height. `first + count == MAX_VIEWPORTS` is LEGAL - only strictly greater is an error,
+    // which KHR-GL43.viewport_array.api_errors checks explicitly in both directions.
+    static Bool ValidateViewportIndex(GLuint index, const char* functionName) {
+        if (index < RenderStateParameters::MAX_VIEWPORTS) return true;
+
+        MG_State::pGLContext->RecordError(
+            ErrorCode::InvalidValue,
+            MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", functionName,
+                                         "Viewport index " + std::to_string(index) +
+                                             " is out of range. Max supported is " +
+                                             std::to_string(RenderStateParameters::MAX_VIEWPORTS - 1) + "."));
+        return false;
+    }
+
+    static Bool ValidateViewportRange(GLuint first, GLsizei count, const char* functionName) {
+        if (count < 0) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidValue,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", functionName, "count must not be negative."));
+            return false;
+        }
+        // Widened before adding: first is a GLuint and count a GLsizei, so `first + count` in
+        // 32 bits can wrap past MAX_VIEWPORTS and let an out-of-range range through.
+        const Uint64 last = static_cast<Uint64>(first) + static_cast<Uint64>(count);
+        if (last > RenderStateParameters::MAX_VIEWPORTS) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidValue,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", functionName,
+                                             "first (" + std::to_string(first) + ") + count (" +
+                                                 std::to_string(count) + ") exceeds GL_MAX_VIEWPORTS (" +
+                                                 std::to_string(RenderStateParameters::MAX_VIEWPORTS) + ")."));
+            return false;
+        }
+        return true;
+    }
+
+    template <typename T>
+    static Bool ValidateNonNegativeExtent(T width, T height, const char* functionName) {
+        if (width >= T(0) && height >= T(0)) return true;
+
+        MG_State::pGLContext->RecordError(
+            ErrorCode::InvalidValue,
+            MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", functionName, "Width and height must be non-negative."));
+        return false;
+    }
+
+    // The array forms are all-or-nothing: one bad element rejects the whole call with a SINGLE
+    // GL_INVALID_VALUE and leaves every rectangle untouched. api_errors relies on both halves -
+    // it passes a full 16-element array with exactly one negative extent and then asserts the
+    // error queue holds exactly one entry.
+    template <typename T>
+    static Bool ValidateArrayExtents(GLsizei count, const T* v, const char* functionName) {
+        for (GLsizei i = 0; i < count; ++i) {
+            if (v[i * 4 + 2] >= T(0) && v[i * 4 + 3] >= T(0)) continue;
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidValue,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", functionName,
+                                             "Width and height must be non-negative (element " + std::to_string(i) +
+                                                 ")."));
+            return false;
+        }
+        return true;
+    }
+
+    static Bool ValidateNonNullArray(const void* v, const char* functionName) {
+        if (v != nullptr) return true;
+        MG_State::pGLContext->RecordError(
+            ErrorCode::InvalidValue,
+            MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", functionName, "value pointer cannot be null."));
+        return false;
+    }
+
     static Bool TryConvertBlendEquation(GLenum mode, const char* functionName,
                                         ::MobileGL::BlendEquation& outEquation) {
         outEquation = MG_Util::ConvertGLEnumToBlendEquation(mode);
@@ -92,16 +183,70 @@ namespace MobileGL::MG_Impl::GLImpl {
     }
 
     void Viewport_State(GLint x, GLint y, GLsizei width, GLsizei height) {
-        if (width < 0 || height < 0) {
+        if (!ValidateNonNegativeExtent(width, height, "Viewport_State")) return;
-            MG_State::pGLContext->RecordError(ErrorCode::InvalidValue,
-                                              MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", "Viewport_State",
-                                                                           "Width abd height must be non-negative."));
-            return;
-        }
 
         MG_State::pGLContext->SetViewport(IntVec4(x, y, width, height));
     }
 
+    // ------------------ ARB_viewport_array setters ------------------
+    void ViewportArrayv_State(GLuint first, GLsizei count, const GLfloat* v) {
+        if (!ValidateViewportRange(first, count, "ViewportArrayv_State")) return;
+        if (count == 0) return;
+        if (!ValidateNonNullArray(v, "ViewportArrayv_State")) return;
+        if (!ValidateArrayExtents(count, v, "ViewportArrayv_State")) return;
+
+        for (GLsizei i = 0; i < count; ++i) {
+            MG_State::pGLContext->SetViewportIndexed(first + static_cast<GLuint>(i),
+                                                     FloatVec4(v[i * 4 + 0], v[i * 4 + 1], v[i * 4 + 2], v[i * 4 + 3]));
+        }
+    }
+
+    void ViewportIndexedf_State(GLuint index, GLfloat x, GLfloat y, GLfloat w, GLfloat h) {
+        if (!ValidateViewportIndex(index, "ViewportIndexedf_State")) return;
+        if (!ValidateNonNegativeExtent(w, h, "ViewportIndexedf_State")) return;
+
+        MG_State::pGLContext->SetViewportIndexed(index, FloatVec4(x, y, w, h));
+    }
+
+    void ScissorArrayv_State(GLuint first, GLsizei count, const GLint* v) {
+        if (!ValidateViewportRange(first, count, "ScissorArrayv_State")) return;
+        if (count == 0) return;
+        if (!ValidateNonNullArray(v, "ScissorArrayv_State")) return;
+        if (!ValidateArrayExtents(count, v, "ScissorArrayv_State")) return;
+
+        for (GLsizei i = 0; i < count; ++i) {
+            MG_State::pGLContext->SetScissorBoxIndexed(first + static_cast<GLuint>(i),
+                                                       IntVec4(v[i * 4 + 0], v[i * 4 + 1], v[i * 4 + 2], v[i * 4 + 3]));
+        }
+    }
+
+    void ScissorIndexed_State(GLuint index, GLint left, GLint bottom, GLsizei width, GLsizei height) {
+        if (!ValidateViewportIndex(index, "ScissorIndexed_State")) return;
+        if (!ValidateNonNegativeExtent(width, height, "ScissorIndexed_State")) return;
+
+        MG_State::pGLContext->SetScissorBoxIndexed(index, IntVec4(left, bottom, width, height));
+    }
+
+    void DepthRangeArrayv_State(GLuint first, GLsizei count, const GLdouble* v) {
+        if (!ValidateViewportRange(first, count, "DepthRangeArrayv_State")) return;
+        if (count == 0) return;
+        if (!ValidateNonNullArray(v, "DepthRangeArrayv_State")) return;
+
+        for (GLsizei i = 0; i < count; ++i) {
+            MG_State::pGLContext->SetDepthRangeIndexed(
+                first + static_cast<GLuint>(i),
+                FloatVec2(ClampUnitFloat(static_cast<GLfloat>(v[i * 2 + 0])),
+                          ClampUnitFloat(static_cast<GLfloat>(v[i * 2 + 1]))));
+        }
+    }
+
+    void DepthRangeIndexed_State(GLuint index, GLdouble n, GLdouble f) {
+        if (!ValidateViewportIndex(index, "DepthRangeIndexed_State")) return;
+
+        MG_State::pGLContext->SetDepthRangeIndexed(
+            index, FloatVec2(ClampUnitFloat(static_cast<GLfloat>(n)), ClampUnitFloat(static_cast<GLfloat>(f))));
+    }
+
     void StencilOpSeparate_State(GLenum face, GLenum sfail, GLenum dpfail, GLenum dppass) {
         Bool applyFront = false;
         Bool applyBack = false;
@@ -174,12 +319,7 @@ namespace MobileGL::MG_Impl::GLImpl {
     }
 
     void Scissor_State(GLint x, GLint y, GLsizei width, GLsizei height) {
-        if (width < 0 || height < 0) {
+        if (!ValidateNonNegativeExtent(width, height, "Scissor_State")) return;
-            MG_State::pGLContext->RecordError(ErrorCode::InvalidValue,
-                                              MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", "Scissor_State",
-                                                                           "Width abd height must be non-negative."));
-            return;
-        }
 
         MG_State::pGLContext->SetScissorBox(IntVec4(x, y, width, height));
     }
@@ -335,7 +475,7 @@ namespace MobileGL::MG_Impl::GLImpl {
     }
 
     GLboolean IsEnabledi_State(GLenum target, GLuint index) {
-        if (!ValidateIndexedBlendCapability(target, index, "IsEnabledi_State")) {
+        if (!ValidateIndexedCapability(target, index, "IsEnabledi_State")) {
             return GL_FALSE;
         }
 
@@ -380,7 +520,25 @@ namespace MobileGL::MG_Impl::GLImpl {
             return;
         }
 
+        // 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);
+        // The ARB_viewport_array rectangles are the only multi-component indexed state that
+        // reaches here; writing element 0 alone would leave the caller's other three untouched.
+        const GLsizei components = target == GL_VIEWPORT || target == GL_SCISSOR_BOX
+            ? 4
+            : (target == GL_DEPTH_RANGE ? 2 : 1);
+        for (GLsizei i = 0; i < components; ++i) {
+            data[i] = values[i] != 0 ? GL_TRUE : GL_FALSE;
+        }
     }
 
     GLboolean IsEnabled_State(GLenum cap) {
@@ -713,7 +871,7 @@ namespace MobileGL::MG_Impl::GLImpl {
     }
 
     void Disablei_State(GLenum target, GLuint index) {
-        if (!ValidateIndexedBlendCapability(target, index, "Disablei_State")) {
+        if (!ValidateIndexedCapability(target, index, "Disablei_State")) {
             return;
         }
 
@@ -731,7 +889,7 @@ namespace MobileGL::MG_Impl::GLImpl {
     }
 
     void Enablei_State(GLenum target, GLuint index) {
-        if (!ValidateIndexedBlendCapability(target, index, "Enablei_State")) {
+        if (!ValidateIndexedCapability(target, index, "Enablei_State")) {
             return;
         }
 
@@ -785,6 +943,44 @@ namespace MobileGL::MG_Impl::GLImpl {
         Viewport_State(x, y, width, height);
     }
 
+    void ViewportArrayv(GLuint first, GLsizei count, const GLfloat* v) {
+        ViewportArrayv_State(first, count, v);
+    }
+
+    void ViewportIndexedf(GLuint index, GLfloat x, GLfloat y, GLfloat w, GLfloat h) {
+        ViewportIndexedf_State(index, x, y, w, h);
+    }
+
+    void ViewportIndexedfv(GLuint index, const GLfloat* v) {
+        // The index is validated before the pointer is touched: glViewportIndexedfv(MAX, nullptr)
+        // must be one GL_INVALID_VALUE, not a null dereference.
+        if (!ValidateViewportIndex(index, "ViewportIndexedfv")) return;
+        if (!ValidateNonNullArray(v, "ViewportIndexedfv")) return;
+        ViewportIndexedf_State(index, v[0], v[1], v[2], v[3]);
+    }
+
+    void ScissorArrayv(GLuint first, GLsizei count, const GLint* v) {
+        ScissorArrayv_State(first, count, v);
+    }
+
+    void ScissorIndexed(GLuint index, GLint left, GLint bottom, GLsizei width, GLsizei height) {
+        ScissorIndexed_State(index, left, bottom, width, height);
+    }
+
+    void ScissorIndexedv(GLuint index, const GLint* v) {
+        if (!ValidateViewportIndex(index, "ScissorIndexedv")) return;
+        if (!ValidateNonNullArray(v, "ScissorIndexedv")) return;
+        ScissorIndexed_State(index, v[0], v[1], v[2], v[3]);
+    }
+
+    void DepthRangeArrayv(GLuint first, GLsizei count, const GLdouble* v) {
+        DepthRangeArrayv_State(first, count, v);
+    }
+
+    void DepthRangeIndexed(GLuint index, GLdouble n, GLdouble f) {
+        DepthRangeIndexed_State(index, n, f);
+    }
+
     void StencilOpSeparate(GLenum face, GLenum sfail, GLenum dpfail, GLenum dppass) {
         StencilOpSeparate_State(face, sfail, dpfail, dppass);
     }

MobileGL/MG_Impl/GLImpl/RenderState/GL_RenderState.h

@@ -20,6 +20,16 @@ namespace MobileGL::MG_Impl::GLImpl {
     void Enablei(GLenum target, GLuint index);
     void BlendFunc(GLenum sfactor, GLenum dfactor);
     void Viewport(GLint x, GLint y, GLsizei width, GLsizei height);
+    // ARB_viewport_array (core since GL 4.1). Every one of these addresses the same 16-element
+    // indexed state the classic glViewport/glScissor/glDepthRange trio broadcasts to.
+    void ViewportArrayv(GLuint first, GLsizei count, const GLfloat* v);
+    void ViewportIndexedf(GLuint index, GLfloat x, GLfloat y, GLfloat w, GLfloat h);
+    void ViewportIndexedfv(GLuint index, const GLfloat* v);
+    void ScissorArrayv(GLuint first, GLsizei count, const GLint* v);
+    void ScissorIndexed(GLuint index, GLint left, GLint bottom, GLsizei width, GLsizei height);
+    void ScissorIndexedv(GLuint index, const GLint* v);
+    void DepthRangeArrayv(GLuint first, GLsizei count, const GLdouble* v);
+    void DepthRangeIndexed(GLuint index, GLdouble n, GLdouble f);
     void StencilOpSeparate(GLenum face, GLenum sfail, GLenum dpfail, GLenum dppass);
     void StencilOp(GLenum fail, GLenum zfail, GLenum zpass);
     void StencilMaskSeparate(GLenum face, GLuint mask);

MobileGL/MG_Impl/GLImpl/Sampler/GL_Sampler.cpp

@@ -9,6 +9,7 @@
 #include "GL_Sampler.h"
 #include "Validators.h"
 #include "../Getter/GL_Getter.h"
+#include "../Texture/GL_Texture.h"
 #include <MG_State/GLState/Core.h>
 #include <MG_Util/Converters/GLToMG/TextureEnumConverter.h>
 #include <MG_Util/Converters/MGToGL/TextureEnumConverter.h>
@@ -269,15 +270,13 @@ namespace MobileGL::MG_Impl::GLImpl {
         }
     }
 
-    // The number of texture units a sampler may be bound to. GL 3.3 core 3.8.2 names
+    // The number of texture units a sampler may be bound to is the same count a TEXTURE may be
-    // GL_MAX_COMBINED_TEXTURE_IMAGE_UNITS, which is what the backend advertises; the frontend's
+    // bound to - GL 3.3 core 3.8.2 names GL_MAX_COMBINED_TEXTURE_IMAGE_UNITS for both - so it is
-    // MAX_TEXTURE_IMAGE_UNITS is only the capacity of the unit array, so it is a clamp on the
+    // computed once, in GetCombinedTextureImageUnitCount, and named here for the sampler-side
-    // answer and never the answer itself - gating on it alone accepts every unit up to 192 no
+    // readers below. Two copies of that arithmetic is how glBindSamplers and glBindTextures would
-    // matter what the driver reports.
+    // come to disagree about which units exist.
     static GLint GetSamplerBindableTextureUnitCount() {
-        GLint maxTextureUnits = 0;
+        return GetCombinedTextureImageUnitCount();
-        GetIntegerv(GL_MAX_COMBINED_TEXTURE_IMAGE_UNITS, &maxTextureUnits);
-        return std::min<GLint>(std::max(maxTextureUnits, 0), MG_State::GLState::TextureState::MAX_TEXTURE_IMAGE_UNITS);
     }
 
     void BindSampler_State(GLuint unit, GLuint sampler) {
@@ -336,8 +335,22 @@ 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_W_ONCE("%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) {
@@ -853,7 +870,7 @@ namespace MobileGL::MG_Impl::GLImpl {
                 MG_Util::GetInputBytesPerPixel(MG_Util::ConvertGLEnumToTextureInputFormat(format),
                                                MG_Util::ConvertGLEnumToTexturePixelDataType(type));
             if (readBytesPerTexel != bytesPerTexel) {
-                MGLOG_I("%s: cannot copy into a %zu-byte texel from a %zu-byte readback layout", caller,
+                MGLOG_W_ONCE("%s: cannot copy into a %zu-byte texel from a %zu-byte readback layout", caller,
                         bytesPerTexel, readBytesPerTexel);
                 return false;
             }
@@ -1359,6 +1376,47 @@ namespace MobileGL::MG_Impl::GLImpl {
         return true;
     }
 
+    // The same rules for the COMPRESSED entry points, whose payload size is the imageSize the
+    // caller passed rather than something derived from a (format, type) pair - and which have no
+    // datum size, so the alignment rule above does not apply to them. Shared by
+    // glCompressedTexImage2D and glCompressedTexSubImage2D so the two cannot drift; the point
+    // that is easy to get wrong and that KHR-GL44.buffer_storage.map_persistent_texture exists to
+    // check is the first one: a PERSISTENT mapping stays a legal transfer source.
+    Bool ValidateCompressedUnpackBufferSource(const void* data, SizeT imageSize, const char* caller) {
+        const auto& unpackBuffer =
+            MG_State::pGLContext->GetBufferBindingSlot(BufferTarget::PixelUnpack).GetBoundObject();
+        if (!unpackBuffer) return true;
+
+        if (unpackBuffer->IsMapped() && !(unpackBuffer->GetMappingAccess() & BufferMappingAccessBit::Persistent)) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidOperation,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", caller, "Pixel unpack buffer is currently mapped."));
+            return false;
+        }
+
+        const SizeT offset = reinterpret_cast<SizeT>(data);
+        const SizeT bufferSize = unpackBuffer->GetSize();
+        if (offset > bufferSize || imageSize > bufferSize - offset) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidOperation,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", caller,
+                                             "Unpacking would read past the end of the pixel unpack buffer."));
+            return false;
+        }
+        return true;
+    }
+
+    // Where a compressed upload reads its blocks from: `data` is an offset into the bound unpack
+    // buffer when there is one, and a client pointer otherwise. Only meaningful once
+    // ValidateCompressedUnpackBufferSource has passed. Null means there is nothing to read, which
+    // GL leaves undefined and which callers must not dereference.
+    const void* CompressedUnpackSource(const void* data) {
+        const auto& unpackBuffer =
+            MG_State::pGLContext->GetBufferBindingSlot(BufferTarget::PixelUnpack).GetBoundObject();
+        if (!unpackBuffer) return data;
+        return reinterpret_cast<const char*>(unpackBuffer->MappedData()) + reinterpret_cast<SizeT>(data);
+    }
+
     void TexSubImage3D_State(GLenum target, GLint level, GLint xoffset, GLint yoffset, GLint zoffset, GLsizei width,
                              GLsizei height, GLsizei depth, GLenum format, GLenum type, const void* pixels) {
         TextureUploadTarget textureUploadTarget = MG_Util::ConvertGLEnumToTextureUploadTarget(target);
@@ -1432,7 +1490,7 @@ namespace MobileGL::MG_Impl::GLImpl {
         if (xoffset + width > static_cast<GLsizei>(texelSize.x()) ||
             yoffset + height > static_cast<GLsizei>(texelSize.y()) ||
             zoffset + depth > static_cast<GLsizei>(texelSize.z())) {
-            MGLOG_E("TexSubImage3D_State: Specified region exceeds texture level dimensions");
+            MGLOG_E_ONCE("TexSubImage3D_State: Specified region exceeds texture level dimensions");
             free(processedPixels);
             return;
         }
@@ -1541,7 +1599,7 @@ namespace MobileGL::MG_Impl::GLImpl {
             {width, height, 1}, false, inputSize);
 
         if (!processedPixels || inputSize == 0) {
-            MGLOG_E("TexSubImage2D_State: Failed to process pixel data for TexSubImage2D, width: %d, height: %d", width,
+            MGLOG_E_ONCE("TexSubImage2D_State: Failed to process pixel data for TexSubImage2D, width: %d, height: %d", width,
                     height);
             if (processedPixels) free(processedPixels);
             return;
@@ -1555,7 +1613,7 @@ namespace MobileGL::MG_Impl::GLImpl {
 
         if (xoffset + width > static_cast<GLsizei>(texelSize.x()) ||
             yoffset + height > static_cast<GLsizei>(texelSize.y())) {
-            MGLOG_E("TexSubImage2D_State: Specified region exceeds texture dimensions");
+            MGLOG_E_ONCE("TexSubImage2D_State: Specified region exceeds texture dimensions");
             free(processedPixels);
             return;
         }
@@ -2106,7 +2164,7 @@ namespace MobileGL::MG_Impl::GLImpl {
 
         if (processedPixels && imageSize > 0) {
             if (imageSize != internalBytes) {
-                MGLOG_W("%s: Processed pixel data size (%zu) does not match expected size (%zu). "
+                MGLOG_W_ONCE("%s: Processed pixel data size (%zu) does not match expected size (%zu). "
                         "This may indicate an alignment or processing issue.",
                         __func__, imageSize, internalBytes);
             }
@@ -2221,6 +2279,23 @@ namespace MobileGL::MG_Impl::GLImpl {
             DiscardMipmapChainOnBaseRespecification(textureMipmapObject, textureUploadTarget, level);
             textureMipmapObject->AllocateStorage(textureUploadTarget, level,
                                                  {{width, height, 1}, internalBytes});
+            // GL 4.6 core 8.5: a SPECIFIC compressed internalformat (unlike a generic
+            // GL_COMPRESSED_* one, where the implementation is free to choose) commits the
+            // level to that format - GL_TEXTURE_COMPRESSED must then answer true for it and
+            // GL_TEXTURE_INTERNAL_FORMAT must report it, which is how an application asks for
+            // the size to hand glCompressedTexSubImage2D afterwards. Only the tag and the size
+            // are recorded: there is no BC/ETC codec here, so the texel shadow keeps the
+            // uncompressed storage this format resolved to (which is also what lets the level
+            // sample as the application's texels), and the compressed image the tag describes
+            // is zero-filled - the one reproducible answer glGetCompressedTexImage can give for
+            // an image nothing ever compressed. AllocateStorage above clears the tag, so this
+            // has to follow it.
+            const auto compressedInfo = MG_Util::GetCompressedFormatInfo(static_cast<GLenum>(internalformat));
+            if (compressedInfo.blockWidth != 0) {
+                textureMipmapObject->SetMipmapCompressedImage(
+                    textureUploadTarget, level, static_cast<GLenum>(internalformat), nullptr,
+                    MG_Util::CalculateCompressedTextureImageSize(compressedInfo, {width, height, 1}));
+            }
         }
 
         if (!originalPixels) {
@@ -2235,7 +2310,7 @@ namespace MobileGL::MG_Impl::GLImpl {
 
         if (processedPixels && imageSize > 0) {
             if (imageSize != internalBytes) {
-                MGLOG_W("TexImage2D_State: Processed pixel data size (%zu) does not match expected size (%zu). "
+                MGLOG_W_ONCE("TexImage2D_State: Processed pixel data size (%zu) does not match expected size (%zu). "
                         "This may indicate an alignment or processing issue.",
                         imageSize, internalBytes);
             }
@@ -2910,7 +2985,8 @@ namespace MobileGL::MG_Impl::GLImpl {
                     break;
                 }
                 default:
-                    THROW_UNIMPL_EXCEPTION;
+                    RecordUnsupportedLevelQueryStorage("GetTexLevelParameteriv_State", pname);
+                    break;
                 }
             }
             break;
@@ -2924,7 +3000,8 @@ namespace MobileGL::MG_Impl::GLImpl {
                     break;
                 }
                 default:
-                    THROW_UNIMPL_EXCEPTION;
+                    RecordUnsupportedLevelQueryStorage("GetTexLevelParameteriv_State", pname);
+                    break;
                 }
             }
             break;
@@ -2938,16 +3015,17 @@ namespace MobileGL::MG_Impl::GLImpl {
                     break;
                 }
                 default:
-                    THROW_UNIMPL_EXCEPTION;
+                    RecordUnsupportedLevelQueryStorage("GetTexLevelParameteriv_State", pname);
+                    break;
                 }
             }
             break;
         case GL_TEXTURE_INTERNAL_FORMAT:
             if (params) {
                 // A level stored compressed must report the token it was given, not the
-                // uncompressed format backing it (GL 4.6 core 8.11). Only glCompressedTexImage* sets
+                // uncompressed format backing it (GL 4.6 core 8.11). glCompressedTexImage2D sets
-                // that tag, so every level created by glTexImage*D - including one given a compressed
+                // that tag, and so does a glTexImage2D given a SPECIFIC compressed internalformat;
-                // internalformat - still answers with its resolved storage format.
+                // every other level answers with its resolved storage format.
                 const GLenum compressedFormat = GetCompressedLevelFormat(textureObject, textureUploadTarget, level);
                 *params = (compressedFormat != GL_NONE)
                               ? (GLint)compressedFormat
@@ -3045,7 +3123,8 @@ namespace MobileGL::MG_Impl::GLImpl {
                     break;
                 }
                 default:
-                    THROW_UNIMPL_EXCEPTION;
+                    RecordUnsupportedLevelQueryStorage("GetTexLevelParameterfv_State", pname);
+                    break;
                 }
             }
             break;
@@ -3059,7 +3138,8 @@ namespace MobileGL::MG_Impl::GLImpl {
                     break;
                 }
                 default:
-                    THROW_UNIMPL_EXCEPTION;
+                    RecordUnsupportedLevelQueryStorage("GetTexLevelParameterfv_State", pname);
+                    break;
                 }
             }
             break;
@@ -3073,16 +3153,17 @@ namespace MobileGL::MG_Impl::GLImpl {
                     break;
                 }
                 default:
-                    THROW_UNIMPL_EXCEPTION;
+                    RecordUnsupportedLevelQueryStorage("GetTexLevelParameterfv_State", pname);
+                    break;
                 }
             }
             break;
         case GL_TEXTURE_INTERNAL_FORMAT:
             if (params) {
                 // A level stored compressed must report the token it was given, not the
-                // uncompressed format backing it (GL 4.6 core 8.11). Only glCompressedTexImage* sets
+                // uncompressed format backing it (GL 4.6 core 8.11). glCompressedTexImage2D sets
-                // that tag, so every level created by glTexImage*D - including one given a compressed
+                // that tag, and so does a glTexImage2D given a SPECIFIC compressed internalformat;
-                // internalformat - still answers with its resolved storage format.
+                // every other level answers with its resolved storage format.
                 const GLenum compressedFormat = GetCompressedLevelFormat(textureObject, textureUploadTarget, level);
                 *params = (GLfloat)((compressedFormat != GL_NONE)
                                         ? compressedFormat
@@ -3301,12 +3382,84 @@ namespace MobileGL::MG_Impl::GLImpl {
                          dstY, dstZ, srcWidth, srcHeight, srcDepth);
     }
 
+    namespace {
+        // The eleven targets GL 4.6 core 18.3.2 accepts. GL_TEXTURE_BUFFER, the six cube FACE
+        // enums and every PROXY enum all convert to a TextureTarget this frontend recognises,
+        // so ValidateTextureTarget lets them through; here they are INVALID_ENUM.
+        Bool ValidateCopyImageTarget(GLenum target, const char* endpointName) {
+            switch (target) {
+            case GL_RENDERBUFFER:
+            case GL_TEXTURE_1D:
+            case GL_TEXTURE_1D_ARRAY:
+            case GL_TEXTURE_2D:
+            case GL_TEXTURE_2D_ARRAY:
+            case GL_TEXTURE_2D_MULTISAMPLE:
+            case GL_TEXTURE_2D_MULTISAMPLE_ARRAY:
+            case GL_TEXTURE_3D:
+            case GL_TEXTURE_CUBE_MAP:
+            case GL_TEXTURE_CUBE_MAP_ARRAY:
+            case GL_TEXTURE_RECTANGLE:
+                return true;
+            default:
+                break;
+            }
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidEnum,
+                MakeUnique<GenericErrorInfo>(
+                    "MG_Impl/GLImpl", "ValidateCopyImageSubData_State",
+                    std::format("{} is not a target glCopyImageSubData accepts as the {}.",
+                                MG_Util::ConvertGLEnumToString(target), endpointName)));
+            return false;
+        }
+
+        IntVec3 GetCopyImageLevelSize(const SharedPtr<MG_State::GLState::ITextureObject>& textureObject,
+                                      TextureUploadTarget uploadTarget, GLint level) {
+            const auto* mipmapTexture = MG_State::GLState::AsMipmapTexture(textureObject.get());
+            if (!mipmapTexture) return textureObject->GetBaseSize();
+            return mipmapTexture->GetMipmapTexelSize(uploadTarget, static_cast<Uint>(level));
+        }
+
+        // glCopyImageSubData names an object that must already exist, and GL 4.6 core 18.3.2
+        // spells the failure INVALID_VALUE - "if either name does not correspond to a valid
+        // object". The shared ValidateTextureObject says INVALID_OPERATION, which is right for
+        // the ~30 entry points that reach it through a BOUND object (where the name was never
+        // in question and the fault is the binding), so this is a local rule rather than a
+        // change to the helper.
+        Bool ValidateCopyImageObjectExists(const SharedPtr<MG_State::GLState::ITextureObject>& textureObject,
+                                           const char* endpointName) {
+            if (textureObject) return true;
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidValue,
+                MakeUnique<GenericErrorInfo>(
+                    "MG_Impl/GLImpl", "ValidateCopyImageSubData_State",
+                    std::format("The {} name does not correspond to an existing image object.", endpointName)));
+            return false;
+        }
+
+        // Same split for the target/object disagreement: GL 4.6 core 18.3.2 makes a target that
+        // does not match the object INVALID_ENUM, where the shared uniformity helper records
+        // INVALID_OPERATION for the upload paths that share it.
+        Bool ValidateCopyImageTargetMatchesObject(const SharedPtr<MG_State::GLState::ITextureObject>& textureObject,
+                                                  TextureTarget target, const char* endpointName) {
+            if (!textureObject || textureObject->GetTarget() == target) return true;
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidEnum,
+                MakeUnique<GenericErrorInfo>(
+                    "MG_Impl/GLImpl", "ValidateCopyImageSubData_State",
+                    std::format("The {} target {} does not match the target the object was created with ({}).",
+                                endpointName, MG_Util::ConvertTextureTargetToString(target),
+                                MG_Util::ConvertTextureTargetToString(textureObject->GetTarget()))));
+            return false;
+        }
+    } // namespace
+
     Bool ValidateCopyImageSubData_State(const SharedPtr<MG_State::GLState::ITextureObject>& srcTexture,
-                                        GLenum srcTarget, GLint srcLevel,
+                                        GLenum srcTarget, GLint srcLevel, GLint srcX, GLint srcY,
                                         const SharedPtr<MG_State::GLState::ITextureObject>& dstTexture,
-                                        GLenum dstTarget, GLint dstLevel,
+                                        GLenum dstTarget, GLint dstLevel, GLint dstX, GLint dstY,
                                         GLsizei srcWidth, GLsizei srcHeight, GLsizei srcDepth) {
-        if (!TextureImpl::ValidateTextureObject(srcTexture) || !TextureImpl::ValidateTextureObject(dstTexture)) {
+        if (!ValidateCopyImageObjectExists(srcTexture, "source") ||
+            !ValidateCopyImageObjectExists(dstTexture, "destination")) {
             return false;
         }
         const auto srcTextureTarget = MG_Util::ConvertGLEnumToTextureTarget(srcTarget);
@@ -3315,14 +3468,30 @@ namespace MobileGL::MG_Impl::GLImpl {
             !TextureImpl::ValidateTextureTarget(dstTextureTarget)) {
             return false;
         }
-        if (!TextureImpl::ValidateTextureTargetUniformity(srcTexture, srcTextureTarget) ||
+        // GL_TEXTURE_BUFFER and the cube FACE enums convert to a target this frontend knows, but
-            !TextureImpl::ValidateTextureTargetUniformity(dstTexture, dstTextureTarget)) {
+        // 18.3.2 does not accept them here - only the eleven whole-image targets do.
+        if (!ValidateCopyImageTarget(srcTarget, "source") || !ValidateCopyImageTarget(dstTarget, "destination")) {
+            return false;
+        }
+        if (!ValidateCopyImageTargetMatchesObject(srcTexture, srcTextureTarget, "source") ||
+            !ValidateCopyImageTargetMatchesObject(dstTexture, dstTextureTarget, "destination")) {
             return false;
         }
         if (!TextureImpl::ValidateTextureLevelNumber(srcLevel) ||
             !TextureImpl::ValidateTextureLevelNumber(dstLevel)) {
             return false;
         }
+        // ValidateTextureLevelNumber only bounds the index by GL_MAX_TEXTURE_SIZE; it cannot
+        // see that this particular texture stops at level 0. Both backends turn <level> into an
+        // image subresource with no further checking (DirectVulkan builds a VkImageCopy from it,
+        // DirectGLES forwards it to the ES copy), so a level the texture never had reached the
+        // driver as an out-of-range mip index - on Adreno that is a SIGSEGV inside
+        // vkCmdCopyImage, which is what KHR-GL43.copy_image.non_existent_mipmap used to do to
+        // the whole glcts process. The answer the spec asks for is GL_INVALID_VALUE.
+        if (!TextureImpl::ValidateTextureLevelExists(srcTexture, srcLevel, __func__) ||
+            !TextureImpl::ValidateTextureLevelExists(dstTexture, dstLevel, __func__)) {
+            return false;
+        }
         if (srcWidth < 0 || srcHeight < 0 || srcDepth < 0) {
             MG_State::pGLContext->RecordError(
                 ErrorCode::InvalidValue,
@@ -3333,7 +3502,44 @@ namespace MobileGL::MG_Impl::GLImpl {
         if (srcWidth == 0 || srcHeight == 0 || srcDepth == 0) {
             return false;
         }
-        if (!TextureImpl::ValidateBaseInternalFormatMatch(srcTexture->GetFormat(), dstTexture->GetFormat())) {
+        // A multisample image can only be copied to one with the same sample count, and a
+        // single-sample image reports zero - so this one comparison is also what rejects
+        // copying between a multisample target and a non-multisample one.
+        if (srcTexture->GetSamples() != dstTexture->GetSamples()) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidOperation,
+                MakeUnique<GenericErrorInfo>(
+                    "MG_Impl/GLImpl", __func__,
+                    std::format("The two images have different sample counts ({} vs. {}).",
+                                srcTexture->GetSamples(), dstTexture->GetSamples())));
+            return false;
+        }
+        // 18.3.2: both images must be complete. An incomplete one has no defined texels to copy
+        // and no defined storage to copy into.
+        if (!srcTexture->IsComplete() || !dstTexture->IsComplete()) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidOperation,
+                MakeUnique<GenericErrorInfo>(
+                    "MG_Impl/GLImpl", __func__,
+                    std::format("A copied image is incomplete (source complete: {}, destination complete: {}).",
+                                srcTexture->IsComplete(), dstTexture->IsComplete())));
+            return false;
+        }
+        const auto srcUploadTarget = GetPrimaryUploadTarget(srcTexture);
+        const auto dstUploadTarget = GetPrimaryUploadTarget(dstTexture);
+        const auto srcBlock = TextureImpl::ResolveCopyImageTexelBlock(
+            srcTexture->GetFormat(), GetCompressedLevelFormat(srcTexture, srcUploadTarget, srcLevel));
+        const auto dstBlock = TextureImpl::ResolveCopyImageTexelBlock(
+            dstTexture->GetFormat(), GetCompressedLevelFormat(dstTexture, dstUploadTarget, dstLevel));
+        if (!TextureImpl::ValidateCopyImageFormatCompatibility(srcBlock, dstBlock)) {
+            return false;
+        }
+        const IntVec3 srcLevelSize = GetCopyImageLevelSize(srcTexture, srcUploadTarget, srcLevel);
+        const IntVec3 dstLevelSize = GetCopyImageLevelSize(dstTexture, dstUploadTarget, dstLevel);
+        if (!TextureImpl::ValidateCopyImageBlockAlignment(srcBlock, srcX, srcY, srcWidth, srcHeight,
+                                                          srcLevelSize.x(), srcLevelSize.y(), "source") ||
+            !TextureImpl::ValidateCopyImageBlockAlignment(dstBlock, dstX, dstY, srcWidth, srcHeight,
+                                                          dstLevelSize.x(), dstLevelSize.y(), "destination")) {
             return false;
         }
         return true;
@@ -3403,7 +3609,10 @@ 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 +3635,13 @@ namespace MobileGL::MG_Impl::GLImpl {
 
     void CopyTexImage1D_State(GLenum target, GLint level, GLenum internalformat, GLint x, GLint y, GLsizei width,
                               GLint border) {
-        // TODO: implement
+        // 1D textures are not implemented by this backend set. Record the error the way every
-        THROW_UNIMPL_EXCEPTION;
+        // 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,
@@ -3437,10 +3651,170 @@ namespace MobileGL::MG_Impl::GLImpl {
         RecordUnsupportedCompressedFormat(__func__);
     }
 
+    // Replaces a block-aligned rectangle of the compressed image glCompressedTexImage2D (or a
+    // compressed glTexImage2D) shadowed for this level. Same deviation as the image call it
+    // patches: the uncompressed texel shadow beside it is NOT touched, because there is no
+    // BC/ETC codec here to decode the incoming blocks with - so what changes is the image
+    // glGetCompressedTexImage hands back, not what the level samples as. Marking the texels
+    // dirty would therefore only re-upload bytes that did not change.
     void CompressedTexSubImage2D_State(GLenum target, GLint level, GLint xoffset, GLint yoffset, GLsizei width,
                                        GLsizei height, GLenum format, GLsizei imageSize, const void* data) {
-        // TODO: implement compressed upload - see CompressedTexImage2D_State.
+        // ======================= Converting ================================
+        const auto textureUploadTarget = MG_Util::ConvertGLEnumToTextureUploadTarget(target);
+        const auto textureTarget = MG_Util::ConvertGLEnumToTextureTarget(target);
+        // Zero block width doubles as "format is not a specific compressed format", the
+        // INVALID_ENUM case - one lookup answers both questions.
+        const auto compressedInfo = MG_Util::GetCompressedFormatInfo(format);
+
+        // ===================== Error Checking ==============================
+        if (!TextureImpl::ValidateTextureUploadTarget(textureUploadTarget)) return;
+        // A proxy holds no image to modify; only the glTexImage*/glCompressedTexImage* pair
+        // accepts one.
+        if (TextureImpl::IsProxyTextureTarget(textureUploadTarget)) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidEnum,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", __func__,
+                                             "A proxy target has no texture image to modify."));
+            return;
+        }
+        if (!TextureImpl::ValidateTextureLevelNumber(level)) return;
+        if (!TextureImpl::ValidateTextureLevelWithUploadTarget(textureUploadTarget, level)) return;
+        if (width < 0 || height < 0) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidValue,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", __func__, "width and height must be non-negative."));
+            return;
+        }
+        if (compressedInfo.blockWidth == 0) {
             RecordUnsupportedCompressedFormat(__func__);
+            return;
+        }
+
+        auto& textureObject = GetTextureObjectByTarget(textureUploadTarget, textureTarget);
+        if (!TextureImpl::ValidateTextureObject(textureObject)) return;
+        auto* textureMipmapObject = MG_State::GLState::AsMipmapTexture(textureObject.get());
+        if (textureMipmapObject == nullptr) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidOperation,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", __func__, "Texture storage is not mipmap-backed."));
+            return;
+        }
+        // GL 4.6 core 8.7: INVALID_OPERATION unless the image being modified is stored in
+        // exactly this compressed format. That is also what makes the block arithmetic below
+        // sound - the level's grid is measured with THIS format's block size.
+        const GLenum levelFormat =
+            textureMipmapObject->GetMipmapCompressedFormat(textureUploadTarget, static_cast<Uint>(level));
+        if (levelFormat != format) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidOperation,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", __func__,
+                                             "format does not match the internal format of the texture image."));
+            return;
+        }
+
+        const IntVec3 levelSize = textureMipmapObject->GetMipmapTexelSize(textureUploadTarget, static_cast<Uint>(level));
+        // Written as a subtraction rather than `xoffset + width > levelSize.x()`: both operands
+        // are application-supplied GLints, so the sum is free to overflow, and a signed overflow
+        // is undefined behaviour that a compiler may resolve by assuming the check passes.
+        // levelSize is our own and non-negative, and the offsets are known non-negative by the
+        // time the subtraction runs, so this form cannot wrap.
+        if (xoffset < 0 || yoffset < 0 || width > levelSize.x() - xoffset || height > levelSize.y() - yoffset) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidValue,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", __func__,
+                                             "The replaced region does not lie within the texture image."));
+            return;
+        }
+        // GL 4.6 core 8.7 for block-based formats: the region must start on a block boundary
+        // and must either be a whole number of blocks wide/high or run to the image's edge.
+        const Int blockWidth = static_cast<Int>(compressedInfo.blockWidth);
+        const Int blockHeight = static_cast<Int>(compressedInfo.blockHeight);
+        const Bool alignedX = (xoffset % blockWidth == 0) &&
+                              (width % blockWidth == 0 || xoffset + width == levelSize.x());
+        const Bool alignedY = (yoffset % blockHeight == 0) &&
+                              (height % blockHeight == 0 || yoffset + height == levelSize.y());
+        if (!alignedX || !alignedY) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidOperation,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", __func__,
+                                             "The replaced region is not aligned to the format's compressed blocks."));
+            return;
+        }
+        // Exactly the size the format and dimensions imply, which is also what keeps the copy
+        // below in bounds.
+        const SizeT expectedImageSize =
+            MG_Util::CalculateCompressedTextureImageSize(compressedInfo, {width, height, 1});
+        if (imageSize < 0 || static_cast<SizeT>(imageSize) != expectedImageSize) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidValue,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", __func__,
+                                             "imageSize does not match the compressed image size."));
+            return;
+        }
+
+        // ======================= Processing ================================
+        if (!ValidateCompressedUnpackBufferSource(data, expectedImageSize, __func__)) return;
+        const void* compressedBytes = CompressedUnpackSource(data);
+        if (expectedImageSize == 0) return; // a zero-sized region is a legal no-op
+        if (compressedBytes == nullptr) {
+            // No unpack buffer and a null client pointer: there is nothing to read. GL leaves
+            // this undefined rather than erroring, and dereferencing it is the one answer that
+            // is never acceptable.
+            MGLOG_D("%s: null data with no pixel unpack buffer bound, nothing to replace", __func__);
+            return;
+        }
+
+        // Once per process: the call is about to succeed, and what it does is narrower than what
+        // an application has every right to expect from it. Before this existed the call answered
+        // GL_INVALID_ENUM, which was wrong but at least visible; a silent success that leaves the
+        // sampled texels untouched is the kind of thing that costs a day to find from the other
+        // end. MGLOG_W is the right level and now survives at INFO; it sat at MGLOG_I only
+        // while the Log.h ordering compiled warnings out of the builds that ship.
+        static std::atomic<Bool> announcedNoCodec{false};
+        if (!announcedNoCodec.exchange(true)) {
+            MGLOG_W("%s: the compressed blocks are stored verbatim and returned by "
+                    "glGetCompressedTexImage, but there is no BC/ETC decoder here, so they do not "
+                    "reach the texels this level SAMPLES as. Upload through glTexSubImage2D for "
+                    "that.",
+                    __func__);
+        }
+
+        // The level's compressed image is stored as one blob, so the rectangle is patched into
+        // a copy of it and the whole thing handed back. Compressed sub-image uploads are not a
+        // hot path, and this keeps the storage layer's compressed API to the two calls it has.
+        const SizeT blobSize =
+            textureMipmapObject->GetMipmapCompressedByteSize(textureUploadTarget, static_cast<Uint>(level));
+        const void* existing =
+            textureMipmapObject->MapMipmapCompressedImage(textureUploadTarget, static_cast<Uint>(level));
+        if (blobSize == 0 || existing == nullptr) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidOperation,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", __func__,
+                                             "The texture level holds no compressed image to modify."));
+            return;
+        }
+        Vector<Uint8> blob(blobSize);
+        Memcpy(blob.data(), existing, blobSize);
+
+        const SizeT blockByteSize = compressedInfo.blockByteSize;
+        const SizeT levelBlocksX = (static_cast<SizeT>(levelSize.x()) + compressedInfo.blockWidth - 1) /
+                                   compressedInfo.blockWidth;
+        const SizeT levelRowBytes = levelBlocksX * blockByteSize;
+        const SizeT regionBlocksX = (static_cast<SizeT>(width) + compressedInfo.blockWidth - 1) /
+                                    compressedInfo.blockWidth;
+        const SizeT regionBlocksY = (static_cast<SizeT>(height) + compressedInfo.blockHeight - 1) /
+                                    compressedInfo.blockHeight;
+        const SizeT firstBlockX = static_cast<SizeT>(xoffset) / compressedInfo.blockWidth;
+        const SizeT firstBlockY = static_cast<SizeT>(yoffset) / compressedInfo.blockHeight;
+        const SizeT regionRowBytes = regionBlocksX * blockByteSize;
+        const auto* source = static_cast<const Uint8*>(compressedBytes);
+        for (SizeT row = 0; row < regionBlocksY; ++row) {
+            const SizeT destOffset = (firstBlockY + row) * levelRowBytes + firstBlockX * blockByteSize;
+            if (destOffset + regionRowBytes > blobSize) break; // a level whose blob predates its size
+            Memcpy(blob.data() + destOffset, source + row * regionRowBytes, regionRowBytes);
+        }
+        textureMipmapObject->SetMipmapCompressedImage(textureUploadTarget, static_cast<Uint>(level), format,
+                                                      blob.data(), blobSize);
     }
 
     void CompressedTexSubImage1D_State(GLenum target, GLint level, GLint xoffset, GLsizei width, GLenum format,
@@ -3533,28 +3907,8 @@ namespace MobileGL::MG_Impl::GLImpl {
         // SetMipmapCompressedImage re-arms it.
         textureMipmapObject->AllocateStorage(textureUploadTarget, level, {{width, height, 1}, internalBytes});
 
-        const void* compressedBytes = data;
+        if (!ValidateCompressedUnpackBufferSource(data, expectedImageSize, __func__)) return;
-        const auto& pixelUnpackBufferObject =
+        const void* compressedBytes = CompressedUnpackSource(data);
-            MG_State::pGLContext->GetBufferBindingSlot(BufferTarget::PixelUnpack).GetBoundObject();
-        if (pixelUnpackBufferObject) {
-            if (pixelUnpackBufferObject->IsMapped()) {
-                MG_State::pGLContext->RecordError(
-                    ErrorCode::InvalidOperation,
-                    MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", __func__,
-                                                 "Pixel unpack buffer is currently mapped."));
-                return;
-            }
-            const SizeT offset = reinterpret_cast<SizeT>(data);
-            const SizeT bufferSize = pixelUnpackBufferObject->GetSize();
-            if (offset > bufferSize || expectedImageSize > bufferSize - offset) {
-                MG_State::pGLContext->RecordError(
-                    ErrorCode::InvalidOperation,
-                    MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", __func__,
-                                                 "Unpacking would read past the end of the pixel unpack buffer."));
-                return;
-            }
-            compressedBytes = reinterpret_cast<const char*>(pixelUnpackBufferObject->MappedData()) + offset;
-        }
         textureMipmapObject->SetMipmapCompressedImage(textureUploadTarget, level, internalformat, compressedBytes,
                                                       expectedImageSize);
         textureMipmapObject->MarkStorageDirty(textureUploadTarget, level, true);
@@ -4064,6 +4418,13 @@ namespace MobileGL::MG_Impl::GLImpl {
         // core 8.19). Allocating only the primary one left the object cube-incomplete, so every
         // framebuffer it was attached to reported GL_FRAMEBUFFER_INCOMPLETE_ATTACHMENT. Every other
         // 2D target has exactly one upload target, so this loop is a no-op change for them.
+        // A specific compressed internalformat commits every level it allocates to that
+        // format, the same way glTexImage2D does - and here it matters twice over, because
+        // immutable storage plus glCompressedTexSubImage2D IS the modern way to upload a
+        // compressed texture: without the tag that sub-image call finds an uncompressed
+        // level and refuses it. Zero width means a generic (implementation's choice)
+        // format, which MobileGL answers with uncompressed storage, so it is not tagged.
+        const auto compressedInfo = MG_Util::GetCompressedFormatInfo(internalformat);
         for (const auto uploadTarget : textureObject->GetUploadTargets()) {
             for (GLsizei level = 0; level < levels; ++level) {
                 const GLsizei levelWidth = std::max<GLsizei>(1, width >> level);
@@ -4072,6 +4433,13 @@ namespace MobileGL::MG_Impl::GLImpl {
                     static_cast<SizeT>(levelWidth) * static_cast<SizeT>(levelHeight) * bytesPerPixel;
                 textureMipmapObject->AllocateStorage(uploadTarget, level, {{levelWidth, levelHeight, 1}, byteSize});
                 textureMipmapObject->MarkStorageDirty(uploadTarget, level, false);
+                if (compressedInfo.blockWidth != 0) {
+                    // After AllocateStorage, which clears the tag.
+                    textureMipmapObject->SetMipmapCompressedImage(
+                        uploadTarget, static_cast<Uint>(level), internalformat, nullptr,
+                        MG_Util::CalculateCompressedTextureImageSize(compressedInfo,
+                                                                     {levelWidth, levelHeight, 1}));
+                }
             }
             // See TextureStorage1D.
             textureMipmapObject->TruncateMipmapLevels(uploadTarget, static_cast<Uint>(levels));
@@ -4079,10 +4447,46 @@ 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;
@@ -4405,6 +4809,14 @@ namespace MobileGL::MG_Impl::GLImpl {
         free(processedPixels);
     }
 
+    void CompressedTextureSubImage2D(GLuint texture, GLint level, GLint xoffset, GLint yoffset, GLsizei width,
+                                     GLsizei height, GLenum format, GLsizei imageSize, const void* data) {
+        auto textureObject = GetTextureObjectByName(texture, __func__);
+        WithTemporarilyBoundNamedTexture(textureObject, [&](GLenum target) {
+            CompressedTexSubImage2D_State(target, level, xoffset, yoffset, width, height, format, imageSize, data);
+        });
+    }
+
     void TextureSubImage3D(GLuint texture, GLint level, GLint xoffset, GLint yoffset, GLint zoffset, GLsizei width,
                            GLsizei height, GLsizei depth, GLenum format, GLenum type, const void* pixels) {
         auto textureObject = GetTextureObjectByName(texture, __func__);
@@ -4482,6 +4894,132 @@ namespace MobileGL::MG_Impl::GLImpl {
         MG_State::pGLContext->NoteTextureUnitTouched(static_cast<Int>(unit), changed);
     }
 
+    GLint GetCombinedTextureImageUnitCount() {
+        GLint maxTextureUnits = 0;
+        GetIntegerv(GL_MAX_COMBINED_TEXTURE_IMAGE_UNITS, &maxTextureUnits);
+        return std::min<GLint>(std::max(maxTextureUnits, 0), MG_State::GLState::TextureState::MAX_TEXTURE_IMAGE_UNITS);
+    }
+
+    namespace {
+        // ARB_multi_bind checks the whole [first, first + count) range before binding anything and
+        // reports an overrun as INVALID_OPERATION - not the INVALID_VALUE the single-bind entry
+        // points report for an out-of-range unit, and not after binding the in-range prefix.
+        Bool ValidateMultiBindUnitRange(GLuint first, GLsizei count, GLint unitCount, const char* funcName) {
+            if (count < 0) {
+                MG_State::pGLContext->RecordError(
+                    ErrorCode::InvalidValue,
+                    MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", funcName, "count must be non-negative."));
+                return false;
+            }
+            if (static_cast<Uint64>(first) + static_cast<Uint64>(count) > static_cast<Uint64>(unitCount)) {
+                MG_State::pGLContext->RecordError(
+                    ErrorCode::InvalidOperation,
+                    MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", funcName,
+                                                 std::format("first + count ({} + {}) exceeds the {} available units.",
+                                                             first, count, unitCount)));
+                return false;
+            }
+            return true;
+        }
+
+        // ARB_multi_bind states the equivalence to a loop of single binds "except that <textures>
+        // will not be created if they do not exist": glBindTexture instantiates a name GenTextures
+        // merely reserved, the multi-bind entry points must refuse it. The error class is
+        // INVALID_OPERATION for both of them, where the scalar glBindImageTexture reports
+        // INVALID_VALUE - hence the check here rather than inside BindImageTexture.
+        //
+        // Deliberately PER ELEMENT: the extension defines these calls as a loop, so a bad entry
+        // costs its own unit and leaves the rest of the range bound.
+        SharedPtr<MG_State::GLState::ITextureObject> ResolveMultiBindTexture(GLuint texture, GLsizei index,
+                                                                            const char* funcName) {
+            SharedPtr<MG_State::GLState::ITextureObject> textureObject =
+                MG_State::pGLContext->GetTextureObject(texture);
+            if (!textureObject) {
+                MG_State::pGLContext->RecordError(
+                    ErrorCode::InvalidOperation,
+                    MakeUnique<GenericErrorInfo>(
+                        "MG_Impl/GLImpl", funcName,
+                        std::format("textures[{}] ({}) is not the name of an existing texture object.", index,
+                                    texture)));
+            }
+            return textureObject;
+        }
+
+        // ARB_multi_bind: an element naming texture zero unbinds EVERY target of its unit, i.e.
+        // rebinds each target's default texture object - the unit's initial state. Same rule
+        // glBindTextureUnit(unit, 0) follows.
+        void UnbindAllTargetsOnUnit(Int unit) {
+            auto& textureUnit = MG_State::pGLContext->GetTextureUnitObject(unit);
+            Bool changed = false;
+            for (auto& slot : textureUnit.GetAllBindingSlots()) {
+                if (slot.Bind(MG_State::pGLContext->GetDefaultTextureObject(slot.GetTarget()))) changed = true;
+            }
+            MG_State::pGLContext->NoteTextureUnitTouched(unit, changed);
+        }
+    } // namespace
+
+    // ARB_multi_bind: glBindTextures binds each texture to ITS OWN target on unit <first> + i, so
+    // there is no target parameter and no way to express it through glBindTexture - the per-unit,
+    // by-object form glBindTextureUnit uses is the one that matches. A NULL <textures> unbinds the
+    // whole range.
+    void BindTextures(GLuint first, GLsizei count, const GLuint* textures) {
+        if (!ValidateMultiBindUnitRange(first, count, GetCombinedTextureImageUnitCount(), __func__)) return;
+
+        for (GLsizei i = 0; i < count; ++i) {
+            const GLuint texture = textures ? textures[i] : 0;
+            const Int unit = static_cast<Int>(first) + i;
+            if (texture == 0) {
+                UnbindAllTargetsOnUnit(unit);
+                continue;
+            }
+            const SharedPtr<MG_State::GLState::ITextureObject> textureObject =
+                ResolveMultiBindTexture(texture, i, __func__);
+            if (!textureObject) continue;
+
+            auto& textureUnit = MG_State::pGLContext->GetTextureUnitObject(unit);
+            const Bool changed = textureUnit.GetBindingSlot(textureObject->GetTarget()).Bind(textureObject);
+            MG_State::pGLContext->NoteTextureUnitTouched(unit, changed);
+        }
+    }
+
+    // ARB_multi_bind: glBindImageTextures is a loop of glBindImageTexture with every parameter but
+    // the unit and the texture fixed by the spec - level 0, layered, layer 0, READ_WRITE, and the
+    // texture's own internal format. An element that names texture zero resets the unit.
+    void BindImageTextures(GLuint first, GLsizei count, const GLuint* textures) {
+        if (!ValidateMultiBindUnitRange(first, count, static_cast<GLint>(GetAdvertisedImageUnitCount()), __func__)) {
+            return;
+        }
+
+        for (GLsizei i = 0; i < count; ++i) {
+            const GLuint texture = textures ? textures[i] : 0;
+            const GLuint unit = first + static_cast<GLuint>(i);
+            if (texture == 0) {
+                BindImageTexture(unit, 0, 0, GL_FALSE, 0, GL_READ_ONLY, GL_R8);
+                continue;
+            }
+            const SharedPtr<MG_State::GLState::ITextureObject> textureObject =
+                ResolveMultiBindTexture(texture, i, __func__);
+            if (!textureObject) continue;
+
+            // "An INVALID_OPERATION error is generated if the internal format of any texture is not
+            // supported for image textures" - a texture that has never been given storage has no
+            // format at all and lands here too, rather than being reported as a bad enum by the
+            // scalar path.
+            const GLenum format = MG_Util::ConvertTextureInternalFormatToGLEnum(textureObject->GetFormat());
+            if (!IsValidImageTextureFormat(format)) {
+                MG_State::pGLContext->RecordError(
+                    ErrorCode::InvalidOperation,
+                    MakeUnique<GenericErrorInfo>(
+                        "MG_Impl/GLImpl", __func__,
+                        std::format("textures[{}] ({}) has an internal format that is not supported for image "
+                                    "textures.",
+                                    i, texture)));
+                continue;
+            }
+            BindImageTexture(unit, texture, 0, GL_TRUE, 0, GL_READ_WRITE, format);
+        }
+    }
+
     void GetTextureImage(GLuint texture, GLint level, GLenum format, GLenum type, GLsizei bufSize, void* pixels) {
         auto textureObject = GetTextureObjectByName(texture, __func__);
         if (!textureObject) return;
@@ -5176,10 +5714,15 @@ namespace MobileGL::MG_Impl::GLImpl {
     void CopyImageSubData(GLuint srcName, GLenum srcTarget, GLint srcLevel, GLint srcX, GLint srcY, GLint srcZ,
                           GLuint dstName, GLenum dstTarget, GLint dstLevel, GLint dstX, GLint dstY, GLint dstZ,
                           GLsizei srcWidth, GLsizei srcHeight, GLsizei srcDepth) {
-        auto srcTexture = GetTextureObjectByName(srcName, __func__);
+        // A missing name is INVALID_VALUE here, where GetTextureObjectByName's own diagnostic is
-        auto dstTexture = GetTextureObjectByName(dstName, __func__);
+        // INVALID_OPERATION - so resolve through the plain lookup, which answers a null
-        if (!ValidateCopyImageSubData_State(srcTexture, srcTarget, srcLevel, dstTexture, dstTarget, dstLevel,
+        // SharedPtr, and let the validator record the error this entry point owes.
-                                            srcWidth, srcHeight, srcDepth)) {
+        const SharedPtr<MG_State::GLState::ITextureObject> srcTexture =
+            MG_State::pGLContext->GetTextureObject(srcName);
+        const SharedPtr<MG_State::GLState::ITextureObject> dstTexture =
+            MG_State::pGLContext->GetTextureObject(dstName);
+        if (!ValidateCopyImageSubData_State(srcTexture, srcTarget, srcLevel, srcX, srcY, dstTexture, dstTarget,
+                                            dstLevel, dstX, dstY, srcWidth, srcHeight, srcDepth)) {
             return;
         }
         CopyImageSubData_Backend(srcTexture, srcTarget, srcLevel, srcX, srcY, srcZ, dstTexture, dstTarget, dstLevel,

MobileGL/MG_Impl/GLImpl/Texture/GL_Texture.h

@@ -37,6 +37,8 @@ namespace MobileGL::MG_Impl::GLImpl {
                            GLenum format, GLenum type, const void* pixels);
     void TextureSubImage3D(GLuint texture, GLint level, GLint xoffset, GLint yoffset, GLint zoffset, GLsizei width,
                            GLsizei height, GLsizei depth, GLenum format, GLenum type, const void* pixels);
+    void CompressedTextureSubImage2D(GLuint texture, GLint level, GLint xoffset, GLint yoffset, GLsizei width,
+                                     GLsizei height, GLenum format, GLsizei imageSize, const void* data);
     void TextureParameterf(GLuint texture, GLenum pname, GLfloat param);
     void TextureParameterfv(GLuint texture, GLenum pname, const GLfloat* params);
     void TextureParameteri(GLuint texture, GLenum pname, GLint param);
@@ -132,5 +134,11 @@ namespace MobileGL::MG_Impl::GLImpl {
     void CompressedTexImage1D(GLenum target, GLint level, GLenum internalformat, GLsizei width, GLint border,
                               GLsizei imageSize, const void* data);
     void BindTexture(GLenum target, GLuint texture);
+    void BindTextures(GLuint first, GLsizei count, const GLuint* textures);
+    void BindImageTextures(GLuint first, GLsizei count, const GLuint* textures);
     void ActiveTexture(GLenum texture);
+    // The number of texture image units a texture or a sampler may be bound to: what the backend
+    // advertises as GL_MAX_COMBINED_TEXTURE_IMAGE_UNITS, clamped by the frontend's fixed unit-array
+    // capacity. Shared so the texture and sampler multi-bind range checks cannot drift apart.
+    GLint GetCombinedTextureImageUnitCount();
 } // namespace MobileGL::MG_Impl::GLImpl

MobileGL/MG_Impl/GLImpl/Texture/Validators.cpp

@@ -15,6 +15,7 @@
 #include <MG_Util/Converters/MGToGL/TextureEnumConverter.h>
 #include <MG_Util/Converters/MGToMG/TextureEnumConverter.h>
 #include <MG_Util/Converters/MGToStr/TextureEnumConverter.h>
+#include <MG_Util/Metrics/TextureMetrics.h>
 
 namespace MobileGL::MG_Impl::GLImpl::TextureImpl {
     Bool ValidateTextureTarget(TextureTarget target) {
@@ -353,6 +354,63 @@ namespace MobileGL::MG_Impl::GLImpl::TextureImpl {
         return true;
     }
 
+    Bool ValidateTextureLevelExists(const SharedPtr<MG_State::GLState::ITextureObject>& textureObject, Int level,
+                                    const char* caller) {
+        // A null object is somebody else's error to report - ValidateTextureObject runs
+        // first at every call site and has already recorded it.
+        if (!textureObject) return false;
+
+        const auto* mipmapTexture = MG_State::GLState::AsMipmapTexture(textureObject.get());
+        if (mipmapTexture == nullptr) {
+            // The only non-mipmap storage class is a buffer texture, and GL_TEXTURE_BUFFER is
+            // not a target glCopyImageSubData accepts at all (it is in the CTS's invalid-target
+            // set). Declining here is not the error code the spec asks for - that would be
+            // INVALID_ENUM from a target check this validator is not - but it does keep a
+            // texture with no image levels whatsoever from reaching a backend that would
+            // dereference a backend texture it never created.
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidOperation,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", caller,
+                                             "Texture has no mipmap levels to address."));
+            return false;
+        }
+
+        // What this number is, exactly, because two other things are almost it and neither is
+        // safe to assume: it is the number of level SLOTS the shadow has allocated - holes
+        // included, since MipmapStorage::AllocateLevel grows to level+1 and never fills the gap.
+        // For a cube map MipmapUploadTargetArray reports face +X's chain rather than the union.
+        //
+        // The guarantee that matters is one-sided: this count is always >= the level count the
+        // backends derive (VkTextureManager::GetUploadMipLevelCount stops at the first level
+        // with a non-positive extent, so it can only be shorter). That is the safe direction -
+        // no copy to a level the texture genuinely has is ever rejected here. It is NOT an
+        // exact match, so the backends keep their own range guard for the band in between: a
+        // chain with a hole (level 0 and 2 defined, 1 not) is accepted by this predicate and
+        // declined by the backend, which is a silent no-op rather than a copy. That band is a
+        // backend storage limitation, not a validation one - rejecting it here with
+        // INVALID_VALUE would be refusing a copy the spec permits.
+        const Uint levelCount = mipmapTexture->GetMipmapLevelCount();
+
+        if (levelCount == 0) {
+            // No image has ever been defined on this texture, so the fault is the texture,
+            // not the number: GL 4.6 core 18.3.2 asks for INVALID_OPERATION when an object a
+            // copy names is an incomplete texture.
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidOperation,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", caller,
+                                             "Texture has no image defined at any level."));
+            return false;
+        }
+        if (level < 0 || static_cast<Uint>(level) >= levelCount) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidValue,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", caller,
+                                             "Texture level does not exist in this texture."));
+            return false;
+        }
+        return true;
+    }
+
     Bool ValidateTextureObject(const SharedPtr<MG_State::GLState::ITextureObject>& textureObject) {
         if (!textureObject) {
             MG_State::pGLContext->RecordError(
@@ -424,19 +482,141 @@ namespace MobileGL::MG_Impl::GLImpl::TextureImpl {
         return true;
     }
 
-    Bool ValidateBaseInternalFormatMatch(TextureInternalFormat format1, TextureInternalFormat format2) {
+    namespace {
-        auto unsizedFormat1 = MG_Util::ConvertInternalFormatToUnsized(format1);
+        // Component set of an UNSIZED base internal format, as the bitmask GL 4.6 SS 8.6
-        auto unsizedFormat2 = MG_Util::ConvertInternalFormatToUnsized(format2);
+        // reasons about. Colour components are independent bits so "subset" is a plain
-        if (unsizedFormat1 != unsizedFormat2) {
+        // 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
+
+    CopyImageTexelBlock ResolveCopyImageTexelBlock(TextureInternalFormat format, GLenum compressedFormat) {
+        CopyImageTexelBlock block{};
+        if (compressedFormat != GL_NONE) {
+            const auto info = MG_Util::GetCompressedFormatInfo(compressedFormat);
+            if (info.blockByteSize != 0) {
+                block.byteSize = info.blockByteSize;
+                block.blockWidth = info.blockWidth;
+                block.blockHeight = info.blockHeight;
+                block.compressed = true;
+                return block;
+            }
+        }
+        // The size MobileGL actually stores a texel of this format in, which for every format GL
+        // gives a required size is that required size. The handful of legacy formats GL leaves
+        // implementation-defined (R3_G3_B2, RGB4/5/10/12, RGBA2/12) have no view class in table
+        // 8.22 to be compared against anyway, and this is the size that decides whether a raw
+        // copy between them would in fact preserve the bytes.
+        block.byteSize = MG_Util::GetSizedInternalFormatSizeInBytes(format);
+        return block;
+    }
+
+    Bool ValidateCopyImageFormatCompatibility(const CopyImageTexelBlock& srcBlock,
+                                              const CopyImageTexelBlock& dstBlock) {
+        if (srcBlock.byteSize == 0 || dstBlock.byteSize == 0) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidOperation,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", "ValidateCopyImageFormatCompatibility",
+                                             "A copied image has no storage whose texel size is known."));
+            return false;
+        }
+        if (srcBlock.byteSize != dstBlock.byteSize) {
             MG_State::pGLContext->RecordError(
                 ErrorCode::InvalidOperation,
                 MakeUnique<GenericErrorInfo>(
-                    std::format("MG_Impl/GLImpl", "ValidateBaseInternalFormatMatch",
+                    "MG_Impl/GLImpl", "ValidateCopyImageFormatCompatibility",
-                                "The base internal format of the two formats do not match ({} vs. {})",
+                    std::format("The two images' texel blocks are different sizes ({} vs. {} bytes), so the "
-                                MG_Util::ConvertTextureInternalFormatToString(unsizedFormat1).c_str(),
+                                "formats are not copy-compatible.",
-                                MG_Util::ConvertTextureInternalFormatToString(unsizedFormat2).c_str())));
+                                srcBlock.byteSize, dstBlock.byteSize)));
+            return false;
+        }
+        // Two compressed images additionally have to agree on the SHAPE of the block, not only
+        // its size: an 8-byte 4x4 block and a hypothetical 8-byte 8x8 one hold different texel
+        // counts, and GL 4.6 core 18.3.2 requires both dimensions to match.
+        if (srcBlock.compressed && dstBlock.compressed &&
+            (srcBlock.blockWidth != dstBlock.blockWidth || srcBlock.blockHeight != dstBlock.blockHeight)) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidOperation,
+                MakeUnique<GenericErrorInfo>(
+                    "MG_Impl/GLImpl", "ValidateCopyImageFormatCompatibility",
+                    std::format("The two compressed images have different block dimensions ({}x{} vs. {}x{}).",
+                                srcBlock.blockWidth, srcBlock.blockHeight, dstBlock.blockWidth,
+                                dstBlock.blockHeight)));
             return false;
         }
         return true;
-    } // namespace TextureImpl
+    }
+
+    Bool ValidateCopyImageBlockAlignment(const CopyImageTexelBlock& block, Int x, Int y, Int width, Int height,
+                                         Int imageWidth, Int imageHeight, const char* endpointName) {
+        if (!block.compressed) return true;
+        const Int blockWidth = static_cast<Int>(block.blockWidth);
+        const Int blockHeight = static_cast<Int>(block.blockHeight);
+        if (blockWidth <= 1 && blockHeight <= 1) return true;
+        // The origin is unconditional; the extent gets the "or it reaches the edge of the image"
+        // exemption GL 4.6 core 18.3.2 grants, which is what lets a 16x16 BPTC image be copied
+        // whole even when the last block is partial.
+        const Bool originAligned = (x % blockWidth == 0) && (y % blockHeight == 0);
+        const Bool widthOk = (width % blockWidth == 0) || (x + width == imageWidth);
+        const Bool heightOk = (height % blockHeight == 0) || (y + height == imageHeight);
+        if (originAligned && widthOk && heightOk) return true;
+        MG_State::pGLContext->RecordError(
+            ErrorCode::InvalidValue,
+            MakeUnique<GenericErrorInfo>(
+                "MG_Impl/GLImpl", "ValidateCopyImageBlockAlignment",
+                std::format("The {} region [{}, {}] + [{} x {}] is not aligned to the {}x{} compressed block "
+                            "grid of a {} x {} image.",
+                            endpointName, x, y, width, height, blockWidth, blockHeight, imageWidth, imageHeight)));
+        return false;
+    }
+
+    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

@@ -30,6 +30,16 @@ namespace MobileGL::MG_Impl::GLImpl::TextureImpl {
                                                           TextureInternalFormat internalFormat,
                                                           TexturePixelDataType type);
     Bool ValidateTextureLevelWithUploadTarget(TextureUploadTarget target, Int level);
+    // "Is <level> a level this texture actually has?", which ValidateTextureLevelNumber above
+    // does NOT answer - that one only bounds the index by GL_MAX_TEXTURE_SIZE and knows nothing
+    // about the object. Entry points that resolve a level straight into a backend image
+    // subresource need this one: a level the texture never had is GL_INVALID_VALUE (GL 4.6 core
+    // 18.3.2), and passing it through instead reaches the driver as an out-of-range subresource.
+    // Note the error split is per-entry-point, so this is not universally reusable:
+    // glClearTexImage owes INVALID_OPERATION for the same out-of-range level and spells its own
+    // copy of this predicate in GL_Texture.cpp (GetClearTextureObject).
+    Bool ValidateTextureLevelExists(const SharedPtr<MG_State::GLState::ITextureObject>& textureObject, Int level,
+                                    const char* caller);
     Bool ValidateTextureObject(const SharedPtr<MG_State::GLState::ITextureObject>& textureObject);
     // Rejects the per-target default texture objects (name 0) with GL_INVALID_OPERATION for entry
     // points that require a GenTextures-created texture, e.g. TexStorage* ("An INVALID_OPERATION
@@ -40,5 +50,33 @@ 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);
-    Bool ValidateBaseInternalFormatMatch(TextureInternalFormat format1, TextureInternalFormat format2);
+    // The texel block of one glCopyImageSubData endpoint, resolved to the two things the
+    // compatibility rule actually asks about. `compressed` is not redundant with a block bigger
+    // than 1x1: it is what distinguishes "compressed, and so the region is measured in texels of
+    // a blocked image" from "uncompressed, and so it is measured in texels".
+    struct CopyImageTexelBlock {
+        SizeT byteSize = 0;
+        Uint blockWidth = 1;
+        Uint blockHeight = 1;
+        Bool compressed = false;
+    };
+    // `compressedFormat` is the GLenum a glCompressedTexImage* upload recorded for the level, or
+    // GL_NONE. It has to be asked for separately because MobileGL stores every compressed format
+    // in uncompressed storage (ConvertGLEnumToTextureInternalFormat), so the TextureInternalFormat
+    // alone can no longer tell a BPTC image from the RGBA8 backing it.
+    CopyImageTexelBlock ResolveCopyImageTexelBlock(TextureInternalFormat format, GLenum compressedFormat);
+    // GL 4.6 core 18.3.2: the two images must be COMPATIBLE, and compatible means their texel
+    // blocks are the same SIZE - not that they share a base internal format. RGBA32UI into
+    // RGBA32F is legal (both 128-bit) while RGBA8 into RGBA32F is not, and a compressed image
+    // pairs with an uncompressed one whose texel is as big as the compressed block.
+    Bool ValidateCopyImageFormatCompatibility(const CopyImageTexelBlock& srcBlock,
+                                              const CopyImageTexelBlock& dstBlock);
+    // GL 4.6 core 18.3.2: for a compressed image the region's origin must sit on a block
+    // boundary and its size must be a whole number of blocks - unless the edge it runs to is
+    // the edge of the image.
+    Bool ValidateCopyImageBlockAlignment(const CopyImageTexelBlock& block, Int x, Int y, Int width, Int height,
+                                         Int imageWidth, Int imageHeight, const char* endpointName);
+    // 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:
@@ -293,9 +315,10 @@ namespace MobileGL::MG_Impl::GLImpl {
 
         auto offset = reinterpret_cast<SizeT>(pointer);
 
-        vao->SetAttributeFormat(index, size, dataType, false, stride, offset, true, false);
+        const int effectiveStride = EffectiveVertexStride(stride, size, type);
+        vao->SetAttributeFormat(index, size, dataType, false, stride, offset, true, false, effectiveStride);
         vao->BindAttributeBuffer(index, vbo);
-        vao->MirrorPointerIntoBinding(index, vbo, offset, EffectiveVertexStride(stride, size, type));
+        vao->MirrorPointerIntoBinding(index, vbo, offset, effectiveStride);
     }
 
     void VertexAttribPointer_State(GLuint index, GLint size, GLenum type, GLboolean normalized, GLsizei stride,
@@ -323,9 +346,11 @@ namespace MobileGL::MG_Impl::GLImpl {
         // backend can pick the reversed VkFormat / pass GL_BGRA through to a GLES driver.
         const bool isBgra = (size == static_cast<GLint>(GL_BGRA));
         const int effectiveSize = isBgra ? 4 : size;
-        vao->SetAttributeFormat(index, effectiveSize, dataType, normalized, stride, offset, false, isBgra);
+        const int effectiveStride = EffectiveVertexStride(stride, effectiveSize, type);
+        vao->SetAttributeFormat(index, effectiveSize, dataType, normalized, stride, offset, false, isBgra,
+                                effectiveStride);
         vao->BindAttributeBuffer(index, vbo);
-        vao->MirrorPointerIntoBinding(index, vbo, offset, EffectiveVertexStride(stride, effectiveSize, type));
+        vao->MirrorPointerIntoBinding(index, vbo, offset, effectiveStride);
     }
 
     void BindVertexArray_State(GLuint array) {
@@ -502,7 +527,7 @@ namespace MobileGL::MG_Impl::GLImpl {
 
         if (!MG_Backend::pActiveBackendObject ||
             !MG_Backend::pActiveBackendObject->GetDynamicParameters().SupportsFloat64VertexAttributes) {
-            MGLOG_I("VertexAttribLFormat: attribute %u asked for a 64-bit (GL_DOUBLE) format, but this "
+            MGLOG_W_ONCE("VertexAttribLFormat: attribute %u asked for a 64-bit (GL_DOUBLE) format, but this "
                     "backend has no double-precision vertex attribute support - see the "
                     "\"64-bit vertex attributes\" / \"shaderFloat64\" POST row for what that costs",
                     attribindex);
@@ -944,7 +969,7 @@ 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 +1039,7 @@ 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 +1104,11 @@ 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 +1166,7 @@ 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 +1250,7 @@ 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 +1322,14 @@ 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 +1343,21 @@ 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 +1369,7 @@ 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 +1377,7 @@ 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_Impl/GLXImpl/Exporting/Definitions.cpp

@@ -166,32 +166,32 @@ MOBILEGL_GLX_API int glXSwapIntervalSGI(int interval) {
 
 // Legacy entry points some loaders probe for; harmless no-op stubs.
 MOBILEGL_GLX_API void glXCopyContext(Display*, void*, void*, unsigned long) {
-    MGLOG_W("glx: glXCopyContext is not supported");
+    MGLOG_W_ONCE("glx: glXCopyContext is not supported");
 }
 
 MOBILEGL_GLX_API unsigned long glXCreateGLXPixmap(Display*, void*, unsigned long) {
-    MGLOG_W("glx: glXCreateGLXPixmap is not supported");
+    MGLOG_W_ONCE("glx: glXCreateGLXPixmap is not supported");
     return 0;
 }
 
 MOBILEGL_GLX_API void glXDestroyGLXPixmap(Display*, unsigned long) {}
 
 MOBILEGL_GLX_API unsigned long glXCreatePixmap(Display*, void*, unsigned long, const int*) {
-    MGLOG_W("glx: glXCreatePixmap is not supported");
+    MGLOG_W_ONCE("glx: glXCreatePixmap is not supported");
     return 0;
 }
 
 MOBILEGL_GLX_API void glXDestroyPixmap(Display*, unsigned long) {}
 
 MOBILEGL_GLX_API unsigned long glXCreatePbuffer(Display*, void*, const int*) {
-    MGLOG_W("glx: glXCreatePbuffer is not supported");
+    MGLOG_W_ONCE("glx: glXCreatePbuffer is not supported");
     return 0;
 }
 
 MOBILEGL_GLX_API void glXDestroyPbuffer(Display*, unsigned long) {}
 
 MOBILEGL_GLX_API void glXUseXFont(unsigned long, int, int, int) {
-    MGLOG_W("glx: glXUseXFont is not supported");
+    MGLOG_W_ONCE("glx: glXUseXFont is not supported");
 }
 
 MOBILEGL_GLX_API void glXSelectEvent(Display*, unsigned long, unsigned long) {}

MobileGL/MG_Impl/GLXImpl/GLXImpl.cpp

@@ -149,7 +149,7 @@ namespace MobileGL::MG_Impl::GLXImpl {
                     fns->Sync = reinterpret_cast<decltype(fns->Sync)>(dlsym(fns->Library, "XSync"));
                 }
                 if (!fns->Valid()) {
-                    MGLOG_E("glx: failed to load libX11 entry points");
+                    MGLOG_E_ONCE("glx: failed to load libX11 entry points");
                 }
                 return fns;
             }();
@@ -314,7 +314,7 @@ namespace MobileGL::MG_Impl::GLXImpl {
             Uint32 width = 0;
             Uint32 height = 0;
             if (!QueryDrawableSize(dpy, drawable, width, height)) {
-                MGLOG_E("glx: XGetGeometry failed for drawable 0x%lx", drawable);
+                MGLOG_E_ONCE("glx: XGetGeometry failed for drawable 0x%lx", drawable);
                 return nullptr;
             }
 
@@ -326,7 +326,7 @@ namespace MobileGL::MG_Impl::GLXImpl {
             EGLSurface surface = EGLImpl::CreatePlatformWindowSurface(
                 context.Display, context.Config, reinterpret_cast<void*>(drawable), attribs);
             if (surface == EGL_NO_SURFACE) {
-                MGLOG_E("glx: failed to create window surface for drawable 0x%lx (%ux%u)", drawable,
+                MGLOG_E_ONCE("glx: failed to create window surface for drawable 0x%lx (%ux%u)", drawable,
                         width, height);
                 return nullptr;
             }
@@ -347,7 +347,7 @@ namespace MobileGL::MG_Impl::GLXImpl {
             const std::lock_guard<std::recursive_mutex> lock(RegistryMutex());
             EGLDisplay display = EnsureDisplay();
             if (display == EGL_NO_DISPLAY) {
-                MGLOG_E("glx: no EGL display");
+                MGLOG_E_ONCE("glx: no EGL display");
                 return nullptr;
             }
             EGLImpl::BindAPI(EGL_OPENGL_API);
@@ -376,13 +376,13 @@ namespace MobileGL::MG_Impl::GLXImpl {
             EGLint configCount = 0;
             if (!EGLImpl::ChooseConfig(display, configAttribs, &config, 1, &configCount) ||
                 configCount <= 0) {
-                MGLOG_E("glx: eglChooseConfig failed");
+                MGLOG_E_ONCE("glx: eglChooseConfig failed");
                 return nullptr;
             }
 
             EGLContext eglContext = EGLImpl::CreateContext(display, config, shareContext, contextAttribs);
             if (eglContext == EGL_NO_CONTEXT) {
-                MGLOG_E("glx: eglCreateContext failed");
+                MGLOG_E_ONCE("glx: eglCreateContext failed");
                 return nullptr;
             }
 
@@ -931,7 +931,7 @@ namespace MobileGL::MG_Impl::GLXImpl {
 
         if (!EGLImpl::MakeCurrent(object->Display, surface->Surface, surface->Surface,
                                   object->Context)) {
-            MGLOG_E("glx: eglMakeCurrent failed (drawable=0x%lx, ctx=%p)", drawable, context);
+            MGLOG_E_ONCE("glx: eglMakeCurrent failed (drawable=0x%lx, ctx=%p)", drawable, context);
             return 0;
         }
         t_current = {dpy, drawable, drawable, context};
@@ -943,7 +943,7 @@ namespace MobileGL::MG_Impl::GLXImpl {
         if (context && draw != read) {
             // MobileGL's backends reject split draw/read surfaces; bind the draw
             // drawable for both, which is what every real caller here needs.
-            MGLOG_W("glx: glXMakeContextCurrent draw 0x%lx != read 0x%lx, using draw for both", draw,
+            MGLOG_W_ONCE("glx: glXMakeContextCurrent draw 0x%lx != read 0x%lx, using draw for both", draw,
                     read);
         }
         const int result = MakeCurrent(dpy, draw, context);
@@ -958,7 +958,7 @@ namespace MobileGL::MG_Impl::GLXImpl {
         auto& surfaces = DrawableSurfaces();
         auto it = surfaces.find(drawable);
         if (it == surfaces.end()) {
-            MGLOG_W("glx: glXSwapBuffers with no surface for drawable 0x%lx", drawable);
+            MGLOG_W_ONCE("glx: glXSwapBuffers with no surface for drawable 0x%lx", drawable);
             return;
         }
         SyncSurfaceSize(dpy, drawable, it->second);

MobileGL/MG_Impl/GLXImpl/LookUp/LookUp.cpp

@@ -31,7 +31,7 @@ namespace MG_Impl::GLXImpl {
 #endif
         void* proc = MobileGL::MG_Impl::GetProcAddress(name);
         if (!proc) {
-            MGLOG_W("Failed to get function: %s", (const char*)name);
+            MGLOG_D("Failed to get function: %s", (const char*)name);
             return nullptr;
         }
 

MobileGL/MG_Impl/GetProcAddress.cpp

@@ -1403,7 +1403,7 @@ namespace MobileGL::MG_Impl {
         GETPROC(glFramebufferTextureMultiviewOVR, name);
         // GETPROC(glNamedFramebufferTextureMultiviewOVR, name);
 
-        MGLOG_W("GetProcAddress(%s) = nullptr!", name);
+        MGLOG_D("GetProcAddress(%s) = nullptr!", name);
         return nullptr;
     }
 } // namespace MobileGL::MG_Impl

MobileGL/MG_Impl/NSOpenGLImpl/NSOpenGLImpl.cpp

@@ -269,7 +269,7 @@ namespace MobileGL::MG_Impl::NSOpenGLImpl {
             }
             id metalLayerClass = reinterpret_cast<id>(objc_getClass("CAMetalLayer"));
             if (!metalLayerClass) {
-                MGLOG_E("NSOpenGLImpl: CAMetalLayer class not found");
+                MGLOG_E_ONCE("NSOpenGLImpl: CAMetalLayer class not found");
                 return nil;
             }
 
@@ -310,7 +310,7 @@ namespace MobileGL::MG_Impl::NSOpenGLImpl {
                                                        static_cast<GLint>(geometry.DrawableSize.width),
                                                        static_cast<GLint>(geometry.DrawableSize.height));
             if (error != kCGLNoError) {
-                MGLOG_E("NSOpenGLImpl: failed to attach drawable: %s", CGLImpl::ErrorString(error));
+                MGLOG_E_ONCE("NSOpenGLImpl: failed to attach drawable: %s", CGLImpl::ErrorString(error));
             }
         }
 
@@ -325,7 +325,7 @@ namespace MobileGL::MG_Impl::NSOpenGLImpl {
             }
             const auto error = CGLImpl::SetCurrentContext(context);
             if (error != kCGLNoError) {
-                MGLOG_E("NSOpenGLImpl: makeCurrentContext failed: %s", CGLImpl::ErrorString(error));
+                MGLOG_E_ONCE("NSOpenGLImpl: makeCurrentContext failed: %s", CGLImpl::ErrorString(error));
             }
         }
 
@@ -345,7 +345,7 @@ namespace MobileGL::MG_Impl::NSOpenGLImpl {
             }
             const auto error = CGLImpl::FlushDrawable(context);
             if (error != kCGLNoError) {
-                MGLOG_E("NSOpenGLImpl: flushBuffer failed: %s", CGLImpl::ErrorString(error));
+                MGLOG_E_ONCE("NSOpenGLImpl: flushBuffer failed: %s", CGLImpl::ErrorString(error));
             }
         }
 
@@ -377,7 +377,7 @@ namespace MobileGL::MG_Impl::NSOpenGLImpl {
                                                        static_cast<GLint>(geometry.DrawableSize.width),
                                                        static_cast<GLint>(geometry.DrawableSize.height));
             if (error != kCGLNoError) {
-                MGLOG_E("NSOpenGLImpl: update failed to attach drawable: %s", CGLImpl::ErrorString(error));
+                MGLOG_E_ONCE("NSOpenGLImpl: update failed to attach drawable: %s", CGLImpl::ErrorString(error));
                 return;
             }
             CGLImpl::UpdateContext(context);
@@ -421,7 +421,7 @@ namespace MobileGL::MG_Impl::NSOpenGLImpl {
             SEL selector = sel_registerName(selectorName);
             Method method = class_getInstanceMethod(cls, selector);
             if (!method) {
-                MGLOG_W("NSOpenGLImpl: missing instance method %s", selectorName);
+                MGLOG_W_ONCE("NSOpenGLImpl: missing instance method %s", selectorName);
                 return;
             }
             if (original) {
@@ -434,7 +434,7 @@ namespace MobileGL::MG_Impl::NSOpenGLImpl {
             SEL selector = sel_registerName(selectorName);
             Method method = class_getClassMethod(cls, selector);
             if (!method) {
-                MGLOG_W("NSOpenGLImpl: missing class method %s", selectorName);
+                MGLOG_W_ONCE("NSOpenGLImpl: missing class method %s", selectorName);
                 return;
             }
             method_setImplementation(method, replacement);
@@ -444,7 +444,7 @@ namespace MobileGL::MG_Impl::NSOpenGLImpl {
             Class pixelFormatClass = objc_getClass("NSOpenGLPixelFormat");
             Class contextClass = objc_getClass("NSOpenGLContext");
             if (!pixelFormatClass || !contextClass) {
-                MGLOG_W("NSOpenGLImpl: NSOpenGL classes are not loaded; hooks not installed");
+                MGLOG_W_ONCE("NSOpenGLImpl: NSOpenGL classes are not loaded; hooks not installed");
                 return false;
             }
 

MobileGL/MG_Impl/WGLImpl/Exporting/Definitions.cpp

@@ -56,7 +56,7 @@ extern "C" HGLRC WINAPI wglCreateLayerContext(HDC hdc, int iLayerPlane) {
 }
 
 extern "C" BOOL WINAPI wglCopyContext(HGLRC, HGLRC, UINT) {
-    MGLOG_W("wglCopyContext is not supported");
+    MGLOG_W_ONCE("wglCopyContext is not supported");
     SetLastError(ERROR_NOT_SUPPORTED);
     return FALSE;
 }
@@ -132,24 +132,24 @@ extern "C" DWORD WINAPI wglSwapMultipleBuffers(UINT n, CONST WGLSWAP* ps) {
 // ---- Font rendering (legacy immediate-mode feature; not supported) ----
 
 extern "C" BOOL WINAPI wglUseFontBitmapsA(HDC, DWORD, DWORD, DWORD) {
-    MGLOG_W("wglUseFontBitmapsA is not supported");
+    MGLOG_W_ONCE("wglUseFontBitmapsA is not supported");
     return FALSE;
 }
 
 extern "C" BOOL WINAPI wglUseFontBitmapsW(HDC, DWORD, DWORD, DWORD) {
-    MGLOG_W("wglUseFontBitmapsW is not supported");
+    MGLOG_W_ONCE("wglUseFontBitmapsW is not supported");
     return FALSE;
 }
 
 extern "C" BOOL WINAPI wglUseFontOutlinesA(HDC, DWORD, DWORD, DWORD, FLOAT, FLOAT, int,
                                                  LPGLYPHMETRICSFLOAT) {
-    MGLOG_W("wglUseFontOutlinesA is not supported");
+    MGLOG_W_ONCE("wglUseFontOutlinesA is not supported");
     return FALSE;
 }
 
 extern "C" BOOL WINAPI wglUseFontOutlinesW(HDC, DWORD, DWORD, DWORD, FLOAT, FLOAT, int,
                                                  LPGLYPHMETRICSFLOAT) {
-    MGLOG_W("wglUseFontOutlinesW is not supported");
+    MGLOG_W_ONCE("wglUseFontOutlinesW is not supported");
     return FALSE;
 }
 

MobileGL/MG_Impl/WGLImpl/WGLImpl.cpp

@@ -215,7 +215,7 @@ namespace MobileGL::MG_Impl::WGLImpl {
             Uint32 width = 0;
             Uint32 height = 0;
             if (!QueryClientSize(hwnd, width, height)) {
-                MGLOG_E("wgl: GetClientRect failed for HWND %p", hwnd);
+                MGLOG_E_ONCE("wgl: GetClientRect failed for HWND %p", hwnd);
                 return nullptr;
             }
 
@@ -227,7 +227,7 @@ namespace MobileGL::MG_Impl::WGLImpl {
             EGLSurface surface =
                 EGLImpl::CreatePlatformWindowSurface(context.Display, context.Config, hwnd, attribs);
             if (surface == EGL_NO_SURFACE) {
-                MGLOG_E("wgl: failed to create window surface for HWND %p (%ux%u)", hwnd, width, height);
+                MGLOG_E_ONCE("wgl: failed to create window surface for HWND %p (%ux%u)", hwnd, width, height);
                 return nullptr;
             }
 
@@ -244,7 +244,7 @@ namespace MobileGL::MG_Impl::WGLImpl {
             const std::lock_guard<std::recursive_mutex> lock(RegistryMutex());
             EGLDisplay display = EnsureDisplay();
             if (display == EGL_NO_DISPLAY) {
-                MGLOG_E("wgl: no EGL display");
+                MGLOG_E_ONCE("wgl: no EGL display");
                 return nullptr;
             }
             EGLImpl::BindAPI(EGL_OPENGL_API);
@@ -275,13 +275,13 @@ namespace MobileGL::MG_Impl::WGLImpl {
             EGLConfig config = nullptr;
             EGLint configCount = 0;
             if (!EGLImpl::ChooseConfig(display, configAttribs, &config, 1, &configCount) || configCount <= 0) {
-                MGLOG_E("wgl: eglChooseConfig failed");
+                MGLOG_E_ONCE("wgl: eglChooseConfig failed");
                 return nullptr;
             }
 
             EGLContext eglContext = EGLImpl::CreateContext(display, config, shareContext, contextAttribs);
             if (eglContext == EGL_NO_CONTEXT) {
-                MGLOG_E("wgl: eglCreateContext failed");
+                MGLOG_E_ONCE("wgl: eglCreateContext failed");
                 return nullptr;
             }
 
@@ -612,7 +612,7 @@ namespace MobileGL::MG_Impl::WGLImpl {
         auto& surfaces = WindowSurfaces();
         auto it = surfaces.find(hwnd);
         if (it == surfaces.end()) {
-            MGLOG_W("wglSwapBuffers: no surface for HWND %p", hwnd);
+            MGLOG_W_ONCE("wglSwapBuffers: no surface for HWND %p", hwnd);
             return FALSE;
         }
         SyncSurfaceSize(hwnd, it->second);
@@ -685,7 +685,7 @@ namespace MobileGL::MG_Impl::WGLImpl {
         }
 
         if (!EGLImpl::MakeCurrent(object->Display, surface->Surface, surface->Surface, object->Context)) {
-            MGLOG_E("wglMakeCurrent: eglMakeCurrent failed (hdc=%p, hglrc=%p)", hdc, hglrc);
+            MGLOG_E_ONCE("wglMakeCurrent: eglMakeCurrent failed (hdc=%p, hglrc=%p)", hdc, hglrc);
             return FALSE;
         }
         t_current = {hdc, hglrc};

MobileGL/MG_IntegrationTest/CMakeLists.txt

@@ -24,9 +24,14 @@ set(CMAKE_CXX_STANDARD_REQUIRED ON)
 
 set(MGL_ITEST_ROOT ${CMAKE_CURRENT_LIST_DIR}/../..)
 
-# Only meaningful where MobileGL_s exists (i.e. not Android).
+# Desktop links the static implementation directly. Android runs the same
-if (NOT TARGET MobileGL_s)
+# executable from adb shell and links the shipping shared library instead.
-    message(STATUS "MobileGL_s is not available; skipping the integration test module")
+if (ANDROID)
+    set(MGL_ITEST_MOBILEGL_TARGET MobileGL)
+elseif (TARGET MobileGL_s)
+    set(MGL_ITEST_MOBILEGL_TARGET MobileGL_s)
+else()
+    message(STATUS "No MobileGL library target is available; skipping the integration test module")
     return()
 endif()
 
@@ -50,9 +55,41 @@ add_executable(MobileGLIntegrationTest
     Scenarios/CrossFrameBufferScenario.cpp
     Scenarios/ResidentIndexScenario.cpp
     Scenarios/MultiDrawScenario.cpp
+    Scenarios/DrawParametersScenario.cpp
     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/DepthStencilReadbackMatrixScenario.cpp
+    Scenarios/DepthStencilReadbackAttachmentShapeScenario.cpp
+    Scenarios/ClipDistanceScenario.cpp
+    Scenarios/ViewportArrayScenario.cpp
+    Scenarios/SsboArrayLengthScenario.cpp
+    Scenarios/DoublePrecisionScenario.cpp
+    Scenarios/UniformInitializerScenario.cpp
+    Scenarios/SwizzleAccessRoutineScenario.cpp
+    Scenarios/IterationRPFirstReductionScenario.cpp
+    Scenarios/IterationRPProgram203Scenario.cpp
+    Scenarios/IterationRPScratchFixScenario.cpp
+    Scenarios/ProgramPipelineScenario.cpp
+    Scenarios/ImageLoadStoreSsoScenario.cpp
+    Scenarios/ImageTargetKindScenario.cpp
+    Scenarios/ImageFormatQualifierScenario.cpp
+    Scenarios/SsboDeclarationFormScenario.cpp
+    Scenarios/Glsl420DeclarationScenario.cpp
+    Scenarios/FragmentOutputArrayIndexScenario.cpp
+    Scenarios/BufferTextureScenario.cpp
+    Scenarios/VertexAttribBindingScenario.cpp
+    Scenarios/XfbCaptureBufferReuseScenario.cpp
+    Scenarios/VertexArrayEnableDisableScenario.cpp
+    Scenarios/CopyImageLevelRangeScenario.cpp
+    Scenarios/CopyImageLayeredScenario.cpp
+    Scenarios/LayeredAttachmentBarrierScenario.cpp
 )
 
 target_include_directories(MobileGLIntegrationTest PRIVATE
@@ -63,9 +100,20 @@ target_include_directories(MobileGLIntegrationTest PRIVATE
 # gtest, not gtest_main: Main.cpp installs the harness banner itself.
 target_link_libraries(MobileGLIntegrationTest PRIVATE
     GTest::gtest
-    MobileGL_s
+    ${MGL_ITEST_MOBILEGL_TARGET}
 )
 
+if (ANDROID)
+    find_library(MGL_ITEST_ANDROID_LIBRARY android REQUIRED)
+    find_library(MGL_ITEST_LOG_LIBRARY log REQUIRED)
+    find_library(MGL_ITEST_MEDIANDK_LIBRARY mediandk REQUIRED)
+    target_link_libraries(MobileGLIntegrationTest PRIVATE
+        ${MGL_ITEST_ANDROID_LIBRARY}
+        ${MGL_ITEST_LOG_LIBRARY}
+        ${MGL_ITEST_MEDIANDK_LIBRARY}
+    )
+endif()
+
 if (MSVC)
     # Same reason as MG_Test/Backend/DirectVulkan: the GLES headers declare gl*
     # as dllimport on Windows, so the in-library GL entry-point definitions only
@@ -74,6 +122,10 @@ if (MSVC)
 endif()
 target_compile_definitions(MobileGLIntegrationTest PRIVATE -DNOMINMAX)
 
+if (ANDROID)
+    return()
+endif()
+
 # --- ctest wiring --------------------------------------------------------
 # A bare libEGL on a glvnd box resolves to whatever vendor comes first, which is
 # usually Mesa/llvmpipe - a software rasteriser silently replacing the GPU under
@@ -169,25 +221,24 @@ 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
+# No EGL_PLATFORM knob here on purpose. The harness pins EGL_PLATFORM=surfaceless
-# platform is not X11 unless it is said out loud (run_driver_bench.sh sets the
+# itself before its first EGL call (HeadlessGL.cpp, EnsureHeadlessPlatform) so a
-# same variable). Wrong platform here is not a soft failure: eglCreatePbuffer
+# developer's machine and a CI runner take the SAME path whether or not a window
-# fails and every scenario skips.
+# system happens to be running. This used to inject "x11", which is how the lane
-if (UNIX AND NOT APPLE AND NOT ANDROID)
+# came up green on a workstation with WSLg and died on a runner with no X server.
-    set(MOBILEGL_ITEST_EGL_PLATFORM "x11" CACHE STRING
+#
-        "EGL_PLATFORM for the integration tests (empty: leave the loader alone)")
+# A build-system knob would not just be redundant, it would be a trap: `set(...
-else()
+# CACHE ...)` does not rewrite an existing cache, so every build directory
-    set(MOBILEGL_ITEST_EGL_PLATFORM "" CACHE STRING
+# configured before this change would keep injecting EGL_PLATFORM=x11 and go on
-        "EGL_PLATFORM for the integration tests (empty: leave the loader alone)")
+# binding to a window system - silently, and only on the machines that have one.
-endif()
+# 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=${MOBILEGL_ITEST_EGL_VENDOR}")
 endif()
-if (MOBILEGL_ITEST_EGL_PLATFORM)
+unset(MOBILEGL_ITEST_EGL_PLATFORM CACHE) # see above: an old cache must not resurrect x11
-    list(APPEND MGL_ITEST_COMMON_ENV "EGL_PLATFORM=${MOBILEGL_ITEST_EGL_PLATFORM}")
-endif()
 if (MOBILEGL_ITEST_REQUIRE_GPU)
     list(APPEND MGL_ITEST_COMMON_ENV "MOBILEGL_ITEST_REQUIRE_GPU=1")
 endif()
@@ -221,6 +272,8 @@ mgl_itest_join_environment(MGL_ITEST_VULKAN_ENVIRONMENT
     "MOBILEGL_BACKEND_TYPE=DirectVulkan" ${MGL_ITEST_VULKAN_ENV})
 mgl_itest_join_environment(MGL_ITEST_VULKAN_ASYNC_ENVIRONMENT
     "MOBILEGL_BACKEND_TYPE=DirectVulkan" "MOBILEGL_ASYNC_SHADER_COMPILE=1" ${MGL_ITEST_VULKAN_ENV})
+mgl_itest_join_environment(MGL_ITEST_GLES_FORCED_DS_ENVIRONMENT
+    "MOBILEGL_BACKEND_TYPE=DirectGLES" "MOBILEGL_ESPRYT_FORCE_DS_READBACK_EMULATION=1" ${MGL_ITEST_COMMON_ENV})
 
 # TIMEOUT on every entry: a GPU test that wedges must fail the run, not hang it.
 set(MGL_ITEST_TIMEOUT 120)
@@ -268,3 +321,21 @@ gtest_discover_tests(MobileGLIntegrationTest
         TIMEOUT ${MGL_ITEST_TIMEOUT}
         ENVIRONMENT "${MGL_ITEST_VULKAN_ASYNC_ENVIRONMENT}"
 )
+
+# A fourth registration, of the depth/stencil readback scenarios, with the ES
+# shader-sampling emulation forced on. Not paranoia - without it these scenarios are
+# UNFALSIFIABLE on the machines this suite runs on: OpenGL ES has no depth or stencil
+# readback in core, but Mesa accepts the reads anyway, so on llvmpipe every one of them
+# goes green through a native path that the Adreno device does not have. Deleting the
+# entire emulation left all of them passing. With the flag the native spellings are off
+# the table and only the path the device actually takes remains. DirectGLES only - the
+# emulation is DirectGLES's.
+gtest_discover_tests(MobileGLIntegrationTest
+    TEST_PREFIX "DirectGLES.ForcedDepthStencilEmulation."
+    TEST_FILTER "DepthStencilReadback*Scenario.*"
+    DISCOVERY_TIMEOUT 30
+    PROPERTIES
+        LABELS integration-gpu
+        TIMEOUT ${MGL_ITEST_TIMEOUT}
+        ENVIRONMENT "${MGL_ITEST_GLES_FORCED_DS_ENVIRONMENT}"
+)

MobileGL/MG_IntegrationTest/Harness/HeadlessGL.cpp

@@ -15,6 +15,16 @@
 #include <ostream>
 #include <sstream>
 
+#if defined(_WIN32)
+#define WIN32_LEAN_AND_MEAN
+#include <windows.h>
+#elif defined(__ANDROID__)
+#include <android/hardware_buffer.h>
+#include <android/native_window.h>
+#include <media/NdkImage.h>
+#include <media/NdkImageReader.h>
+#endif
+
 // MobileGL's own headers, in the order MobileGL/Includes.h uses them: GL/gl.h
 // first, then glcorearb.h for the 3.x+ entry points. This binary links
 // MobileGL_s, so every gl*/egl* below binds to MobileGL's implementation, not
@@ -32,7 +42,7 @@
 // the only construction that is actually predictive here: MobileGL ABORTS
 // (MOBILEGL_ASSERT -> SIGTRAP) rather than returning an error on an unusable
 // platform, so nothing the parent can call in-process is allowed to be wrong.
-#if !defined(_WIN32) && !defined(__APPLE__) && __has_include(<sys/wait.h>)
+#if !defined(_WIN32) && !defined(__APPLE__) && !defined(__ANDROID__) && __has_include(<sys/wait.h>)
 #define MGITEST_HAVE_FORK_PREFLIGHT 1
 #include <csignal>
 #include <ctime>
@@ -53,6 +63,83 @@ namespace MGITest {
         constexpr int kSurfaceWidth = 128;
         constexpr int kSurfaceHeight = 96;
 
+#if defined(_WIN32)
+        HWND g_testWindow = nullptr;
+
+        HWND CreateTestWindow() {
+            static const wchar_t* const kClassName = L"MobileGLIntegrationTestWindow";
+            static bool registered = false;
+            if (!registered) {
+                WNDCLASSW windowClass{};
+                windowClass.lpfnWndProc = DefWindowProcW;
+                windowClass.hInstance = GetModuleHandleW(nullptr);
+                windowClass.lpszClassName = kClassName;
+                if (RegisterClassW(&windowClass) == 0 && GetLastError() != ERROR_CLASS_ALREADY_EXISTS) {
+                    return nullptr;
+                }
+                registered = true;
+            }
+            return CreateWindowExW(0, kClassName, L"MobileGL Integration Test", WS_OVERLAPPEDWINDOW,
+                                   CW_USEDEFAULT, CW_USEDEFAULT, kSurfaceWidth, kSurfaceHeight, nullptr, nullptr,
+                                   GetModuleHandleW(nullptr), nullptr);
+        }
+#elif defined(__ANDROID__)
+        AImageReader* g_imageReader = nullptr;
+        ANativeWindow* g_imageReaderWindow = nullptr;
+
+        void DrainImageReader(void*, AImageReader* reader) {
+            AImage* image = nullptr;
+            if (AImageReader_acquireNextImage(reader, &image) == AMEDIA_OK && image != nullptr) {
+                AImage_delete(image);
+            }
+        }
+
+        bool CreateImageReaderWindow() {
+            if (g_imageReaderWindow != nullptr) return true;
+            constexpr int kMaxImages = 4;
+            const media_status_t status = AImageReader_newWithUsage(
+                kSurfaceWidth, kSurfaceHeight, AIMAGE_FORMAT_RGBA_8888,
+                AHARDWAREBUFFER_USAGE_GPU_SAMPLED_IMAGE | AHARDWAREBUFFER_USAGE_GPU_COLOR_OUTPUT,
+                kMaxImages, &g_imageReader);
+            if (status != AMEDIA_OK || g_imageReader == nullptr) return false;
+
+            AImageReader_ImageListener listener = {nullptr, DrainImageReader};
+            AImageReader_setImageListener(g_imageReader, &listener);
+            if (AImageReader_getWindow(g_imageReader, &g_imageReaderWindow) != AMEDIA_OK ||
+                g_imageReaderWindow == nullptr) {
+                AImageReader_setImageListener(g_imageReader, nullptr);
+                AImageReader_delete(g_imageReader);
+                g_imageReader = nullptr;
+                return false;
+            }
+            ANativeWindow_acquire(g_imageReaderWindow);
+            return true;
+        }
+
+        void DestroyImageReaderWindow() {
+            if (g_imageReaderWindow != nullptr) {
+                ANativeWindow_release(g_imageReaderWindow);
+                g_imageReaderWindow = nullptr;
+            }
+            if (g_imageReader != nullptr) {
+                AImageReader_setImageListener(g_imageReader, nullptr);
+                AImageReader_delete(g_imageReader);
+                g_imageReader = nullptr;
+            }
+        }
+#endif
+
+        bool UseWindowSurface() {
+#if defined(_WIN32)
+            const char* value = std::getenv("MOBILEGL_ITEST_WINDOW_SURFACE");
+            return value != nullptr && value[0] != '\0' && std::strcmp(value, "0") != 0;
+#elif defined(__ANDROID__)
+            return true;
+#else
+            return false;
+#endif
+        }
+
         std::string EnvOr(const char* name, const char* fallback) {
             const char* value = std::getenv(name);
             return (value != nullptr && value[0] != '\0') ? std::string(value) : std::string(fallback);
@@ -74,6 +161,40 @@ 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 on desktop. Android instead supplies an AImageReader
+        // ANativeWindow. DISPLAY/WAYLAND_DISPLAY are cleared as well so that a
+        // driver that consults them directly cannot reintroduce the dependency
+        // behind EGL's back.
+        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 +203,9 @@ 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");
@@ -97,8 +221,9 @@ namespace MGITest {
                 return 3;
             }
 
+            const bool useWindowSurface = UseWindowSurface();
             const EGLint configAttribs[] = {EGL_SURFACE_TYPE,
-                                            EGL_PBUFFER_BIT,
+                                            useWindowSurface ? EGL_WINDOW_BIT : EGL_PBUFFER_BIT,
                                             EGL_RED_SIZE,
                                             8,
                                             EGL_GREEN_SIZE,
@@ -115,7 +240,9 @@ namespace MGITest {
             EGLConfig config = nullptr;
             EGLint configCount = 0;
             if (eglChooseConfig(display, configAttribs, &config, 1, &configCount) != EGL_TRUE || configCount < 1) {
-                outReason = WithEglError("eglChooseConfig found no pbuffer-capable RGBA8/D24 config");
+                outReason = WithEglError(useWindowSurface
+                                             ? "eglChooseConfig found no window-capable RGBA8/D24 config"
+                                             : "eglChooseConfig found no pbuffer-capable RGBA8/D24 config");
                 return 4;
             }
 
@@ -129,10 +256,32 @@ namespace MGITest {
                 return 5;
             }
 
+            EGLSurface surface = EGL_NO_SURFACE;
+            if (useWindowSurface) {
+#if defined(_WIN32)
+                if (g_testWindow == nullptr) g_testWindow = CreateTestWindow();
+                if (g_testWindow == nullptr) {
+                    outReason = "failed to create the Windows integration-test window";
+                    return 6;
+                }
+                surface = eglCreateWindowSurface(display, config, g_testWindow, nullptr);
+#elif defined(__ANDROID__)
+                if (!CreateImageReaderWindow()) {
+                    outReason = "failed to create the Android AImageReader integration-test window";
+                    return 6;
+                }
+                surface = eglCreateWindowSurface(display, config, g_imageReaderWindow, nullptr);
+#endif
+            } else {
                 const EGLint pbufferAttribs[] = {EGL_WIDTH, kSurfaceWidth, EGL_HEIGHT, kSurfaceHeight, EGL_NONE};
-            EGLSurface surface = eglCreatePbufferSurface(display, config, pbufferAttribs);
+                surface = eglCreatePbufferSurface(display, config, pbufferAttribs);
+            }
             if (surface == EGL_NO_SURFACE) {
-                outReason = WithEglError("eglCreatePbufferSurface failed");
+#if defined(__ANDROID__)
+                DestroyImageReaderWindow();
+#endif
+                outReason = WithEglError(useWindowSurface ? "eglCreateWindowSurface failed"
+                                                         : "eglCreatePbufferSurface failed");
                 return 6;
             }
             // The step that brings the whole backend up (DirectVulkan creates its
@@ -199,11 +348,10 @@ namespace MGITest {
             }
             if (child == 0) {
                 close(channel[0]);
-                // The child is EXPECTED to die on a signal on an unusable
+                // No core suppression here, deliberately: when the child dies on a
-                // platform; that is the measurement. Do not let each such
+                // signal, the core IS the diagnosis (an rlimit that used to sit here
-                // measurement drop a core file next to the test binary.
+                // made a CI-only crash undebuggable). Machines that do not want
-                const rlimit noCore{0, 0};
+                // cores control that with the usual ulimit/core_pattern knobs.
-                setrlimit(RLIMIT_CORE, &noCore);
                 std::fprintf(stderr, "[itest] pre-flight child: attempting a full EGL bring-up\n");
                 EglBringUp local;
                 std::string reason;
@@ -284,10 +432,20 @@ namespace MGITest {
         }
     } // namespace
 
-    bool RequireGpu() {
+    namespace {
-        const char* value = std::getenv("MOBILEGL_ITEST_REQUIRE_GPU");
+        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() {
+        return EnvFlag("MOBILEGL_ITEST_REQUIRE_GPU");
+    }
+
+    bool RequireHardwareGpu() {
+        return EnvFlag("MOBILEGL_ITEST_REQUIRE_HARDWARE_GPU");
+    }
 
     std::ostream& operator<<(std::ostream& os, const Rgba8& c) {
         os << "rgba(" << int(c.r) << "," << int(c.g) << "," << int(c.b) << "," << int(c.a) << ")";
@@ -390,6 +548,10 @@ 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();
     }
@@ -441,6 +603,14 @@ namespace MGITest {
         if (m_context != nullptr) eglDestroyContext(display, static_cast<EGLContext>(m_context));
         if (m_surface != nullptr) eglDestroySurface(display, static_cast<EGLSurface>(m_surface));
         eglTerminate(display);
+#if defined(_WIN32)
+        if (g_testWindow != nullptr) {
+            DestroyWindow(g_testWindow);
+            g_testWindow = nullptr;
+        }
+#elif defined(__ANDROID__)
+        DestroyImageReaderWindow();
+#endif
         m_context = nullptr;
         m_surface = nullptr;
         m_display = nullptr;
@@ -551,9 +721,13 @@ 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

@@ -14,11 +14,11 @@
 // inspects backend state - both bugs this module pins were invisible to
 // state-level assertions and visible only in pixels.
 //
-// Headless by construction, following MG_Benchmark/Driver/DriverBench.c: an EGL
+// Headless by construction: desktop uses an EGL pbuffer and Android uses an
-// context on a PBUFFER surface. No window, no window manager, no human. Unlike
+// AImageReader-backed ANativeWindow that needs no Activity. No window manager,
-// DriverBench the scenarios do draw to the DEFAULT framebuffer (that is where
+// no human. Unlike DriverBench the scenarios do draw to the DEFAULT framebuffer
-// the Y-flip lives) and do call eglSwapBuffers (that is the frame boundary the
+// (that is where the Y-flip lives) and do call eglSwapBuffers (that is the frame
-// cross-frame scenarios need to be real).
+// boundary the cross-frame scenarios need to be real).
 //
 // One process is one backend: MOBILEGL_BACKEND_TYPE is latched at
 // initialization, so the CMake wiring runs this binary once per backend rather
@@ -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
+    // Reads back the whole currently bound READ framebuffer.
-    // be the target's full size - DirectVulkan's default-framebuffer readback
-    // only re-orients a full-extent read.
     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())) {
+            if (RequireHardwareGpu() && LooksLikeSoftwareRasterizer(gl.RendererString())) {
-                // "Ran on llvmpipe" must not be able to pass as "ran on the GPU":
+                // Only when hardware was asked for BY NAME. REQUIRE_GPU means "an
-                // a misconfigured vendor pin silently lands on the software
+                // unusable harness is a failure, not a silent skip" - it is the
-                // rasterizer, and REQUIRE_GPU exists precisely to make that loud.
+                // falsifiability switch, and CI is exactly where it belongs. But CI
-                FAIL() << "MOBILEGL_ITEST_REQUIRE_GPU is set but the context landed on a software rasterizer: "
+                // runners have no GPU, so folding "must not be llvmpipe" into the
-                       << gl.RendererString();
+                // 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 +26,20 @@ 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",
+                // EGL_PLATFORM is echoed because it is the invariant this harness
-                             gl.RendererString().c_str(), gl.Width(), gl.Height());
+                // 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");
+#if defined(__ANDROID__)
+                constexpr const char* surfaceKind = "AImageReader window";
+#else
+                constexpr const char* surfaceKind = "pbuffer";
+#endif
+                std::fprintf(stderr, "  renderer: %s\n  surface:  %dx%d %s (headless, EGL_PLATFORM=%s)\n",
+                             gl.RendererString().c_str(), gl.Width(), gl.Height(),
+                             surfaceKind,
+                             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 +1,259 @@
+// 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";
+            }
+        }
+
+        // A per-stage block count is an amount of BINDING POINTS an application will use, so it
+        // can never exceed the number of binding points that exist. GL 4.6 Table 23.64 states the
+        // relation the other way round (MAX_UNIFORM_BUFFER_BINDINGS >= MAX_COMBINED_UNIFORM_BLOCKS
+        // >= every per-stage count), and DirectVulkan broke it by clamping the two families
+        // independently: a device reporting 256 compute uniform blocks and 84 uniform binding
+        // points passes both ceilings and still cannot serve
+        // KHR-GL44.multi_bind.dispatch_bind_buffers_base, which reads the block count and binds
+        // that many buffers in one glBindBuffersBase - INVALID_OPERATION before a single bind.
+        TEST_F(AdvertisedLimitsScenario, PerStageBlockCountsFitInTheirBindingPoints) {
+            struct Relation {
+                GLenum blocks;
+                const char* blocksName;
+                GLenum bindings;
+                const char* bindingsName;
+            };
+            const Relation relations[] = {
+                {GL_MAX_COMPUTE_UNIFORM_BLOCKS, "GL_MAX_COMPUTE_UNIFORM_BLOCKS", GL_MAX_UNIFORM_BUFFER_BINDINGS,
+                 "GL_MAX_UNIFORM_BUFFER_BINDINGS"},
+                {GL_MAX_VERTEX_UNIFORM_BLOCKS, "GL_MAX_VERTEX_UNIFORM_BLOCKS", GL_MAX_UNIFORM_BUFFER_BINDINGS,
+                 "GL_MAX_UNIFORM_BUFFER_BINDINGS"},
+                {GL_MAX_FRAGMENT_UNIFORM_BLOCKS, "GL_MAX_FRAGMENT_UNIFORM_BLOCKS", GL_MAX_UNIFORM_BUFFER_BINDINGS,
+                 "GL_MAX_UNIFORM_BUFFER_BINDINGS"},
+                {GL_MAX_COMBINED_UNIFORM_BLOCKS, "GL_MAX_COMBINED_UNIFORM_BLOCKS", GL_MAX_UNIFORM_BUFFER_BINDINGS,
+                 "GL_MAX_UNIFORM_BUFFER_BINDINGS"},
+                {GL_MAX_COMPUTE_SHADER_STORAGE_BLOCKS, "GL_MAX_COMPUTE_SHADER_STORAGE_BLOCKS",
+                 GL_MAX_SHADER_STORAGE_BUFFER_BINDINGS, "GL_MAX_SHADER_STORAGE_BUFFER_BINDINGS"},
+                {GL_MAX_COMBINED_SHADER_STORAGE_BLOCKS, "GL_MAX_COMBINED_SHADER_STORAGE_BLOCKS",
+                 GL_MAX_SHADER_STORAGE_BUFFER_BINDINGS, "GL_MAX_SHADER_STORAGE_BUFFER_BINDINGS"},
+            };
+            for (const Relation& relation : relations) {
+                GLint blocks = -1;
+                GLint bindings = -1;
+                glGetIntegerv(relation.blocks, &blocks);
+                glGetIntegerv(relation.bindings, &bindings);
+                ASSERT_EQ(FirstGLError(), GLenum(GL_NO_ERROR)) << relation.blocksName;
+                EXPECT_LE(blocks, bindings)
+                    << relation.blocksName << " = " << blocks << " exceeds " << relation.bindingsName << " = "
+                    << bindings << "; a shader may declare more blocks than there are binding points to bind them to";
+            }
+        }
+
+        // KHR-GL44.multi_bind.functional_bind_buffers_range sizes each of an indexed target's
+        // binding points at MAX_<target>_SIZE / MAX_<target>_BINDINGS and binds all of them in
+        // one glBindBuffersRange. That quotient has to be a legal BindBufferRange size, which
+        // makes the two limits of every indexed family a PAIR: advertise a size that does not
+        // survive division by the binding count and the call fails with INVALID_VALUE before any
+        // of it binds.
+        TEST_F(AdvertisedLimitsScenario, IndexedTargetSizeSurvivesDivisionByItsBindingCount) {
+            struct IndexedFamily {
+                GLenum maxSize;
+                const char* maxSizeName;
+                GLenum maxBindings;
+                const char* maxBindingsName;
+                GLint sizeGranularity; // BindBufferRange's size rule for the target
+            };
+            const IndexedFamily families[] = {
+                {GL_MAX_ATOMIC_COUNTER_BUFFER_SIZE, "GL_MAX_ATOMIC_COUNTER_BUFFER_SIZE",
+                 GL_MAX_ATOMIC_COUNTER_BUFFER_BINDINGS, "GL_MAX_ATOMIC_COUNTER_BUFFER_BINDINGS", 1},
+                {GL_MAX_TRANSFORM_FEEDBACK_INTERLEAVED_COMPONENTS, "GL_MAX_TRANSFORM_FEEDBACK_INTERLEAVED_COMPONENTS",
+                 GL_MAX_TRANSFORM_FEEDBACK_BUFFERS, "GL_MAX_TRANSFORM_FEEDBACK_BUFFERS", 4},
+                {GL_MAX_UNIFORM_BLOCK_SIZE, "GL_MAX_UNIFORM_BLOCK_SIZE", GL_MAX_UNIFORM_BUFFER_BINDINGS,
+                 "GL_MAX_UNIFORM_BUFFER_BINDINGS", 1},
+                {GL_MAX_SHADER_STORAGE_BLOCK_SIZE, "GL_MAX_SHADER_STORAGE_BLOCK_SIZE",
+                 GL_MAX_SHADER_STORAGE_BUFFER_BINDINGS, "GL_MAX_SHADER_STORAGE_BUFFER_BINDINGS", 1},
+            };
+            for (const IndexedFamily& family : families) {
+                GLint maxSize = -1;
+                GLint maxBindings = -1;
+                glGetIntegerv(family.maxSize, &maxSize);
+                glGetIntegerv(family.maxBindings, &maxBindings);
+                ASSERT_EQ(FirstGLError(), GLenum(GL_NO_ERROR)) << family.maxSizeName;
+                ASSERT_GT(maxBindings, 0) << family.maxBindingsName;
+                const GLint perBinding = maxSize / maxBindings;
+                EXPECT_GT(perBinding, 0)
+                    << family.maxSizeName << " (" << maxSize << ") / " << family.maxBindingsName << " ("
+                    << maxBindings << ") is zero, and BindBufferRange rejects a zero size";
+                EXPECT_EQ(perBinding % family.sizeGranularity, 0)
+                    << family.maxSizeName << " (" << maxSize << ") / " << family.maxBindingsName << " ("
+                    << maxBindings << ") = " << perBinding << " is not a multiple of the "
+                    << family.sizeGranularity << "-byte size granularity BindBufferRange requires for it";
+            }
+        }
+
+        // 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";
+        }
+
+        // ARB_viewport_array's own limits. They are advertised from three different places -
+        // GL_MAX_VIEWPORTS from the frontend's indexed state width, the bounds range and the
+        // subpixel bits from the backend caps table - and each backend fills that table from a
+        // different source, so all three are checked on both lanes.
+        //
+        // GL_VIEWPORT_BOUNDS_RANGE is the one that shipped wrong: GLES has no such query, the
+        // DirectGLES loader's glGetFloatv(GL_VIEWPORT_BOUNDS_RANGE) therefore raised
+        // GL_INVALID_ENUM and left the probe's zero-initialized array in place, and MobileGL
+        // advertised [0, 0] - a range that admits no viewport origin at all, and the check that
+        // kept KHR-GL43.viewport_array.queries red on Espryt after the indexed-state work.
+        TEST_F(AdvertisedLimitsScenario, ViewportArrayLimitsMeetTheirGL43Floors) {
+            GLint maxViewports = -1;
+            glGetIntegerv(GL_MAX_VIEWPORTS, &maxViewports);
+            ASSERT_EQ(FirstGLError(), GLenum(GL_NO_ERROR));
+            EXPECT_GE(maxViewports, 16) << "GL 4.3 core table 23.53 sets the MAX_VIEWPORTS minimum at 16";
+            EXPECT_LE(maxViewports, 256) << "one viewport rectangle of indexed state is allocated per advertised "
+                                            "viewport, and the CTS sizes its arrays off this number";
+
+            GLfloat boundsRange[2] = {1.0f, -1.0f};
+            glGetFloatv(GL_VIEWPORT_BOUNDS_RANGE, boundsRange);
+            ASSERT_EQ(FirstGLError(), GLenum(GL_NO_ERROR));
+            EXPECT_LE(boundsRange[0], -32768.0f)
+                << "GL 4.6 core table 23.60 sets the VIEWPORT_BOUNDS_RANGE minimum at [-32768, 32767]; got ["
+                << boundsRange[0] << ", " << boundsRange[1] << "]";
+            EXPECT_GE(boundsRange[1], 32767.0f)
+                << "GL 4.6 core table 23.60 sets the VIEWPORT_BOUNDS_RANGE minimum at [-32768, 32767]; got ["
+                << boundsRange[0] << ", " << boundsRange[1] << "]";
+
+            // KNOWN INFIDELITY, pinned here rather than hidden. MobileGL reports the driver's own
+            // VIEWPORT_SUBPIXEL_BITS (4 on llvmpipe, i.e. 1/16-pixel viewport precision), but the
+            // float viewport rectangle glViewportIndexedf stores is snapped to integers on its
+            // way to both backends (ComputeGLViewport, DirectGLES SyncRenderState). The STATE
+            // round trip is exact - which is all KHR-GL43.viewport_array.viewport_api checks, and
+            // all this cluster set out to fix - so the gap is in rasterization only: a fractional
+            // viewport origin rasterizes as if it had been rounded. Nothing in the suite or in
+            // Minecraft sets one. Only the spec floor is asserted; tightening this to EQ(0) would
+            // mean advertising no subpixel precision at all, which is a separate decision about a
+            // limit MobileGL currently passes through from the driver.
+            GLint subpixelBits = -1;
+            glGetIntegerv(GL_VIEWPORT_SUBPIXEL_BITS, &subpixelBits);
+            ASSERT_EQ(FirstGLError(), GLenum(GL_NO_ERROR));
+            EXPECT_GE(subpixelBits, 0) << "GL 4.6 core table 23.60: VIEWPORT_SUBPIXEL_BITS has a minimum of 0, and "
+                                          "a negative value is what a sign-flipped uint32 looks like";
+
+            GLint viewportDims[2] = {-1, -1};
+            glGetIntegerv(GL_MAX_VIEWPORT_DIMS, viewportDims);
+            ASSERT_EQ(FirstGLError(), GLenum(GL_NO_ERROR));
+            GLint maxRenderbufferSize = -1;
+            glGetIntegerv(GL_MAX_RENDERBUFFER_SIZE, &maxRenderbufferSize);
+            ASSERT_EQ(FirstGLError(), GLenum(GL_NO_ERROR));
+            // GL 4.6 core 13.6.1: MAX_VIEWPORT_DIMS must be at least as large as the largest
+            // renderable surface, or a full-size framebuffer could not be fully viewported.
+            EXPECT_GE(viewportDims[0], maxRenderbufferSize);
+            EXPECT_GE(viewportDims[1], maxRenderbufferSize);
+        }
+
+    } // 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 +1,302 @@
+// 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 <cstring>
+#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); }
+)";
+
+        // A buffer texture bound as a WRITABLE image: the shader reads one texel and writes
+        // another, so a single dispatch proves the read direction (which already worked) and
+        // the write direction (which is what this exists for) apart from each other.
+        constexpr const char* kImageBufferCS = R"(#version 430 core
+layout(local_size_x = 1) in;
+layout(binding = 0, rgba8) uniform imageBuffer uImage;
+void main() {
+    vec4 read = imageLoad(uImage, 1);
+    imageStore(uImage, 0, vec4(0.0, 1.0, 0.0, 1.0));
+    imageStore(uImage, 2, read);
+}
+)";
+
+        class BufferTextureScenario : public ScenarioTest {
+        protected:
+            bool ComputeImagesAreUsable() const {
+                GLint maxImageUnits = 0;
+                GLint maxComputeImageUniforms = 0;
+                glGetIntegerv(GL_MAX_IMAGE_UNITS, &maxImageUnits);
+                glGetIntegerv(GL_MAX_COMPUTE_IMAGE_UNIFORMS, &maxComputeImageUniforms);
+                while (glGetError() != GL_NO_ERROR) {
+                }
+                return maxImageUnits >= 1 && maxComputeImageUniforms >= 1;
+            }
+
+            unsigned int MakeComputeProgram(const char* source) {
+                const GLuint shader = glCreateShader(GL_COMPUTE_SHADER);
+                glShaderSource(shader, 1, &source, nullptr);
+                glCompileShader(shader);
+                GLint compiled = GL_FALSE;
+                glGetShaderiv(shader, GL_COMPILE_STATUS, &compiled);
+                if (compiled == GL_FALSE) {
+                    char log[4096] = {};
+                    glGetShaderInfoLog(shader, sizeof(log) - 1, nullptr, log);
+                    ADD_FAILURE() << "the compute shader did not compile: " << log;
+                    glDeleteShader(shader);
+                    return 0;
+                }
+                const GLuint program = glCreateProgram();
+                glAttachShader(program, shader);
+                glLinkProgram(program);
+                glDeleteShader(shader);
+                GLint linked = GL_FALSE;
+                glGetProgramiv(program, GL_LINK_STATUS, &linked);
+                if (linked == GL_FALSE) {
+                    char log[4096] = {};
+                    glGetProgramInfoLog(program, sizeof(log) - 1, nullptr, log);
+                    ADD_FAILURE() << "the compute program did not link: " << log;
+                    glDeleteProgram(program);
+                    return 0;
+                }
+                return program;
+            }
+        };
+
+        // 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);
+    }
+
+    // A shader may WRITE a buffer texture too, through an image unit, and the bytes it writes
+    // land in the backend's buffer - not in the frontend's CPU shadow, which is what MapBuffer
+    // and GetBufferSubData hand back. A storage-block write is flagged for exactly this reason
+    // and the shadow is refreshed on the next read; a buffer reached through an image unit is
+    // the same write through a different binding, and Espryt used to flag only the first, so
+    // an imageStore into a buffer texture was invisible to every CPU read that followed it -
+    // silently, with the correct value sitting in the driver's buffer the whole time.
+    //
+    // The read direction is asserted in the same dispatch (texel 2 is a copy of texel 1) so a
+    // failure here cannot be blamed on the image binding not working at all.
+    TEST_F(BufferTextureScenario, AnImageStoreIntoABufferTextureIsVisibleToTheCpu) {
+        if (!Ready()) return;
+        if (!ComputeImagesAreUsable()) GTEST_SKIP() << "no compute image units on this host";
+
+        constexpr GLuint kRed = 0x000000ffu;   // RGBA8 little-endian: r = 255
+        constexpr GLuint kGreen = 0xff00ff00u; // what the shader stores: (0, 1, 0, 1)
+        constexpr int kTexels = 16;
+
+        FirstGLError();
+
+        const unsigned int program = MakeComputeProgram(kImageBufferCS);
+        ASSERT_NE(program, 0u);
+
+        const std::vector<GLuint> texels(kTexels, kRed);
+        GLuint buffer = 0;
+        glGenBuffers(1, &buffer);
+        glBindBuffer(GL_TEXTURE_BUFFER, buffer);
+        glBufferData(GL_TEXTURE_BUFFER, static_cast<GLsizeiptr>(texels.size() * sizeof(GLuint)), texels.data(),
+                     GL_DYNAMIC_COPY);
+
+        GLuint texture = 0;
+        glGenTextures(1, &texture);
+        glBindTexture(GL_TEXTURE_BUFFER, texture);
+        glTexBuffer(GL_TEXTURE_BUFFER, GL_RGBA8, buffer);
+        EXPECT_EQ(FirstGLError(), 0u) << "glTexBuffer(GL_RGBA8) was refused";
+
+        glBindImageTexture(0, texture, 0, GL_FALSE, 0, GL_READ_WRITE, GL_RGBA8);
+        EXPECT_EQ(FirstGLError(), 0u) << "glBindImageTexture on a buffer texture was refused";
+
+        glUseProgram(program);
+        glDispatchCompute(1, 1, 1);
+        glMemoryBarrier(GL_ALL_BARRIER_BITS);
+
+        // Both CPU read paths, because they are two entry points onto the same refresh and a
+        // fix that reaches only one of them is not a fix. Everything below is EXPECT rather than
+        // ASSERT so that a failure still reaches the cleanup at the end: the harness shares one
+        // context across every scenario in the process, and a leaked buffer or image binding
+        // here would surface as a failure somewhere else entirely.
+        std::vector<GLuint> readBack(kTexels, 0u);
+        glBindBuffer(GL_TEXTURE_BUFFER, buffer);
+        glGetBufferSubData(GL_TEXTURE_BUFFER, 0, static_cast<GLsizeiptr>(readBack.size() * sizeof(GLuint)),
+                           readBack.data());
+        EXPECT_EQ(readBack[0], kGreen) << "glGetBufferSubData did not see the imageStore";
+        EXPECT_EQ(readBack[2], kRed) << "the imageLoad side of the same dispatch read the wrong texel";
+
+        const void* mapped = glMapBuffer(GL_TEXTURE_BUFFER, GL_READ_ONLY);
+        EXPECT_NE(mapped, nullptr) << "glMapBuffer(GL_READ_ONLY) on the texture's buffer failed";
+        if (mapped != nullptr) {
+            GLuint mappedTexel0 = 0;
+            std::memcpy(&mappedTexel0, mapped, sizeof(mappedTexel0));
+            EXPECT_EQ(mappedTexel0, kGreen) << "glMapBuffer did not see the imageStore";
+            glUnmapBuffer(GL_TEXTURE_BUFFER);
+        }
+
+        glBindImageTexture(0, 0, 0, GL_FALSE, 0, GL_READ_ONLY, GL_RGBA8);
+        glBindBuffer(GL_TEXTURE_BUFFER, 0);
+        glBindTexture(GL_TEXTURE_BUFFER, 0);
+        glUseProgram(0);
+        glDeleteProgram(program);
+        glDeleteTextures(1, &texture);
+        glDeleteBuffers(1, &buffer);
+        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/ClipDistanceScenario.cpp

@@ -0,0 +1,388 @@
+// MobileGL - MobileGL/MG_IntegrationTest/Scenarios/ClipDistanceScenario.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_ClipDistance ACTUALLY CLIPS, AND ONLY WHERE IT IS ENABLED.
+//
+// CapabilityInput::ClipDistance0..7 existed end to end - the GL enum converted to it, the
+// string converter named it, glEnable(GL_CLIP_DISTANCE0 + i) raised no error - and then
+// RenderState::SetCapability had no case for it and dropped it into `default: break`. Nothing
+// was stored, no version was bumped, and neither backend ever heard about it. The shader half
+// worked all along (SPIRV-Cross emits gl_ClipDistance with a
+// `#extension GL_EXT_clip_cull_distance : require` that Adreno accepts), so the distances were
+// computed and then ignored: no clipping ever happened on DirectGLES, which is the whole of
+// KHR-GLxx.clip_distance.functional. glIsEnabled lied about it too - it returned GL_FALSE
+// immediately after a successful glEnable.
+//
+// The assertions are behavioural, not query-shaped, because a query-only test passes against a
+// backend that stores the bit and never forwards it. Each case draws one full-viewport triangle
+// whose clip distance is positive on one side of the viewport and negative on the other, then
+// checks BOTH sides: the kept side proves the draw happened at all, and the clipped side is the
+// actual claim. The disabled case is the negative control - the identical shader with the
+// identical distances and the enable turned off must leave both sides painted, which is what
+// says the pixels below are being removed by clipping and not by something else.
+//
+// HONEST LIMIT OF THIS FILE IN CI. Of the four cases, only EnableIsObservableThroughIsEnabled is
+// falsifiable on the software rasterizers every automated lane runs on. llvmpipe and lavapipe
+// clip by EVERY declared gl_ClipDistance regardless of the enables, so
+// AnEnabledClipDistanceRemovesTheNegativeHalf goes green there against the broken tree as well,
+// and the two cases that need real per-distance semantics skip (see
+// DriverHonoursPerDistanceEnables). What actually pins the behaviour is Adreno, through
+// KHR-GLxx.clip_distance.functional - whose "without dynamic redeclaration" variants declare all
+// gl_MaxClipDistances slots and enable only the first N, i.e. exactly the subset semantics these
+// skipped cases assert. Read a green CI run here as "the state survives the frontend", not as
+// "clipping is correct"; the second claim is a device claim.
+//
+// Every case disables all eight distances on entry rather than assuming they start off:
+// XfbAfterClipDistanceScenario deliberately leaves one enabled for the rest of the process, and
+// forwarding the enables is what turned that leftover from inert bookkeeping into live driver
+// state.
+
+#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
+
+#ifndef GL_CLIP_DISTANCE0
+#define GL_CLIP_DISTANCE0 0x3000
+#endif
+#ifndef GL_CLIP_DISTANCE1
+#define GL_CLIP_DISTANCE1 0x3001
+#endif
+
+namespace MGITest {
+    namespace {
+
+        // One clip distance per half of the viewport: distance 0 is positive on the right half
+        // (x > 0 in clip space) and distance 1 is positive on the top half. A vertex shader
+        // producing a full-screen triangle from gl_VertexID, so no buffers are needed.
+        const char* const kVertexSource = R"(#version 400 core
+out float gl_ClipDistance[2];
+void main() {
+    vec2 positions[3] = vec2[3](vec2(-1.0, -1.0), vec2(3.0, -1.0), vec2(-1.0, 3.0));
+    vec2 p = positions[gl_VertexID];
+    gl_Position = vec4(p, 0.0, 1.0);
+    gl_ClipDistance[0] = p.x;
+    gl_ClipDistance[1] = p.y;
+}
+)";
+
+        const char* const kFragmentSource = R"(#version 400 core
+out vec4 fragColor;
+void main() { fragColor = vec4(0.0, 1.0, 0.0, 1.0); }
+)";
+
+        class ClipDistanceScenario : public ScenarioTest {
+        protected:
+            GLuint BuildProgram() {
+                const GLuint vs = glCreateShader(GL_VERTEX_SHADER);
+                glShaderSource(vs, 1, &kVertexSource, nullptr);
+                glCompileShader(vs);
+                GLint compiled = 0;
+                glGetShaderiv(vs, GL_COMPILE_STATUS, &compiled);
+                if (!compiled) {
+                    m_buildLog = ShaderLog(vs);
+                    glDeleteShader(vs);
+                    return 0;
+                }
+                const GLuint fs = glCreateShader(GL_FRAGMENT_SHADER);
+                glShaderSource(fs, 1, &kFragmentSource, nullptr);
+                glCompileShader(fs);
+                glGetShaderiv(fs, GL_COMPILE_STATUS, &compiled);
+                if (!compiled) {
+                    m_buildLog = ShaderLog(fs);
+                    glDeleteShader(vs);
+                    glDeleteShader(fs);
+                    return 0;
+                }
+                const GLuint program = glCreateProgram();
+                glAttachShader(program, vs);
+                glAttachShader(program, fs);
+                glLinkProgram(program);
+                GLint linked = 0;
+                glGetProgramiv(program, GL_LINK_STATUS, &linked);
+                glDeleteShader(vs);
+                glDeleteShader(fs);
+                if (!linked) {
+                    GLint length = 0;
+                    glGetProgramiv(program, GL_INFO_LOG_LENGTH, &length);
+                    std::vector<char> log(static_cast<size_t>(length > 1 ? length : 1), '\0');
+                    glGetProgramInfoLog(program, static_cast<GLsizei>(log.size()), nullptr, log.data());
+                    m_buildLog = log.data();
+                    glDeleteProgram(program);
+                    return 0;
+                }
+                return program;
+            }
+
+            const std::string& BuildLog() const { return m_buildLog; }
+
+            // Paints the whole viewport red, then draws the clipped triangle in green.
+            void DrawClippedTriangle(GLuint program, GLuint vao) const {
+                glClearColor(1.0f, 0.0f, 0.0f, 1.0f);
+                glClear(GL_COLOR_BUFFER_BIT);
+                glUseProgram(program);
+                glBindVertexArray(vao);
+                glDrawArrays(GL_TRIANGLES, 0, 3);
+            }
+
+            static bool IsGreen(const unsigned char* px) {
+                return px[0] < 64 && px[1] > 192;
+            }
+
+            static bool IsRed(const unsigned char* px) {
+                return px[0] > 192 && px[1] < 64;
+            }
+
+            void PixelAt(int x, int y, unsigned char* out) const {
+                glReadPixels(x, y, 1, 1, GL_RGBA, GL_UNSIGNED_BYTE, out);
+            }
+
+            // Never assume the eight start disabled - see the header note about
+            // XfbAfterClipDistanceScenario leaving one on for the rest of the process.
+            static void DisableEveryClipDistance() {
+                for (int i = 0; i < 8; ++i) {
+                    glDisable(static_cast<GLenum>(GL_CLIP_DISTANCE0 + i));
+                }
+            }
+
+            // True when the driver under this backend actually implements PER-DISTANCE enable
+            // state, i.e. when a written-but-disabled gl_ClipDistance leaves its fragments
+            // alone. Not every stack does, and the difference is not MobileGL's to hide:
+            //
+            //   - Adreno's ES driver honours GL_CLIP_DISTANCE0_EXT..7_EXT, which is what makes
+            //     KHR-GLxx.clip_distance.functional pass on the device once the enables are
+            //     forwarded at all.
+            //   - Vulkan has no such state: every clip distance a shader declares is active,
+            //     always. DirectVulkan therefore clips by a disabled distance.
+            //   - Mesa's llvmpipe ES driver behaves like Vulkan here.
+            //
+            // Emulating GL's semantics on those two would mean forcing the disabled slots to a
+            // non-negative value inside the shader, which makes the enable mask part of the
+            // pipeline key - a feature, not a fix, and deliberately not attempted here. The
+            // cases that need the real semantics gate on this probe and say so when they skip,
+            // rather than being deleted or silently weakened.
+            bool DriverHonoursPerDistanceEnables(GLuint program, GLuint vao) const {
+                for (int i = 0; i < 8; ++i) {
+                    glDisable(static_cast<GLenum>(GL_CLIP_DISTANCE0 + i));
+                }
+                DrawClippedTriangle(program, vao);
+                unsigned char negativeSide[4] = {0, 0, 0, 0};
+                glReadPixels(Gl().Width() / 4, Gl().Height() / 2, 1, 1, GL_RGBA, GL_UNSIGNED_BYTE, negativeSide);
+                return IsGreen(negativeSide);
+            }
+
+        private:
+            static std::string ShaderLog(GLuint shader) {
+                GLint length = 0;
+                glGetShaderiv(shader, GL_INFO_LOG_LENGTH, &length);
+                std::vector<char> log(static_cast<size_t>(length > 1 ? length : 1), '\0');
+                glGetShaderInfoLog(shader, static_cast<GLsizei>(log.size()), nullptr, log.data());
+                return log.data();
+            }
+
+            std::string m_buildLog;
+        };
+
+    } // namespace
+
+    // The state itself: glEnable must be observable through glIsEnabled. This is the cheap half
+    // of the bug - SetCapability's missing case made the query answer GL_FALSE for a capability
+    // that had just been enabled without error.
+    TEST_F(ClipDistanceScenario, EnableIsObservableThroughIsEnabled) {
+        if (!Ready()) return;
+        HeadlessGL& gl = Gl();
+
+        DisableEveryClipDistance();
+        EXPECT_EQ(glIsEnabled(GL_CLIP_DISTANCE0), GL_FALSE)
+            << "glDisable(GL_CLIP_DISTANCE0) is not observable through glIsEnabled";
+        glEnable(GL_CLIP_DISTANCE0);
+        EXPECT_EQ(FirstGLError(), 0u);
+        EXPECT_EQ(glIsEnabled(GL_CLIP_DISTANCE0), GL_TRUE)
+            << "glEnable(GL_CLIP_DISTANCE0) raised no error but glIsEnabled still reports it disabled";
+        EXPECT_EQ(glIsEnabled(GL_CLIP_DISTANCE1), GL_FALSE)
+            << "enabling distance 0 must not enable distance 1 - the eight are independent";
+
+        glEnable(GL_CLIP_DISTANCE1);
+        glDisable(GL_CLIP_DISTANCE0);
+        EXPECT_EQ(glIsEnabled(GL_CLIP_DISTANCE0), GL_FALSE);
+        EXPECT_EQ(glIsEnabled(GL_CLIP_DISTANCE1), GL_TRUE);
+
+        glDisable(GL_CLIP_DISTANCE1);
+        EXPECT_EQ(FirstGLError(), 0u);
+        gl.EndFrame();
+    }
+
+    // The claim: an enabled clip distance removes the fragments where it is negative.
+    TEST_F(ClipDistanceScenario, AnEnabledClipDistanceRemovesTheNegativeHalf) {
+        if (!Ready()) return;
+        HeadlessGL& gl = Gl();
+        const int width = gl.Width();
+        const int height = gl.Height();
+        ASSERT_GE(width, 8);
+        ASSERT_GE(height, 8);
+
+        GLuint vao = 0;
+        glGenVertexArrays(1, &vao);
+        const GLuint program = BuildProgram();
+        ASSERT_NE(program, 0u) << "the gl_ClipDistance program did not build: " << BuildLog();
+
+        BindDefaultFramebuffer();
+        glViewport(0, 0, width, height);
+        glDisable(GL_SCISSOR_TEST);
+        glDisable(GL_DEPTH_TEST);
+        glDisable(GL_CULL_FACE);
+        glColorMask(GL_TRUE, GL_TRUE, GL_TRUE, GL_TRUE);
+
+        // Distance 1 is positive by a single pixel at the sampled row, so a stray enable on it
+        // would put the "kept" probe right on the clip boundary.
+        DisableEveryClipDistance();
+        glEnable(GL_CLIP_DISTANCE0);
+        DrawClippedTriangle(program, vao);
+        EXPECT_EQ(FirstGLError(), 0u);
+
+        unsigned char right[4] = {0, 0, 0, 0};
+        unsigned char left[4] = {0, 0, 0, 0};
+        PixelAt(width - 1 - width / 4, height / 2, right);
+        PixelAt(width / 4, height / 2, left);
+        EXPECT_EQ(FirstGLError(), 0u);
+
+        EXPECT_TRUE(IsGreen(right)) << "the kept half is not painted (" << int(right[0]) << "," << int(right[1])
+                                    << "," << int(right[2]) << ") - the draw itself did not happen, so the clipped "
+                                       "half below proves nothing";
+        EXPECT_TRUE(IsRed(left)) << "gl_ClipDistance[0] is negative on the left half and GL_CLIP_DISTANCE0 is "
+                                    "enabled, so those fragments must be clipped away; found ("
+                                 << int(left[0]) << "," << int(left[1]) << "," << int(left[2]) << ")";
+
+        glDisable(GL_CLIP_DISTANCE0);
+        glUseProgram(0);
+        glBindVertexArray(0);
+        glDeleteProgram(program);
+        glDeleteVertexArrays(1, &vao);
+        gl.EndFrame();
+    }
+
+    // The negative control: the same shader writing the same distances, with the enable off,
+    // must paint both halves. Without this a backend that clipped everything - or one whose
+    // draw simply failed - would pass the case above.
+    TEST_F(ClipDistanceScenario, ADisabledClipDistanceRemovesNothing) {
+        if (!Ready()) return;
+        HeadlessGL& gl = Gl();
+        const int width = gl.Width();
+        const int height = gl.Height();
+        ASSERT_GE(width, 8);
+        ASSERT_GE(height, 8);
+
+        GLuint vao = 0;
+        glGenVertexArrays(1, &vao);
+        const GLuint program = BuildProgram();
+        ASSERT_NE(program, 0u) << "the gl_ClipDistance program did not build: " << BuildLog();
+
+        BindDefaultFramebuffer();
+        glViewport(0, 0, width, height);
+        glDisable(GL_SCISSOR_TEST);
+        glDisable(GL_DEPTH_TEST);
+        glDisable(GL_CULL_FACE);
+        DisableEveryClipDistance();
+
+        DrawClippedTriangle(program, vao);
+        EXPECT_EQ(FirstGLError(), 0u);
+
+        unsigned char right[4] = {0, 0, 0, 0};
+        unsigned char left[4] = {0, 0, 0, 0};
+        PixelAt(width - 1 - width / 4, height / 2, right);
+        PixelAt(width / 4, height / 2, left);
+        EXPECT_EQ(FirstGLError(), 0u);
+
+        EXPECT_TRUE(IsGreen(right)) << "with every clip distance disabled the whole triangle must survive";
+        const bool driverHonoursEnables = IsGreen(left);
+
+        glUseProgram(0);
+        glBindVertexArray(0);
+        glDeleteProgram(program);
+        glDeleteVertexArrays(1, &vao);
+        gl.EndFrame();
+        if (!driverHonoursEnables) {
+            GTEST_SKIP() << "renderer " << gl.RendererString()
+                         << " clips by a DISABLED gl_ClipDistance - it does not implement per-distance enable state "
+                            "(see DriverHonoursPerDistanceEnables). Emulating GL's semantics there needs shader-side "
+                            "masking keyed on the enable mask, which is a separate feature";
+        }
+    }
+
+    // The eight enables are independent: enabling only distance 1 must clip by distance 1 and
+    // leave distance 0 alone. A backend that forwarded "any clip distance enabled" as a single
+    // bit, or that always enables every declared distance (which is what Vulkan does natively),
+    // passes both cases above and fails this one.
+    TEST_F(ClipDistanceScenario, TheEnablesAreIndependentPerDistance) {
+        if (!Ready()) return;
+        HeadlessGL& gl = Gl();
+        const int width = gl.Width();
+        const int height = gl.Height();
+        ASSERT_GE(width, 8);
+        ASSERT_GE(height, 8);
+
+        GLuint vao = 0;
+        glGenVertexArrays(1, &vao);
+        const GLuint program = BuildProgram();
+        ASSERT_NE(program, 0u) << "the gl_ClipDistance program did not build: " << BuildLog();
+
+        BindDefaultFramebuffer();
+        glViewport(0, 0, width, height);
+        glDisable(GL_SCISSOR_TEST);
+        glDisable(GL_DEPTH_TEST);
+        glDisable(GL_CULL_FACE);
+        if (!DriverHonoursPerDistanceEnables(program, vao)) {
+            glUseProgram(0);
+            glBindVertexArray(0);
+            glDeleteProgram(program);
+            glDeleteVertexArrays(1, &vao);
+            DisableEveryClipDistance();
+            gl.EndFrame();
+            GTEST_SKIP() << "renderer " << gl.RendererString()
+                         << " clips by every declared gl_ClipDistance regardless of the enables, so per-distance "
+                            "independence is not observable here";
+        }
+
+        DisableEveryClipDistance();
+        glEnable(GL_CLIP_DISTANCE1);
+
+        DrawClippedTriangle(program, vao);
+        EXPECT_EQ(FirstGLError(), 0u);
+
+        // Distance 1 is negative on the bottom half, distance 0 on the left half. With only
+        // distance 1 enabled, the bottom-left must survive (distance 0 is off) and the bottom
+        // must not.
+        unsigned char topLeft[4] = {0, 0, 0, 0};
+        unsigned char bottomRight[4] = {0, 0, 0, 0};
+        PixelAt(width / 4, height - 1 - height / 4, topLeft);
+        PixelAt(width - 1 - width / 4, height / 4, bottomRight);
+        EXPECT_EQ(FirstGLError(), 0u);
+
+        EXPECT_TRUE(IsGreen(topLeft)) << "gl_ClipDistance[0] is negative here but GL_CLIP_DISTANCE0 is disabled, so "
+                                         "this fragment must survive";
+        EXPECT_TRUE(IsRed(bottomRight)) << "gl_ClipDistance[1] is negative here and GL_CLIP_DISTANCE1 is enabled, so "
+                                           "this fragment must be clipped";
+
+        glDisable(GL_CLIP_DISTANCE1);
+        glUseProgram(0);
+        glBindVertexArray(0);
+        glDeleteProgram(program);
+        glDeleteVertexArrays(1, &vao);
+        gl.EndFrame();
+    }
+
+} // namespace MGITest

MobileGL/MG_IntegrationTest/Scenarios/CopyImageLayeredScenario.cpp

@@ -0,0 +1,331 @@
+// MobileGL - MobileGL/MG_IntegrationTest/Scenarios/CopyImageLayeredScenario.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 - glCopyImageSubData MOVES EVERY SLICE IT WAS ASKED FOR, NOT JUST SLICE 0.
+//
+// KHR-GL43.copy_image.functional_* copies a whole 12-layer region in one call whenever both
+// endpoints are layered, i.e. for the four target pairs 2d_array->2d_array, 2d_array->3d,
+// 3d->2d_array and 3d->3d. DirectVulkan built its VkImageCopy with baseArrayLayer 0, layerCount 1
+// and srcOffset.z 0 no matter what the call asked for, so slice 0 landed correctly and slices 1..N
+// were never written - 64 conformance cases (16 compatible format pairs x those 4 pairs) failing
+// with "first mismatch at [x, y, 1]", the first texel of the first slice the copy skipped.
+//
+// The reason one hardcode covered both shapes wrongly is that GL states a layered copy ONE way -
+// srcZ/dstZ and srcDepth - while Vulkan states it two ways and picks by image type:
+//
+//   GL_TEXTURE_3D       -> VK_IMAGE_TYPE_3D: slices are z, so srcOffset.z/dstOffset.z select them
+//                          and extent.depth counts them; the layer range must stay (0, 1).
+//   GL_TEXTURE_2D_ARRAY -> VK_IMAGE_TYPE_2D: slices are array layers, so baseArrayLayer selects
+//                          them and layerCount counts them; offset.z stays 0.
+//
+// A mixed pair is legal (maintenance1, core in Vulkan 1.1) but only when the counts correspond:
+// the 3D side's extent.depth has to equal the array side's layerCount. So the four pairs below are
+// four DIFFERENT VkImageCopy shapes, not one shape with different arguments, which is why one
+// scenario per pair is the coverage that matters here.
+//
+// Every case also asserts the slices OUTSIDE the copied range still hold their fill. A backend
+// that "fixed" the miss by copying the whole image regardless of srcZ/srcDepth would pass a
+// slices-landed check and fail this one.
+//
+// The verification path is an FBO attachment per slice plus glReadPixels, not glGetTexImage: it is
+// the readback both backends share, and glFramebufferTextureLayer names an array layer and a 3D
+// slice through the same call, so the two texture kinds are read back identically.
+//
+// DirectGLES is the control - it forwards to the driver's own glCopyImageSubData - so a failure on
+// both backends means the scenario is wrong, and a failure on DirectVulkan alone means Magma is.
+
+#include <array>
+#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 kWidth = 4;
+        constexpr int kHeight = 4;
+        // Six is enough for a copy that starts and ends away from both edges of both endpoints
+        // while still leaving untouched slices on either side to assert against.
+        constexpr int kSlices = 6;
+
+        struct Rgba8 {
+            GLubyte r = 0, g = 0, b = 0, a = 0;
+
+            bool operator==(const Rgba8& other) const {
+                return r == other.r && g == other.g && b == other.b && a == other.a;
+            }
+        };
+
+        std::string Describe(const Rgba8& color) {
+            return "(" + std::to_string(color.r) + ", " + std::to_string(color.g) + ", " + std::to_string(color.b) +
+                   ", " + std::to_string(color.a) + ")";
+        }
+
+        // Per-slice constants, uniform within a slice. A uniform fill is deliberate: the defect is
+        // in which SLICE the copy addresses, and a value that also varied within the slice would
+        // make the assertions depend on the framebuffer row order as well.
+        Rgba8 SourceColor(int slice) {
+            return {static_cast<GLubyte>(10 + slice * 20), static_cast<GLubyte>(40 + slice * 10),
+                    static_cast<GLubyte>(200 - slice * 15), 255};
+        }
+
+        Rgba8 DestinationFill(int slice) {
+            return {static_cast<GLubyte>(3 + slice), static_cast<GLubyte>(250 - slice * 7),
+                    static_cast<GLubyte>(120 + slice * 5), 255};
+        }
+
+        class CopyImageLayeredScenario : public ScenarioTest {
+        protected:
+            void SetUp() override {
+                ScenarioTest::SetUp();
+                if (!Ready()) return;
+                if (!CopyImageSubDataUsable()) {
+                    GTEST_SKIP() << "glCopyImageSubData is unavailable on backend " << Gl().BackendName();
+                }
+            }
+
+            void TearDown() override {
+                if (!Ready()) return;
+                for (const GLuint texture : m_textures) {
+                    glDeleteTextures(1, &texture);
+                }
+                m_textures.clear();
+                if (m_fbo != 0) {
+                    glBindFramebuffer(GL_FRAMEBUFFER, 0);
+                    glDeleteFramebuffers(1, &m_fbo);
+                    m_fbo = 0;
+                }
+            }
+
+            // A trivial 1x1x1 array-to-array copy: it exercises the entry point without depending
+            // on any of the behaviour under test, so a driver (or a backend function table) that
+            // simply does not have the call skips instead of failing every case below.
+            bool CopyImageSubDataUsable() {
+                GLuint probe[2] = {0, 0};
+                glGenTextures(2, probe);
+                for (const GLuint texture : probe) {
+                    glBindTexture(GL_TEXTURE_2D_ARRAY, texture);
+                    glTexStorage3D(GL_TEXTURE_2D_ARRAY, 1, GL_RGBA8, 1, 1, 1);
+                }
+                glBindTexture(GL_TEXTURE_2D_ARRAY, 0);
+                while (glGetError() != GL_NO_ERROR) {
+                }
+                glCopyImageSubData(probe[0], GL_TEXTURE_2D_ARRAY, 0, 0, 0, 0, probe[1], GL_TEXTURE_2D_ARRAY, 0, 0, 0,
+                                   0, 1, 1, 1);
+                const bool usable = glGetError() == GL_NO_ERROR;
+                glDeleteTextures(2, probe);
+                return usable;
+            }
+
+            // `target` is GL_TEXTURE_2D_ARRAY or GL_TEXTURE_3D; both take glTexStorage3D and
+            // glTexSubImage3D with the slice on the same axis, which is the whole reason GL can
+            // copy between them. `levels` > 1 puts a real mip chain behind the level the copy
+            // names, so the level's own extent - a 3D level's depth included - has to be resolved
+            // rather than assumed to be the image's.
+            GLuint MakeTexture(GLenum target, int levels, Rgba8 (*colorForSlice)(int)) {
+                GLuint texture = 0;
+                glGenTextures(1, &texture);
+                m_textures.push_back(texture);
+                glBindTexture(target, texture);
+                glTexStorage3D(target, levels, GL_RGBA8, kWidth << (levels - 1), kHeight << (levels - 1),
+                               target == GL_TEXTURE_3D ? (kSlices << (levels - 1)) : kSlices);
+                glTexParameteri(target, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
+                glTexParameteri(target, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
+
+                // Fill every level, so nothing below can pass by reading a level that was never
+                // written and happened to hold the expected bytes.
+                for (int level = 0; level < levels; ++level) {
+                    const int levelWidth = kWidth << (levels - 1 - level);
+                    const int levelHeight = kHeight << (levels - 1 - level);
+                    const int levelSlices =
+                        target == GL_TEXTURE_3D ? (kSlices << (levels - 1 - level)) : kSlices;
+                    for (int slice = 0; slice < levelSlices; ++slice) {
+                        const Rgba8 color = colorForSlice(slice % kSlices);
+                        std::vector<Rgba8> texels(static_cast<size_t>(levelWidth) * levelHeight, color);
+                        glTexSubImage3D(target, level, 0, 0, slice, levelWidth, levelHeight, 1, GL_RGBA,
+                                        GL_UNSIGNED_BYTE, texels.data());
+                    }
+                }
+                glBindTexture(target, 0);
+                return texture;
+            }
+
+            // One slice of one level, through an FBO attachment. glFramebufferTextureLayer takes an
+            // array layer and a 3D slice through the same argument, so both targets read back the
+            // same way.
+            Rgba8 ReadSlice(GLuint texture, int level, int slice, int width, int height) {
+                if (m_fbo == 0) {
+                    glGenFramebuffers(1, &m_fbo);
+                }
+                glBindFramebuffer(GL_FRAMEBUFFER, m_fbo);
+                glFramebufferTextureLayer(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, texture, level, slice);
+                EXPECT_EQ(glCheckFramebufferStatus(GL_FRAMEBUFFER), static_cast<GLenum>(GL_FRAMEBUFFER_COMPLETE))
+                    << "slice " << slice << " of level " << level << " is not attachable";
+                std::vector<Rgba8> pixels(static_cast<size_t>(width) * height, Rgba8{});
+                glReadBuffer(GL_COLOR_ATTACHMENT0);
+                glPixelStorei(GL_PACK_ALIGNMENT, 1);
+                glReadPixels(0, 0, width, height, GL_RGBA, GL_UNSIGNED_BYTE, pixels.data());
+                glBindFramebuffer(GL_FRAMEBUFFER, 0);
+
+                // The fill is uniform within a slice, so any disagreement between texels is itself
+                // a failure - reported here rather than silently reduced to pixels[0].
+                for (size_t i = 1; i < pixels.size(); ++i) {
+                    EXPECT_TRUE(pixels[i] == pixels[0])
+                        << "slice " << slice << " of level " << level << " is not uniform: texel 0 is "
+                        << Describe(pixels[0]) << ", texel " << i << " is " << Describe(pixels[i]);
+                }
+                return pixels[0];
+            }
+
+            // The assertion every case ends with: slices inside [dstZ, dstZ + depth) hold the
+            // source slice they were fed, and every slice outside it still holds its own fill.
+            void ExpectCopied(GLuint destination, int level, int width, int height, int sliceCount, int srcZ,
+                              int dstZ, int depth, const char* what) {
+                for (int slice = 0; slice < sliceCount; ++slice) {
+                    const bool inRange = slice >= dstZ && slice < dstZ + depth;
+                    const Rgba8 expected =
+                        inRange ? SourceColor(srcZ + (slice - dstZ)) : DestinationFill(slice);
+                    const Rgba8 actual = ReadSlice(destination, level, slice, width, height);
+                    EXPECT_TRUE(actual == expected)
+                        << what << ": destination slice " << slice << (inRange ? " (copied)" : " (untouched)")
+                        << " is " << Describe(actual) << ", expected " << Describe(expected);
+                }
+            }
+
+            std::vector<GLuint> m_textures;
+            GLuint m_fbo = 0;
+        };
+
+        // 2d_array -> 2d_array. Both endpoints put the slices on the layer axis, so BOTH layer
+        // counts carry the depth and extent.depth must stay 1.
+        TEST_F(CopyImageLayeredScenario, ArrayToArrayCopiesEverySlice) {
+            if (!Ready() || IsSkipped()) return;
+
+            const GLuint source = MakeTexture(GL_TEXTURE_2D_ARRAY, 1, SourceColor);
+            const GLuint destination = MakeTexture(GL_TEXTURE_2D_ARRAY, 1, DestinationFill);
+            ASSERT_EQ(glGetError(), static_cast<GLenum>(GL_NO_ERROR)) << "texture setup failed";
+
+            glCopyImageSubData(source, GL_TEXTURE_2D_ARRAY, 0, 0, 0, 0, destination, GL_TEXTURE_2D_ARRAY, 0, 0, 0, 0,
+                               kWidth, kHeight, kSlices);
+            ASSERT_EQ(glGetError(), static_cast<GLenum>(GL_NO_ERROR)) << "glCopyImageSubData raised an error";
+
+            ExpectCopied(destination, 0, kWidth, kHeight, kSlices, 0, 0, kSlices, "array->array, all slices");
+        }
+
+        // The same pair with the layer ranges offset differently on the two sides: the shape that
+        // separates "copies more than slice 0" from "copies the RIGHT slices". A backend that read
+        // the source range but wrote from layer 0 (or vice versa) passes the case above.
+        TEST_F(CopyImageLayeredScenario, ArrayToArrayHonoursDifferentLayerOffsets) {
+            if (!Ready() || IsSkipped()) return;
+
+            const GLuint source = MakeTexture(GL_TEXTURE_2D_ARRAY, 1, SourceColor);
+            const GLuint destination = MakeTexture(GL_TEXTURE_2D_ARRAY, 1, DestinationFill);
+            ASSERT_EQ(glGetError(), static_cast<GLenum>(GL_NO_ERROR)) << "texture setup failed";
+
+            constexpr int kSrcZ = 3;
+            constexpr int kDstZ = 1;
+            constexpr int kDepth = 2;
+            glCopyImageSubData(source, GL_TEXTURE_2D_ARRAY, 0, 0, 0, kSrcZ, destination, GL_TEXTURE_2D_ARRAY, 0, 0, 0,
+                               kDstZ, kWidth, kHeight, kDepth);
+            ASSERT_EQ(glGetError(), static_cast<GLenum>(GL_NO_ERROR)) << "glCopyImageSubData raised an error";
+
+            ExpectCopied(destination, 0, kWidth, kHeight, kSlices, kSrcZ, kDstZ, kDepth,
+                         "array->array, offset layer ranges");
+        }
+
+        // 3d -> 3d. Neither endpoint has array layers at all: the depth travels on extent.depth and
+        // the offsets on srcOffset.z/dstOffset.z, with both layer counts pinned to 1.
+        TEST_F(CopyImageLayeredScenario, VolumeToVolumeHonoursNonZeroZ) {
+            if (!Ready() || IsSkipped()) return;
+
+            const GLuint source = MakeTexture(GL_TEXTURE_3D, 1, SourceColor);
+            const GLuint destination = MakeTexture(GL_TEXTURE_3D, 1, DestinationFill);
+            ASSERT_EQ(glGetError(), static_cast<GLenum>(GL_NO_ERROR)) << "texture setup failed";
+
+            constexpr int kSrcZ = 1;
+            constexpr int kDstZ = 3;
+            constexpr int kDepth = 3;
+            glCopyImageSubData(source, GL_TEXTURE_3D, 0, 0, 0, kSrcZ, destination, GL_TEXTURE_3D, 0, 0, 0, kDstZ,
+                               kWidth, kHeight, kDepth);
+            ASSERT_EQ(glGetError(), static_cast<GLenum>(GL_NO_ERROR)) << "glCopyImageSubData raised an error";
+
+            ExpectCopied(destination, 0, kWidth, kHeight, kSlices, kSrcZ, kDstZ, kDepth, "3d->3d, non-zero z");
+        }
+
+        // The same pair one mip level down. A 3D level's DEPTH halves with its width and height, so
+        // this is the only case where the slice count the copy may name is not the image's own -
+        // the bound a layered endpoint is checked against has to come from the level.
+        TEST_F(CopyImageLayeredScenario, VolumeToVolumeAtNonZeroMipLevel) {
+            if (!Ready() || IsSkipped()) return;
+
+            const GLuint source = MakeTexture(GL_TEXTURE_3D, 2, SourceColor);
+            const GLuint destination = MakeTexture(GL_TEXTURE_3D, 2, DestinationFill);
+            ASSERT_EQ(glGetError(), static_cast<GLenum>(GL_NO_ERROR)) << "texture setup failed";
+
+            constexpr int kLevel = 1;
+            constexpr int kSrcZ = 2;
+            constexpr int kDstZ = 0;
+            constexpr int kDepth = 4;
+            glCopyImageSubData(source, GL_TEXTURE_3D, kLevel, 0, 0, kSrcZ, destination, GL_TEXTURE_3D, kLevel, 0, 0,
+                               kDstZ, kWidth, kHeight, kDepth);
+            ASSERT_EQ(glGetError(), static_cast<GLenum>(GL_NO_ERROR)) << "glCopyImageSubData raised an error";
+
+            ExpectCopied(destination, kLevel, kWidth, kHeight, kSlices, kSrcZ, kDstZ, kDepth,
+                         "3d->3d at mip level 1");
+        }
+
+        // 2d_array -> 3d. The mixed shape: the source counts its slices as layers, the destination
+        // as depth, and Vulkan requires extent.depth to equal the source's layerCount.
+        TEST_F(CopyImageLayeredScenario, ArrayToVolumeCopiesEverySlice) {
+            if (!Ready() || IsSkipped()) return;
+
+            const GLuint source = MakeTexture(GL_TEXTURE_2D_ARRAY, 1, SourceColor);
+            const GLuint destination = MakeTexture(GL_TEXTURE_3D, 1, DestinationFill);
+            ASSERT_EQ(glGetError(), static_cast<GLenum>(GL_NO_ERROR)) << "texture setup failed";
+
+            constexpr int kSrcZ = 2;
+            constexpr int kDstZ = 1;
+            constexpr int kDepth = 4;
+            glCopyImageSubData(source, GL_TEXTURE_2D_ARRAY, 0, 0, 0, kSrcZ, destination, GL_TEXTURE_3D, 0, 0, 0, kDstZ,
+                               kWidth, kHeight, kDepth);
+            ASSERT_EQ(glGetError(), static_cast<GLenum>(GL_NO_ERROR)) << "glCopyImageSubData raised an error";
+
+            ExpectCopied(destination, 0, kWidth, kHeight, kSlices, kSrcZ, kDstZ, kDepth, "2d_array->3d");
+        }
+
+        // 3d -> 2d_array, the mirror image: the depth now has to reach the DESTINATION's layerCount
+        // while the source states it as extent.depth from a z offset.
+        TEST_F(CopyImageLayeredScenario, VolumeToArrayCopiesEverySlice) {
+            if (!Ready() || IsSkipped()) return;
+
+            const GLuint source = MakeTexture(GL_TEXTURE_3D, 1, SourceColor);
+            const GLuint destination = MakeTexture(GL_TEXTURE_2D_ARRAY, 1, DestinationFill);
+            ASSERT_EQ(glGetError(), static_cast<GLenum>(GL_NO_ERROR)) << "texture setup failed";
+
+            constexpr int kSrcZ = 1;
+            constexpr int kDstZ = 2;
+            constexpr int kDepth = 4;
+            glCopyImageSubData(source, GL_TEXTURE_3D, 0, 0, 0, kSrcZ, destination, GL_TEXTURE_2D_ARRAY, 0, 0, 0, kDstZ,
+                               kWidth, kHeight, kDepth);
+            ASSERT_EQ(glGetError(), static_cast<GLenum>(GL_NO_ERROR)) << "glCopyImageSubData raised an error";
+
+            ExpectCopied(destination, 0, kWidth, kHeight, kSlices, kSrcZ, kDstZ, kDepth, "3d->2d_array");
+        }
+
+    } // namespace
+} // namespace MGITest

MobileGL/MG_IntegrationTest/Scenarios/CopyImageLevelRangeScenario.cpp

@@ -0,0 +1,209 @@
+// MobileGL - MobileGL/MG_IntegrationTest/Scenarios/CopyImageLevelRangeScenario.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
+//
+// KHR-GL43.copy_image.non_existent_mipmap, and what it cost.
+//
+// The CTS case is a pure negative test: two 16x16 textures that have level 0 and
+// nothing else, and a glCopyImageSubData naming level 1. The answer is
+// GL_INVALID_VALUE (GL 4.6 core 18.3.2 / ARB_copy_image: "srcLevel/dstLevel is not
+// a valid level"). MobileGL's frontend only checked the level against
+// GL_MAX_TEXTURE_SIZE, so level 1 sailed through into the backends, DirectVulkan
+// resolved it into a VkImageCopy subresource on a VkImage that was created with
+// exactly one mip level, and the Adreno driver dereferenced the level it was
+// promised - SIGSEGV inside vkCmdCopyImage, taking the whole glcts process down
+// mid-run. A negative case must never do that.
+//
+// So the level-1-on-a-one-level-texture rejection is the regression proper, and the
+// rest of this file is what keeps the fix honest. A validator that answered
+// GL_INVALID_VALUE to every level would satisfy the regression tests alone, so the
+// scenarios below pin the BOUNDARY rather than the symptom:
+//
+//   * a texture that really does have two levels must accept a copy at level 1,
+//   * the same texture must still reject level 2,
+//   * and a plain level-0 copy must move pixels, which is checked by reading the
+//     destination back rather than by trusting glGetError.
+//
+// Both backends are covered because the fix is in the shared frontend: DirectGLES
+// forwards to the ES glCopyImageSubData (whose own error lands in the ES context,
+// not in MobileGL's, so it never reached the application either) and DirectVulkan
+// records the copy itself.
+
+#include <array>
+#include <cstring>
+#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 GLsizei kSize = 16;
+
+        struct Rgba8 {
+            GLubyte r, g, b, a;
+            bool operator==(const Rgba8& other) const {
+                return r == other.r && g == other.g && b == other.b && a == other.a;
+            }
+        };
+
+        std::vector<Rgba8> SolidImage(GLsizei width, GLsizei height, Rgba8 color) {
+            return std::vector<Rgba8>(static_cast<std::size_t>(width) * static_cast<std::size_t>(height), color);
+        }
+
+        class CopyImageLevelRangeScenario : public ScenarioTest {
+        protected:
+            void SetUp() override {
+                ScenarioTest::SetUp();
+                if (!Ready()) return;
+                DrainErrors();
+            }
+
+            void TearDown() override {
+                if (!Ready()) return;
+                DeleteTextures();
+                if (m_fbo != 0) {
+                    glBindFramebuffer(GL_FRAMEBUFFER, 0);
+                    glDeleteFramebuffers(1, &m_fbo);
+                    m_fbo = 0;
+                }
+                DrainErrors();
+                ScenarioTest::TearDown();
+            }
+
+            static void DrainErrors() {
+                for (int i = 0; i < 16 && glGetError() != GL_NO_ERROR; ++i) {
+                }
+            }
+
+            void DeleteTextures() {
+                if (m_src != 0) glDeleteTextures(1, &m_src);
+                if (m_dst != 0) glDeleteTextures(1, &m_dst);
+                m_src = 0;
+                m_dst = 0;
+            }
+
+            // One 16x16 RGBA8 texture with `levelCount` levels defined through
+            // glTexImage2D - the same way the CTS case builds its textures, and
+            // deliberately NOT glTexStorage2D: an immutable allocation would define the
+            // whole chain up front and could not express "level 1 does not exist".
+            GLuint MakeTexture(int levelCount, Rgba8 baseColor) {
+                GLuint texture = 0;
+                glGenTextures(1, &texture);
+                glBindTexture(GL_TEXTURE_2D, texture);
+                for (int level = 0; level < levelCount; ++level) {
+                    const GLsizei extent = kSize >> level;
+                    const std::vector<Rgba8> pixels = SolidImage(extent, extent, baseColor);
+                    glTexImage2D(GL_TEXTURE_2D, level, GL_RGBA8, extent, extent, 0, GL_RGBA, GL_UNSIGNED_BYTE,
+                                 pixels.data());
+                }
+                // What Utils::makeTextureComplete does in the CTS case: the texture is
+                // complete for the levels it actually has, not for a chain it does not.
+                glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_BASE_LEVEL, 0);
+                glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAX_LEVEL, levelCount - 1);
+                glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
+                glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
+                glBindTexture(GL_TEXTURE_2D, 0);
+                return texture;
+            }
+
+            void MakePair(int levelCount) {
+                DeleteTextures();
+                m_src = MakeTexture(levelCount, Rgba8{11, 22, 33, 255});
+                m_dst = MakeTexture(levelCount, Rgba8{200, 100, 50, 255});
+                ASSERT_EQ(glGetError(), GL_NO_ERROR) << "texture setup with " << levelCount << " level(s)";
+            }
+
+            // The call under test, at whatever levels the caller wants, over a 1x1
+            // region so the region check can never be what rejects it.
+            GLenum CopyAt(GLint srcLevel, GLint dstLevel, GLsizei extent = 1) {
+                DrainErrors();
+                glCopyImageSubData(m_src, GL_TEXTURE_2D, srcLevel, 0, 0, 0, m_dst, GL_TEXTURE_2D, dstLevel, 0, 0, 0,
+                                   extent, extent, 1);
+                const GLenum error = glGetError();
+                // A second pending error would mean the entry point queued more than one,
+                // and the extra would be handed out at an unrelated call site later.
+                EXPECT_EQ(glGetError(), GL_NO_ERROR) << "the copy recorded more than one error";
+                return error;
+            }
+
+            Rgba8 ReadBackDestinationLevel0() {
+                if (m_fbo == 0) glGenFramebuffers(1, &m_fbo);
+                glBindFramebuffer(GL_FRAMEBUFFER, m_fbo);
+                glFramebufferTexture2D(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_TEXTURE_2D, m_dst, 0);
+                const GLenum status = glCheckFramebufferStatus(GL_FRAMEBUFFER);
+                if (status != GL_FRAMEBUFFER_COMPLETE) {
+                    ADD_FAILURE() << "readback framebuffer incomplete: " << status;
+                    glBindFramebuffer(GL_FRAMEBUFFER, 0);
+                    return Rgba8{0, 0, 0, 0};
+                }
+                Rgba8 texel{0, 0, 0, 0};
+                glReadPixels(0, 0, 1, 1, GL_RGBA, GL_UNSIGNED_BYTE, &texel);
+                glBindFramebuffer(GL_FRAMEBUFFER, 0);
+                return texel;
+            }
+
+            GLuint m_src = 0;
+            GLuint m_dst = 0;
+            GLuint m_fbo = 0;
+        };
+
+        // The regression. Level 1 of a texture that has only level 0 is not a level, and
+        // saying so is the whole job: before the fix this reached DirectVulkan, which
+        // handed mipLevel=1 to vkCmdCopyImage on a one-level VkImage and died inside the
+        // Adreno driver.
+        TEST_F(CopyImageLevelRangeScenario, LevelOneOfASingleLevelTextureIsRejected) {
+            if (!Ready()) GTEST_SKIP();
+            MakePair(1);
+
+            EXPECT_EQ(CopyAt(1, 0), static_cast<GLenum>(GL_INVALID_VALUE)) << "source level 1";
+            EXPECT_EQ(CopyAt(0, 1), static_cast<GLenum>(GL_INVALID_VALUE)) << "destination level 1";
+            EXPECT_EQ(CopyAt(1, 1), static_cast<GLenum>(GL_INVALID_VALUE)) << "both levels 1";
+        }
+
+        // The negative control that makes the test above falsifiable: the same level
+        // index, on textures that genuinely have it, must be accepted. A validator that
+        // rejected every non-zero level would pass the regression test and fail here.
+        TEST_F(CopyImageLevelRangeScenario, LevelOneOfATwoLevelTextureIsAccepted) {
+            if (!Ready()) GTEST_SKIP();
+            MakePair(2);
+
+            EXPECT_EQ(CopyAt(1, 1), static_cast<GLenum>(GL_NO_ERROR));
+        }
+
+        // And the boundary from the other side: two levels means 0 and 1, not 2.
+        TEST_F(CopyImageLevelRangeScenario, LevelTwoOfATwoLevelTextureIsRejected) {
+            if (!Ready()) GTEST_SKIP();
+            MakePair(2);
+
+            EXPECT_EQ(CopyAt(2, 0), static_cast<GLenum>(GL_INVALID_VALUE)) << "source level 2";
+            EXPECT_EQ(CopyAt(0, 2), static_cast<GLenum>(GL_INVALID_VALUE)) << "destination level 2";
+        }
+
+        // Errors alone cannot tell an accepted copy from a silently dropped one, so the
+        // ordinary case is checked by reading the destination back: the copy has to move
+        // the source's texel, not merely decline to complain.
+        TEST_F(CopyImageLevelRangeScenario, AValidLevelZeroCopyStillMovesPixels) {
+            if (!Ready()) GTEST_SKIP();
+            MakePair(1);
+
+            ASSERT_EQ(ReadBackDestinationLevel0(), (Rgba8{200, 100, 50, 255})) << "destination before the copy";
+            EXPECT_EQ(CopyAt(0, 0, kSize), static_cast<GLenum>(GL_NO_ERROR));
+            EXPECT_EQ(ReadBackDestinationLevel0(), (Rgba8{11, 22, 33, 255})) << "destination after the copy";
+        }
+
+    } // namespace
+} // namespace MGITest

MobileGL/MG_IntegrationTest/Scenarios/DepthStencilReadbackAttachmentShapeScenario.cpp

@@ -0,0 +1,370 @@
+// MobileGL - MobileGL/MG_IntegrationTest/Scenarios/DepthStencilReadbackAttachmentShapeScenario.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 - DEPTH/STENCIL READBACK WHEN THE ATTACHMENT IS NOT A PLAIN GL_TEXTURE_2D,
+// AND THE DEFAULT FRAMEBUFFER'S ADVERTISED DEPTH/STENCIL FORMAT.
+//
+// Three shipped defects, all of them invisible to a test that only ever attaches a 2D texture
+// or only ever asks the default framebuffer for a colour value.
+//
+// (1) The ES depth/stencil readback emulation identifies the source format by binding the
+//     attachment's texture NAME to GL_TEXTURE_2D and asking that target for its internal
+//     format. A name whose target is GL_TEXTURE_2D_ARRAY (attached by
+//     glFramebufferTextureLayer) makes the bind answer GL_INVALID_OPERATION and change
+//     nothing - so the query then truthfully describes whatever texture was already on
+//     GL_TEXTURE_2D, which on that path is the emulation's own staging scratch. A wrong
+//     answer that looks like a right one: the staging blit is issued between mismatched
+//     depth formats, ES rejects it, and the read reports nothing at all.
+//
+// (2) Adreno answers GL_NONE for GL_FRAMEBUFFER_ATTACHMENT_OBJECT_TYPE on an attachment made
+//     by glFramebufferTexture (a cube map, attached layered) while still reporting its depth
+//     and stencil bits correctly. The emulation took OBJECT_TYPE as the sole witness for "is
+//     there an aspect here at all" and declined the whole read.
+//
+// (3) DirectGLES never told the frontend what its default framebuffer's depth/stencil format
+//     actually is, so the placeholder from MG_Impl/Init.cpp - GL_DEPTH32F_STENCIL8 - was what
+//     every attachment query answered, whatever the surface really had. That is not cosmetic:
+//     GL blits depth/stencil only between IDENTICAL formats, so an application that reads
+//     GL_FRAMEBUFFER_ATTACHMENT_DEPTH_SIZE, allocates the buffer it was just told about and
+//     blits gets GL_INVALID_OPERATION - and a rejected glBlitFramebuffer transfers NOTHING,
+//     colour bits included. DirectVulkan has published its real format since the swapchain
+//     work; this is the half that was missing.
+//
+// Every case poisons its destination with a value the correct answer cannot be, so "the
+// backend wrote nothing" fails loudly instead of passing on stale memory. The plain
+// GL_TEXTURE_2D case at the end is the built-in control: it shares every line of the readback
+// path with the array and cube cases, so its passing is what says a failure above is about the
+// attachment's SHAPE and not about depth readback in general.
+//
+// The scenario name starts with DepthStencilReadback on purpose - that is the filter the
+// forced-emulation ctest registration uses (MG_IntegrationTest/CMakeLists.txt), and without
+// that registration these cases are unfalsifiable on llvmpipe, which accepts the native ES
+// depth reads that the Adreno device does not have.
+
+#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 float kDepthPoison = 0.2f;
+        constexpr int kStencilPoison = 50;
+        constexpr float kDepthValue = 0.75f;
+        constexpr int kStencilValue = 7;
+        constexpr int kSize = 16;
+
+        class DepthStencilReadbackAttachmentShapeScenario : public ScenarioTest {
+        protected:
+            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;
+            }
+
+            // Clears the currently bound framebuffer's depth and stencil to the shared
+            // reference values, with both write masks explicitly open (glClear honours them,
+            // and a leftover mask from another scenario in this shared context would look
+            // exactly like the bug under test).
+            void ClearDepthStencil() const {
+                glDepthMask(GL_TRUE);
+                glStencilMask(0xFFu);
+                glDisable(GL_SCISSOR_TEST);
+                glClearDepth(kDepthValue);
+                glClearStencil(kStencilValue);
+                glClear(GL_DEPTH_BUFFER_BIT | GL_STENCIL_BUFFER_BIT);
+            }
+        };
+
+        // Fails the calling test if the framebuffer bound at both targets is not complete;
+        // an incomplete framebuffer would make every read below return the poison for a
+        // reason that has nothing to do with what is being tested.
+        ::testing::AssertionResult FramebufferIsComplete() {
+            const GLenum status = glCheckFramebufferStatus(GL_FRAMEBUFFER);
+            if (status == GL_FRAMEBUFFER_COMPLETE) return ::testing::AssertionSuccess();
+            return ::testing::AssertionFailure() << "framebuffer status 0x" << std::hex << status;
+        }
+
+    } // namespace
+
+    // (1) A depth slice of a 2D ARRAY texture, attached with glFramebufferTextureLayer.
+    // Pre-fix this read back the poison: the format probe answered with the staging scratch's
+    // GL_DEPTH24_STENCIL8 instead of the array's GL_DEPTH_COMPONENT24, and the mismatched
+    // staging blit was rejected.
+    TEST_F(DepthStencilReadbackAttachmentShapeScenario, DepthOfAnArrayLayerAttachmentReadsBack) {
+        if (!Ready()) return;
+        HeadlessGL& gl = Gl();
+
+        GLuint fbo = 0;
+        GLuint depthArray = 0;
+        glGenFramebuffers(1, &fbo);
+        glGenTextures(1, &depthArray);
+        glBindTexture(GL_TEXTURE_2D_ARRAY, depthArray);
+        glTexStorage3D(GL_TEXTURE_2D_ARRAY, 1, GL_DEPTH_COMPONENT24, kSize, kSize, 4);
+        glBindTexture(GL_TEXTURE_2D_ARRAY, 0);
+
+        glBindFramebuffer(GL_FRAMEBUFFER, fbo);
+        // Layer 2, not layer 0: a backend that silently reads the wrong slice would still
+        // agree with a single-layer texture.
+        glFramebufferTextureLayer(GL_FRAMEBUFFER, GL_DEPTH_ATTACHMENT, depthArray, 0, 2);
+        glDrawBuffer(GL_NONE);
+        glReadBuffer(GL_NONE);
+        EXPECT_EQ(FirstGLError(), 0u);
+        ASSERT_TRUE(FramebufferIsComplete());
+
+        glViewport(0, 0, kSize, kSize);
+        ClearDepthStencil();
+
+        const float depth = ReadDepthAt(kSize / 2, kSize / 2);
+        EXPECT_EQ(FirstGLError(), 0u);
+        EXPECT_NEAR(depth, kDepthValue, 1.0f / 4096.0f)
+            << "glReadPixels(GL_DEPTH_COMPONENT) of a GL_TEXTURE_2D_ARRAY layer attachment returned " << depth
+            << (std::fabs(depth - kDepthPoison) < 1e-6f ? " - the destination was never written at all" : "");
+
+        BindDefaultFramebuffer();
+        glDeleteFramebuffers(1, &fbo);
+        glDeleteTextures(1, &depthArray);
+        gl.EndFrame();
+    }
+
+    // (2) A depth cube map, attached whole with glFramebufferTexture - a LAYERED attachment.
+    // Pre-fix the emulation declined outright, because the driver reports GL_NONE for that
+    // attachment's OBJECT_TYPE.
+    TEST_F(DepthStencilReadbackAttachmentShapeScenario, DepthOfALayeredCubeAttachmentReadsBack) {
+        if (!Ready()) return;
+        HeadlessGL& gl = Gl();
+
+        GLuint fbo = 0;
+        GLuint depthCube = 0;
+        glGenFramebuffers(1, &fbo);
+        glGenTextures(1, &depthCube);
+        glBindTexture(GL_TEXTURE_CUBE_MAP, depthCube);
+        glTexStorage2D(GL_TEXTURE_CUBE_MAP, 1, GL_DEPTH_COMPONENT24, kSize, kSize);
+        glBindTexture(GL_TEXTURE_CUBE_MAP, 0);
+
+        glBindFramebuffer(GL_FRAMEBUFFER, fbo);
+        glFramebufferTexture(GL_FRAMEBUFFER, GL_DEPTH_ATTACHMENT, depthCube, 0);
+        glDrawBuffer(GL_NONE);
+        glReadBuffer(GL_NONE);
+        EXPECT_EQ(FirstGLError(), 0u);
+        ASSERT_TRUE(FramebufferIsComplete());
+
+        glViewport(0, 0, kSize, kSize);
+        ClearDepthStencil();
+
+        const float depth = ReadDepthAt(kSize / 2, kSize / 2);
+        EXPECT_EQ(FirstGLError(), 0u);
+        EXPECT_NEAR(depth, kDepthValue, 1.0f / 4096.0f)
+            << "glReadPixels(GL_DEPTH_COMPONENT) of a layered GL_TEXTURE_CUBE_MAP attachment returned " << depth
+            << (std::fabs(depth - kDepthPoison) < 1e-6f ? " - the destination was never written at all" : "");
+
+        BindDefaultFramebuffer();
+        glDeleteFramebuffers(1, &fbo);
+        glDeleteTextures(1, &depthCube);
+        gl.EndFrame();
+    }
+
+    // Both aspects of a packed array attachment. The stencil half goes through a different
+    // sampling mode than the depth half, and only the depth half was covered above.
+    TEST_F(DepthStencilReadbackAttachmentShapeScenario, PackedArrayLayerAttachmentReadsBackBothAspects) {
+        if (!Ready()) return;
+        HeadlessGL& gl = Gl();
+
+        GLuint fbo = 0;
+        GLuint packedArray = 0;
+        glGenFramebuffers(1, &fbo);
+        glGenTextures(1, &packedArray);
+        glBindTexture(GL_TEXTURE_2D_ARRAY, packedArray);
+        glTexStorage3D(GL_TEXTURE_2D_ARRAY, 1, GL_DEPTH24_STENCIL8, kSize, kSize, 3);
+        glBindTexture(GL_TEXTURE_2D_ARRAY, 0);
+
+        glBindFramebuffer(GL_FRAMEBUFFER, fbo);
+        glFramebufferTextureLayer(GL_FRAMEBUFFER, GL_DEPTH_STENCIL_ATTACHMENT, packedArray, 0, 1);
+        glDrawBuffer(GL_NONE);
+        glReadBuffer(GL_NONE);
+        EXPECT_EQ(FirstGLError(), 0u);
+        ASSERT_TRUE(FramebufferIsComplete());
+
+        glViewport(0, 0, kSize, kSize);
+        ClearDepthStencil();
+
+        const float depth = ReadDepthAt(kSize / 2, kSize / 2);
+        const int stencil = ReadStencilAt(kSize / 2, kSize / 2);
+        EXPECT_EQ(FirstGLError(), 0u);
+        EXPECT_NEAR(depth, kDepthValue, 1.0f / 4096.0f)
+            << "depth of a packed GL_TEXTURE_2D_ARRAY layer attachment returned " << depth;
+        EXPECT_EQ(stencil, kStencilValue)
+            << "stencil of a packed GL_TEXTURE_2D_ARRAY layer attachment returned " << stencil
+            << (stencil == kStencilPoison ? " - the destination was never written at all" : "");
+
+        BindDefaultFramebuffer();
+        glDeleteFramebuffers(1, &fbo);
+        glDeleteTextures(1, &packedArray);
+        gl.EndFrame();
+    }
+
+    // The control: the plain GL_TEXTURE_2D shape, which always worked. If this one ever fails
+    // alongside the three above, the fault is in depth readback generally rather than in how
+    // the attachment's format and presence are discovered.
+    TEST_F(DepthStencilReadbackAttachmentShapeScenario, DepthOfAPlainTexture2DAttachmentReadsBack) {
+        if (!Ready()) return;
+        HeadlessGL& gl = Gl();
+
+        GLuint fbo = 0;
+        GLuint depthTex = 0;
+        glGenFramebuffers(1, &fbo);
+        glGenTextures(1, &depthTex);
+        glBindTexture(GL_TEXTURE_2D, depthTex);
+        glTexStorage2D(GL_TEXTURE_2D, 1, GL_DEPTH_COMPONENT24, kSize, kSize);
+        glBindTexture(GL_TEXTURE_2D, 0);
+
+        glBindFramebuffer(GL_FRAMEBUFFER, fbo);
+        glFramebufferTexture2D(GL_FRAMEBUFFER, GL_DEPTH_ATTACHMENT, GL_TEXTURE_2D, depthTex, 0);
+        glDrawBuffer(GL_NONE);
+        glReadBuffer(GL_NONE);
+        EXPECT_EQ(FirstGLError(), 0u);
+        ASSERT_TRUE(FramebufferIsComplete());
+
+        glViewport(0, 0, kSize, kSize);
+        ClearDepthStencil();
+
+        const float depth = ReadDepthAt(kSize / 2, kSize / 2);
+        EXPECT_EQ(FirstGLError(), 0u);
+        EXPECT_NEAR(depth, kDepthValue, 1.0f / 4096.0f)
+            << "the control case failed: even a plain GL_TEXTURE_2D depth attachment read back " << depth;
+
+        BindDefaultFramebuffer();
+        glDeleteFramebuffers(1, &fbo);
+        glDeleteTextures(1, &depthTex);
+        gl.EndFrame();
+    }
+
+    // (3) The default framebuffer must describe its depth/stencil truthfully enough that a
+    // buffer allocated from that description is blit-compatible with it. This is the exact
+    // sequence KHR-GLxx.framebuffer_blit performs, and the exact reason 22 of its cases died
+    // on DirectGLES: the frontend answered 32-bit float depth for a 24-bit fixed-point
+    // surface, so the renderbuffer the caller allocated could never be blitted to.
+    TEST_F(DepthStencilReadbackAttachmentShapeScenario, DefaultFramebufferDepthStencilFormatIsBlitCompatible) {
+        if (!Ready()) return;
+        HeadlessGL& gl = Gl();
+        const int width = gl.Width();
+        const int height = gl.Height();
+
+        BindDefaultFramebuffer();
+        GLint depthBits = 0;
+        GLint stencilBits = 0;
+        GLint componentType = GL_UNSIGNED_NORMALIZED;
+        glGetFramebufferAttachmentParameteriv(GL_DRAW_FRAMEBUFFER, GL_DEPTH,
+                                              GL_FRAMEBUFFER_ATTACHMENT_DEPTH_SIZE, &depthBits);
+        glGetFramebufferAttachmentParameteriv(GL_DRAW_FRAMEBUFFER, GL_STENCIL,
+                                              GL_FRAMEBUFFER_ATTACHMENT_STENCIL_SIZE, &stencilBits);
+        glGetFramebufferAttachmentParameteriv(GL_DRAW_FRAMEBUFFER, GL_DEPTH,
+                                              GL_FRAMEBUFFER_ATTACHMENT_COMPONENT_TYPE, &componentType);
+        EXPECT_EQ(FirstGLError(), 0u);
+        if (depthBits <= 0 || stencilBits <= 0) {
+            GTEST_SKIP() << "this surface has no packed depth/stencil (depth=" << depthBits
+                         << " stencil=" << stencilBits << "); the blit-compatibility contract needs both";
+        }
+
+        // The one sized format the reported description names. Getting here with the wrong
+        // answer is the bug: the two candidates are not interchangeable for a blit.
+        const GLenum reported = (componentType == GL_FLOAT || depthBits > 24) ? GL_DEPTH32F_STENCIL8
+                                                                             : GL_DEPTH24_STENCIL8;
+
+        GLuint fbo = 0;
+        GLuint colorRbo = 0;
+        GLuint depthRbo = 0;
+        glGenFramebuffers(1, &fbo);
+        glGenRenderbuffers(1, &colorRbo);
+        glGenRenderbuffers(1, &depthRbo);
+        glBindRenderbuffer(GL_RENDERBUFFER, colorRbo);
+        glRenderbufferStorage(GL_RENDERBUFFER, GL_RGBA8, width, height);
+        glBindRenderbuffer(GL_RENDERBUFFER, depthRbo);
+        glRenderbufferStorage(GL_RENDERBUFFER, reported, width, height);
+        glBindRenderbuffer(GL_RENDERBUFFER, 0);
+        glBindFramebuffer(GL_FRAMEBUFFER, fbo);
+        glFramebufferRenderbuffer(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_RENDERBUFFER, colorRbo);
+        glFramebufferRenderbuffer(GL_FRAMEBUFFER, GL_DEPTH_STENCIL_ATTACHMENT, GL_RENDERBUFFER, depthRbo);
+        EXPECT_EQ(FirstGLError(), 0u);
+        ASSERT_TRUE(FramebufferIsComplete());
+
+        // Put a known depth in the default framebuffer, then blit colour+depth+stencil out of
+        // it into the buffer that its own description asked for.
+        BindDefaultFramebuffer();
+        glViewport(0, 0, width, height);
+        glColorMask(GL_TRUE, GL_TRUE, GL_TRUE, GL_TRUE);
+        glClearColor(0.0f, 1.0f, 0.0f, 1.0f);
+        glClear(GL_COLOR_BUFFER_BIT);
+        ClearDepthStencil();
+        EXPECT_EQ(FirstGLError(), 0u);
+
+        glBindFramebuffer(GL_READ_FRAMEBUFFER, 0);
+        glBindFramebuffer(GL_DRAW_FRAMEBUFFER, fbo);
+        glBlitFramebuffer(0, 0, width, height, 0, 0, width, height,
+                          GL_COLOR_BUFFER_BIT | GL_DEPTH_BUFFER_BIT | GL_STENCIL_BUFFER_BIT, GL_NEAREST);
+        EXPECT_EQ(FirstGLError(), 0u)
+            << "blitting depth/stencil out of the default framebuffer into a buffer allocated from the format "
+               "the default framebuffer itself reported was rejected - the report and the storage disagree";
+
+        glBindFramebuffer(GL_READ_FRAMEBUFFER, fbo);
+        unsigned char color[4] = {0, 0, 0, 0};
+        glReadPixels(width / 2, height / 2, 1, 1, GL_RGBA, GL_UNSIGNED_BYTE, color);
+        const float depth = ReadDepthAt(width / 2, height / 2);
+        const int stencil = ReadStencilAt(width / 2, height / 2);
+        EXPECT_EQ(FirstGLError(), 0u);
+        // The colour bit is the precondition, not the claim: it says this stack can blit out of
+        // its default framebuffer at all, which has nothing to do with depth/stencil formats.
+        // DirectVulkan on a surfaceless pbuffer cannot - the whole call, colour included, is a
+        // no-op there, while the same blit works on a real surface (KHR-GLxx.framebuffer_blit
+        // exercises exactly it and Magma passes 33/33 on device). Skipping keeps the
+        // depth/stencil claim below falsifiable instead of drowning it in an unrelated
+        // harness limitation.
+        if (int(color[1]) <= 192) {
+            // GTEST_SKIP() expands to a return, so the teardown below it would never run and this
+            // scenario would hand the next one a foreign framebuffer plus three leaked objects -
+            // and this is the path DirectVulkan takes on every headless run, not a rare one.
+            BindDefaultFramebuffer();
+            glDeleteFramebuffers(1, &fbo);
+            glDeleteRenderbuffers(1, &colorRbo);
+            glDeleteRenderbuffers(1, &depthRbo);
+            gl.EndFrame();
+            GTEST_SKIP() << "backend " << gl.BackendName() << " on this surface transferred no colour either (green="
+                         << int(color[1])
+                         << "): it cannot blit out of the default framebuffer here, so the depth/stencil half proves "
+                            "nothing. The GL-error assertion above still ran, and it is the format contract";
+        }
+        EXPECT_NEAR(depth, kDepthValue, 1.0f / 4096.0f)
+            << "depth blitted out of the default framebuffer read back " << depth
+            << (std::fabs(depth - kDepthPoison) < 1e-6f ? " - the blit transferred nothing" : "");
+        EXPECT_EQ(stencil, kStencilValue) << "stencil blitted out of the default framebuffer read back " << stencil;
+
+        BindDefaultFramebuffer();
+        glDeleteFramebuffers(1, &fbo);
+        glDeleteRenderbuffers(1, &colorRbo);
+        glDeleteRenderbuffers(1, &depthRbo);
+        gl.EndFrame();
+    }
+
+} // namespace MGITest

MobileGL/MG_IntegrationTest/Scenarios/DepthStencilReadbackMatrixScenario.cpp

@@ -0,0 +1,784 @@
+// MobileGL - MobileGL/MG_IntegrationTest/Scenarios/DepthStencilReadbackMatrixScenario.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 - THE DEPTH/STENCIL READBACK MATRIX: every verb, every source kind.
+//
+// DepthStencilReadbackScenario pins the default framebuffer. This file pins the rest of
+// the surface a depth/stencil read has to cover, because the three verbs and the four
+// source kinds do NOT share a code path by accident - they share one on purpose, and a
+// change that quietly serves only one of them is exactly what these assertions catch:
+//
+//   verbs         glReadPixels(GL_DEPTH_COMPONENT | GL_STENCIL_INDEX | GL_DEPTH_STENCIL),
+//                 glGetTexImage(GL_DEPTH_STENCIL), glCopyTexImage2D followed by a read
+//   source kinds  depth(-stencil) TEXTURE, RENDERBUFFER (not samplable at all),
+//                 MULTISAMPLE renderbuffer (needs a resolve first), default framebuffer
+//   formats       DEPTH24_STENCIL8, DEPTH32F_STENCIL8, DEPTH_COMPONENT16/24/32F,
+//                 STENCIL_INDEX8
+//   client types  GL_FLOAT / GL_UNSIGNED_INT / GL_UNSIGNED_SHORT depth, GL_INT /
+//                 GL_UNSIGNED_BYTE stencil, both packed GL_DEPTH_STENCIL layouts
+//
+// On DirectGLES none of this exists natively - ES has no depth or stencil readback in
+// core - so every assertion here is really an assertion about the shader-sampling
+// emulation. The catch is that some ES drivers accept the reads anyway (Mesa does,
+// Adreno does not), which would make the emulation dead code on the very stack the
+// headless suite runs on. That is what the second ctest registration is for: the same
+// scenarios run again with MOBILEGL_ESPRYT_FORCE_DS_READBACK_EMULATION=1, which takes the
+// native spellings off the table and leaves only the path the device actually uses.
+//
+// Every destination is poisoned with a value the correct answer cannot be, so "the
+// backend wrote nothing" fails loudly instead of passing on a coincidence - a test that
+// only checked "no GL error" would pass against a readback that never touched the buffer,
+// which is precisely how this whole cluster hid for so long.
+
+#include <cmath>
+#include <cstring>
+#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 float kDepthPoison = 0.2f;
+        constexpr int kStencilPoison = 50;
+        constexpr int kWidth = 64;
+        constexpr int kHeight = 48;
+
+        // A depth-stencil pair no clear in these tests produces, packed both ways.
+        constexpr unsigned int kPacked24_8Poison = 0xAAAAAA33u;
+
+        struct D32fS8 {
+            float depth;
+            unsigned int stencil;
+        };
+
+        // Everything a source needs to be read: the framebuffer to bind, plus the objects
+        // to delete afterwards.
+        struct DepthSource {
+            GLuint fbo = 0;
+            GLuint colorTexture = 0;
+            GLuint depthTexture = 0;
+            GLuint depthRenderbuffer = 0;
+            GLuint colorRenderbuffer = 0;
+        };
+
+        void DestroySource(DepthSource& source) {
+            if (source.fbo != 0) glDeleteFramebuffers(1, &source.fbo);
+            if (source.colorTexture != 0) glDeleteTextures(1, &source.colorTexture);
+            if (source.depthTexture != 0) glDeleteTextures(1, &source.depthTexture);
+            if (source.depthRenderbuffer != 0) glDeleteRenderbuffers(1, &source.depthRenderbuffer);
+            if (source.colorRenderbuffer != 0) glDeleteRenderbuffers(1, &source.colorRenderbuffer);
+            source = DepthSource{};
+        }
+
+        GLenum AttachmentPointFor(GLenum internalFormat) {
+            switch (internalFormat) {
+            case GL_DEPTH24_STENCIL8:
+            case GL_DEPTH32F_STENCIL8: return GL_DEPTH_STENCIL_ATTACHMENT;
+            case GL_STENCIL_INDEX8: return GL_STENCIL_ATTACHMENT;
+            default: return GL_DEPTH_ATTACHMENT;
+            }
+        }
+
+        bool FormatHasDepth(GLenum internalFormat) { return internalFormat != GL_STENCIL_INDEX8; }
+        bool FormatHasStencil(GLenum internalFormat) {
+            return internalFormat == GL_DEPTH24_STENCIL8 || internalFormat == GL_DEPTH32F_STENCIL8 ||
+                   internalFormat == GL_STENCIL_INDEX8;
+        }
+
+        // A framebuffer whose depth/stencil lives in a TEXTURE. The colour attachment is
+        // there so a stencil-only or depth-only framebuffer still has something to size it.
+        DepthSource MakeTextureSource(GLenum internalFormat) {
+            DepthSource source;
+            glGenFramebuffers(1, &source.fbo);
+            glBindFramebuffer(GL_FRAMEBUFFER, source.fbo);
+            glGenTextures(1, &source.colorTexture);
+            glBindTexture(GL_TEXTURE_2D, source.colorTexture);
+            glTexStorage2D(GL_TEXTURE_2D, 1, GL_RGBA8, kWidth, kHeight);
+            glFramebufferTexture2D(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_TEXTURE_2D, source.colorTexture, 0);
+            glGenTextures(1, &source.depthTexture);
+            glBindTexture(GL_TEXTURE_2D, source.depthTexture);
+            glTexStorage2D(GL_TEXTURE_2D, 1, internalFormat, kWidth, kHeight);
+            glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
+            glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
+            glFramebufferTexture2D(GL_FRAMEBUFFER, AttachmentPointFor(internalFormat), GL_TEXTURE_2D,
+                                   source.depthTexture, 0);
+            return source;
+        }
+
+        // The same, with the depth/stencil in a RENDERBUFFER - which cannot be sampled at
+        // all, so the readback has no choice but to copy it somewhere samplable first.
+        // `samples` > 0 makes it multisample, which additionally needs a resolve.
+        DepthSource MakeRenderbufferSource(GLenum internalFormat, int samples) {
+            DepthSource source;
+            glGenFramebuffers(1, &source.fbo);
+            glBindFramebuffer(GL_FRAMEBUFFER, source.fbo);
+            glGenRenderbuffers(1, &source.colorRenderbuffer);
+            glBindRenderbuffer(GL_RENDERBUFFER, source.colorRenderbuffer);
+            if (samples > 0) {
+                glRenderbufferStorageMultisample(GL_RENDERBUFFER, samples, GL_RGBA8, kWidth, kHeight);
+            } else {
+                glRenderbufferStorage(GL_RENDERBUFFER, GL_RGBA8, kWidth, kHeight);
+            }
+            glFramebufferRenderbuffer(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_RENDERBUFFER, source.colorRenderbuffer);
+            glGenRenderbuffers(1, &source.depthRenderbuffer);
+            glBindRenderbuffer(GL_RENDERBUFFER, source.depthRenderbuffer);
+            if (samples > 0) {
+                glRenderbufferStorageMultisample(GL_RENDERBUFFER, samples, internalFormat, kWidth, kHeight);
+            } else {
+                glRenderbufferStorage(GL_RENDERBUFFER, internalFormat, kWidth, kHeight);
+            }
+            glFramebufferRenderbuffer(GL_FRAMEBUFFER, AttachmentPointFor(internalFormat), GL_RENDERBUFFER,
+                                      source.depthRenderbuffer);
+            glBindRenderbuffer(GL_RENDERBUFFER, 0);
+            return source;
+        }
+
+        // Clears the bound framebuffer's depth and stencil to known values, with the masks
+        // and the scissor explicitly out of the way (a leaked scissor from an earlier
+        // scenario would clip the clear and every assertion after it).
+        void ClearDepthStencil(GLenum internalFormat, float depth, int stencil) {
+            glDisable(GL_SCISSOR_TEST);
+            glViewport(0, 0, kWidth, kHeight);
+            GLbitfield mask = 0;
+            if (FormatHasDepth(internalFormat)) {
+                glDepthMask(GL_TRUE);
+                glClearDepth(depth);
+                mask |= GL_DEPTH_BUFFER_BIT;
+            }
+            if (FormatHasStencil(internalFormat)) {
+                glStencilMask(0xFFu);
+                glClearStencil(stencil);
+                mask |= GL_STENCIL_BUFFER_BIT;
+            }
+            glClear(mask);
+        }
+
+        class DepthStencilReadbackMatrixScenario : public ScenarioTest {
+        protected:
+            // Not every ES driver can render to every depth format (DEPTH_COMPONENT32F and
+            // the multisample counts in particular), and an incomplete framebuffer would
+            // turn a legitimate "this machine cannot host the source" into a spurious
+            // failure about the readback.
+            static bool SourceIsUsable() {
+                return glCheckFramebufferStatus(GL_FRAMEBUFFER) == GLenum(GL_FRAMEBUFFER_COMPLETE);
+            }
+
+            static std::vector<float> ReadDepthFloat(int x, int y, int width, int height) {
+                std::vector<float> depth(static_cast<size_t>(width) * height, kDepthPoison);
+                glReadPixels(x, y, width, height, GL_DEPTH_COMPONENT, GL_FLOAT, depth.data());
+                return depth;
+            }
+
+            static std::vector<int> ReadStencilInt(int x, int y, int width, int height) {
+                std::vector<int> stencil(static_cast<size_t>(width) * height, kStencilPoison);
+                glReadPixels(x, y, width, height, GL_STENCIL_INDEX, GL_INT, stencil.data());
+                return stencil;
+            }
+
+            // "every value in the region is `expected`" rather than "the middle pixel is":
+            // a staging blit that lands the wrong rectangle, or a conversion pass with a
+            // half-texel offset, still gets the centre right.
+            static void ExpectAllDepth(const std::vector<float>& values, float expected, const char* what) {
+                size_t bad = 0;
+                float worst = expected;
+                for (float value : values) {
+                    if (std::fabs(value - expected) > 1.0f / 4096.0f) {
+                        if (bad == 0) worst = value;
+                        ++bad;
+                    }
+                }
+                EXPECT_EQ(bad, 0u) << what << ": " << bad << " of " << values.size()
+                                   << " depth values differ from " << expected << "; first bad value " << worst
+                                   << (std::fabs(worst - kDepthPoison) < 1e-6f
+                                           ? " - which is the poison value, so nothing was written at all"
+                                           : "");
+            }
+
+            static void ExpectAllStencil(const std::vector<int>& values, int expected, const char* what) {
+                size_t bad = 0;
+                int worst = expected;
+                for (int value : values) {
+                    if (value != expected) {
+                        if (bad == 0) worst = value;
+                        ++bad;
+                    }
+                }
+                EXPECT_EQ(bad, 0u) << what << ": " << bad << " of " << values.size()
+                                   << " stencil values differ from " << expected << "; first bad value " << worst
+                                   << (worst == kStencilPoison
+                                           ? " - which is the poison value, so nothing was written at all"
+                                           : "");
+            }
+        };
+
+        // ---- glReadPixels across the source kinds -----------------------------------
+
+        struct SourceCase {
+            const char* name;
+            GLenum internalFormat;
+            int samples;
+            bool renderbuffer;
+        };
+
+        const SourceCase kSourceCases[] = {
+            {"texture depth24_stencil8", GL_DEPTH24_STENCIL8, 0, false},
+            {"texture depth32f_stencil8", GL_DEPTH32F_STENCIL8, 0, false},
+            {"texture depth_component16", GL_DEPTH_COMPONENT16, 0, false},
+            {"texture depth_component24", GL_DEPTH_COMPONENT24, 0, false},
+            {"texture depth_component32f", GL_DEPTH_COMPONENT32F, 0, false},
+            {"renderbuffer depth24_stencil8", GL_DEPTH24_STENCIL8, 0, true},
+            {"renderbuffer depth_component24", GL_DEPTH_COMPONENT24, 0, true},
+            {"renderbuffer stencil_index8", GL_STENCIL_INDEX8, 0, true},
+        };
+
+    } // namespace
+
+    TEST_F(DepthStencilReadbackMatrixScenario, EverySourceKindReadsItsClearBack) {
+        if (!Ready()) return;
+        int exercised = 0;
+        for (const SourceCase& testCase : kSourceCases) {
+            SCOPED_TRACE(testCase.name);
+            DepthSource source = testCase.renderbuffer
+                                     ? MakeRenderbufferSource(testCase.internalFormat, testCase.samples)
+                                     : MakeTextureSource(testCase.internalFormat);
+            if (!SourceIsUsable()) {
+                DestroySource(source);
+                continue;
+            }
+            FirstGLError(); // the storage calls above may have probed an unsupported combination
+            ClearDepthStencil(testCase.internalFormat, 0.625f, 9);
+            EXPECT_EQ(FirstGLError(), 0u) << "clearing the source";
+
+            if (FormatHasDepth(testCase.internalFormat)) {
+                const std::vector<float> depth = ReadDepthFloat(0, 0, kWidth, kHeight);
+                EXPECT_EQ(FirstGLError(), 0u) << "glReadPixels(GL_DEPTH_COMPONENT, GL_FLOAT)";
+                ExpectAllDepth(depth, 0.625f, testCase.name);
+            }
+            if (FormatHasStencil(testCase.internalFormat)) {
+                const std::vector<int> stencil = ReadStencilInt(0, 0, kWidth, kHeight);
+                EXPECT_EQ(FirstGLError(), 0u) << "glReadPixels(GL_STENCIL_INDEX, GL_INT)";
+                ExpectAllStencil(stencil, 9, testCase.name);
+            }
+            ++exercised;
+            DestroySource(source);
+        }
+        // A machine that hosted none of the sources would report a vacuous pass.
+        EXPECT_GE(exercised, 4) << "too few depth/stencil source kinds were usable to call this a matrix";
+        glBindFramebuffer(GL_FRAMEBUFFER, 0);
+        Gl().EndFrame();
+    }
+
+    // Depth and stencil in two SEPARATE objects, with two different formats, on the same
+    // framebuffer. Legal GL, and the shape KHR-GL3x.framebuffer_blit builds when its depth
+    // config and its stencil config are configured independently - so a readback that
+    // describes "the" depth/stencil source as one thing serves whichever aspect it happened
+    // to find first and silently abandons the other. Each aspect has to be staged from its
+    // own attachment, in its own format.
+    TEST_F(DepthStencilReadbackMatrixScenario, SeparateDepthAndStencilAttachmentsAreBothReadable) {
+        if (!Ready()) return;
+        DepthSource source;
+        glGenFramebuffers(1, &source.fbo);
+        glBindFramebuffer(GL_FRAMEBUFFER, source.fbo);
+        glGenRenderbuffers(1, &source.colorRenderbuffer);
+        glBindRenderbuffer(GL_RENDERBUFFER, source.colorRenderbuffer);
+        glRenderbufferStorage(GL_RENDERBUFFER, GL_RGBA8, kWidth, kHeight);
+        glFramebufferRenderbuffer(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_RENDERBUFFER, source.colorRenderbuffer);
+        // Depth in a DEPTH_COMPONENT24 renderbuffer...
+        glGenRenderbuffers(1, &source.depthRenderbuffer);
+        glBindRenderbuffer(GL_RENDERBUFFER, source.depthRenderbuffer);
+        glRenderbufferStorage(GL_RENDERBUFFER, GL_DEPTH_COMPONENT24, kWidth, kHeight);
+        glFramebufferRenderbuffer(GL_FRAMEBUFFER, GL_DEPTH_ATTACHMENT, GL_RENDERBUFFER, source.depthRenderbuffer);
+        // ...and stencil in a STENCIL_INDEX8 one of its own.
+        GLuint stencilRenderbuffer = 0;
+        glGenRenderbuffers(1, &stencilRenderbuffer);
+        glBindRenderbuffer(GL_RENDERBUFFER, stencilRenderbuffer);
+        glRenderbufferStorage(GL_RENDERBUFFER, GL_STENCIL_INDEX8, kWidth, kHeight);
+        glFramebufferRenderbuffer(GL_FRAMEBUFFER, GL_STENCIL_ATTACHMENT, GL_RENDERBUFFER, stencilRenderbuffer);
+        glBindRenderbuffer(GL_RENDERBUFFER, 0);
+        if (!SourceIsUsable()) {
+            // Separate depth and stencil images are legal GL but many stacks answer
+            // GL_FRAMEBUFFER_UNSUPPORTED for them; say which, so a skip here is a fact about
+            // the driver rather than an unexplained hole in the matrix.
+            const GLenum status = glCheckFramebufferStatus(GL_FRAMEBUFFER);
+            glDeleteRenderbuffers(1, &stencilRenderbuffer);
+            DestroySource(source);
+            GTEST_SKIP() << "this driver cannot host separate DEPTH_COMPONENT24 and STENCIL_INDEX8 attachments: "
+                         << "glCheckFramebufferStatus = 0x" << std::hex << status;
+        }
+        FirstGLError();
+
+        glDisable(GL_SCISSOR_TEST);
+        glViewport(0, 0, kWidth, kHeight);
+        glDepthMask(GL_TRUE);
+        glStencilMask(0xFFu);
+        glClearDepth(0.3125);
+        glClearStencil(17);
+        glClear(GL_DEPTH_BUFFER_BIT | GL_STENCIL_BUFFER_BIT);
+        ASSERT_EQ(FirstGLError(), 0u);
+
+        const std::vector<float> depth = ReadDepthFloat(0, 0, kWidth, kHeight);
+        EXPECT_EQ(FirstGLError(), 0u) << "reading depth from a separately-attached DEPTH_COMPONENT24";
+        ExpectAllDepth(depth, 0.3125f, "separate depth attachment");
+
+        const std::vector<int> stencil = ReadStencilInt(0, 0, kWidth, kHeight);
+        EXPECT_EQ(FirstGLError(), 0u) << "reading stencil from a separately-attached STENCIL_INDEX8";
+        ExpectAllStencil(stencil, 17, "separate stencil attachment");
+
+        glDeleteRenderbuffers(1, &stencilRenderbuffer);
+        DestroySource(source);
+        glBindFramebuffer(GL_FRAMEBUFFER, 0);
+        Gl().EndFrame();
+    }
+
+    // A multisample source is never read directly - glReadPixels on a multisampled
+    // framebuffer is INVALID_OPERATION in GL as much as in ES, and the state layer says so.
+    // The way multisample depth reaches a reader is a resolve blit into a single-sampled
+    // framebuffer, which is then read; that pair is
+    // KHR-GL3x.framebuffer_blit.multisampled_to_singlesampled_blit_depth_config_test, and
+    // the assertion here is that the resolved depth arrives intact rather than as the
+    // destination's own clear value.
+    TEST_F(DepthStencilReadbackMatrixScenario, AResolvedMultisampleDepthReadsBackFromTheDestination) {
+        if (!Ready()) return;
+        DepthSource multisampled = MakeRenderbufferSource(GL_DEPTH24_STENCIL8, 4);
+        if (!SourceIsUsable()) {
+            DestroySource(multisampled);
+            GTEST_SKIP() << "this driver cannot host a 4x multisample DEPTH24_STENCIL8 renderbuffer";
+        }
+        FirstGLError();
+        ClearDepthStencil(GL_DEPTH24_STENCIL8, 0.875f, 0);
+        ASSERT_EQ(FirstGLError(), 0u);
+
+        // The destination starts at a depth the resolve must overwrite everywhere.
+        DepthSource resolved = MakeTextureSource(GL_DEPTH24_STENCIL8);
+        ASSERT_TRUE(SourceIsUsable());
+        ClearDepthStencil(GL_DEPTH24_STENCIL8, 0.125f, 0);
+        ASSERT_EQ(FirstGLError(), 0u);
+
+        glBindFramebuffer(GL_READ_FRAMEBUFFER, multisampled.fbo);
+        glBindFramebuffer(GL_DRAW_FRAMEBUFFER, resolved.fbo);
+        glDisable(GL_SCISSOR_TEST);
+        glBlitFramebuffer(0, 0, kWidth, kHeight, 0, 0, kWidth, kHeight, GL_DEPTH_BUFFER_BIT, GL_NEAREST);
+        EXPECT_EQ(FirstGLError(), 0u) << "resolving a multisample depth buffer into a single-sampled one";
+
+        glBindFramebuffer(GL_FRAMEBUFFER, resolved.fbo);
+        const std::vector<float> depth = ReadDepthFloat(0, 0, kWidth, kHeight);
+        EXPECT_EQ(FirstGLError(), 0u);
+        ExpectAllDepth(depth, 0.875f, "resolved multisample depth");
+
+        DestroySource(resolved);
+        DestroySource(multisampled);
+        glBindFramebuffer(GL_FRAMEBUFFER, 0);
+        Gl().EndFrame();
+    }
+
+    // A read whose rectangle is NOT the whole attachment. The staging copy has to carry
+    // the requested rect (not the origin) and hand back its rows bottom-up, which a
+    // full-extent uniform read is a fixed point of and therefore cannot see.
+    TEST_F(DepthStencilReadbackMatrixScenario, ASubRectangleReadsTheRightBandInTheRightOrder) {
+        if (!Ready()) return;
+        DepthSource source = MakeTextureSource(GL_DEPTH24_STENCIL8);
+        ASSERT_TRUE(SourceIsUsable());
+
+        // Bottom half 0.25, top half 0.75, and the stencil banded the other way round so a
+        // mix-up between the two aspects cannot pass either.
+        glDisable(GL_SCISSOR_TEST);
+        glViewport(0, 0, kWidth, kHeight);
+        glDepthMask(GL_TRUE);
+        glStencilMask(0xFFu);
+        glEnable(GL_SCISSOR_TEST);
+        glScissor(0, 0, kWidth, kHeight / 2);
+        glClearDepth(0.25);
+        glClearStencil(11);
+        glClear(GL_DEPTH_BUFFER_BIT | GL_STENCIL_BUFFER_BIT);
+        glScissor(0, kHeight / 2, kWidth, kHeight - kHeight / 2);
+        glClearDepth(0.75);
+        glClearStencil(22);
+        glClear(GL_DEPTH_BUFFER_BIT | GL_STENCIL_BUFFER_BIT);
+        glDisable(GL_SCISSOR_TEST);
+        ASSERT_EQ(FirstGLError(), 0u);
+
+        // A rect wholly inside the bottom band, offset from the origin in both axes.
+        const int rectWidth = 8;
+        const int rectHeight = 4;
+        const std::vector<float> bottom = ReadDepthFloat(16, 4, rectWidth, rectHeight);
+        EXPECT_EQ(FirstGLError(), 0u);
+        ExpectAllDepth(bottom, 0.25f, "sub-rect inside the bottom depth band");
+        const std::vector<int> bottomStencil = ReadStencilInt(16, 4, rectWidth, rectHeight);
+        EXPECT_EQ(FirstGLError(), 0u);
+        ExpectAllStencil(bottomStencil, 11, "sub-rect inside the bottom stencil band");
+
+        // And one wholly inside the top band. Reading the mirrored row would answer 0.25.
+        const std::vector<float> top = ReadDepthFloat(16, kHeight - 4 - rectHeight, rectWidth, rectHeight);
+        EXPECT_EQ(FirstGLError(), 0u);
+        ExpectAllDepth(top, 0.75f, "sub-rect inside the top depth band");
+
+        // A rect that STRADDLES the boundary pins the row order itself: its first rows must
+        // be the bottom band and its last rows the top one.
+        const int straddleHeight = 8;
+        const std::vector<float> straddle =
+            ReadDepthFloat(16, kHeight / 2 - straddleHeight / 2, rectWidth, straddleHeight);
+        EXPECT_EQ(FirstGLError(), 0u);
+        ASSERT_EQ(straddle.size(), static_cast<size_t>(rectWidth) * straddleHeight);
+        EXPECT_NEAR(straddle[0], 0.25f, 1.0f / 4096.0f)
+            << "the first row of the returned rect must be its BOTTOM row (GL order), which is in the 0.25 band";
+        EXPECT_NEAR(straddle[straddle.size() - 1], 0.75f, 1.0f / 4096.0f)
+            << "the last row of the returned rect must be its TOP row, which is in the 0.75 band";
+
+        DestroySource(source);
+        glBindFramebuffer(GL_FRAMEBUFFER, 0);
+        Gl().EndFrame();
+    }
+
+    // The packed layouts the packed_depth_stencil family reads its gradients with.
+    TEST_F(DepthStencilReadbackMatrixScenario, PackedDepthStencilReadPixelsCarriesBothAspects) {
+        if (!Ready()) return;
+        struct PackedCase {
+            const char* name;
+            GLenum internalFormat;
+            GLenum type;
+        };
+        const PackedCase cases[] = {
+            {"depth24_stencil8 / GL_UNSIGNED_INT_24_8", GL_DEPTH24_STENCIL8, GL_UNSIGNED_INT_24_8},
+            {"depth32f_stencil8 / GL_FLOAT_32_UNSIGNED_INT_24_8_REV", GL_DEPTH32F_STENCIL8,
+             GL_FLOAT_32_UNSIGNED_INT_24_8_REV},
+        };
+        int exercised = 0;
+        for (const PackedCase& testCase : cases) {
+            SCOPED_TRACE(testCase.name);
+            DepthSource source = MakeTextureSource(testCase.internalFormat);
+            if (!SourceIsUsable()) {
+                DestroySource(source);
+                continue;
+            }
+            FirstGLError();
+            ClearDepthStencil(testCase.internalFormat, 0.5f, 3);
+            ASSERT_EQ(FirstGLError(), 0u);
+
+            const size_t pixels = static_cast<size_t>(kWidth) * kHeight;
+            if (testCase.type == GL_UNSIGNED_INT_24_8) {
+                std::vector<unsigned int> packed(pixels, kPacked24_8Poison);
+                glReadPixels(0, 0, kWidth, kHeight, GL_DEPTH_STENCIL, testCase.type, packed.data());
+                EXPECT_EQ(FirstGLError(), 0u);
+                size_t bad = 0;
+                for (unsigned int value : packed) {
+                    const float depth = static_cast<float>(value >> 8) / 16777215.0f;
+                    const int stencil = static_cast<int>(value & 0xFFu);
+                    if (std::fabs(depth - 0.5f) > 0.01f || stencil != 3) ++bad;
+                }
+                EXPECT_EQ(bad, 0u) << testCase.name << ": " << bad << " of " << pixels
+                                   << " packed words carry the wrong depth or stencil (first word 0x" << std::hex
+                                   << packed[0] << std::dec << ")";
+            } else {
+                std::vector<D32fS8> packed(pixels, D32fS8{kDepthPoison, static_cast<unsigned int>(kStencilPoison)});
+                glReadPixels(0, 0, kWidth, kHeight, GL_DEPTH_STENCIL, testCase.type, packed.data());
+                EXPECT_EQ(FirstGLError(), 0u);
+                size_t bad = 0;
+                for (const D32fS8& value : packed) {
+                    if (std::fabs(value.depth - 0.5f) > 0.01f || (value.stencil & 0xFFu) != 3u) ++bad;
+                }
+                EXPECT_EQ(bad, 0u) << testCase.name << ": " << bad << " of " << pixels
+                                   << " packed pairs carry the wrong depth or stencil (first pair depth "
+                                   << packed[0].depth << " stencil " << (packed[0].stencil & 0xFFu) << ")";
+            }
+            ++exercised;
+            DestroySource(source);
+        }
+        EXPECT_GE(exercised, 1) << "neither packed depth/stencil format was renderable";
+        glBindFramebuffer(GL_FRAMEBUFFER, 0);
+        Gl().EndFrame();
+    }
+
+    // glGetTexImage reads a TEXTURE, not the bound framebuffer - a different entry point
+    // that has to reach the same machinery. This is verify_get_tex_image's shape.
+    TEST_F(DepthStencilReadbackMatrixScenario, GetTexImageReadsAPackedDepthStencilTexture) {
+        if (!Ready()) return;
+        DepthSource source = MakeTextureSource(GL_DEPTH24_STENCIL8);
+        ASSERT_TRUE(SourceIsUsable());
+        FirstGLError();
+        ClearDepthStencil(GL_DEPTH24_STENCIL8, 0.375f, 5);
+        ASSERT_EQ(FirstGLError(), 0u);
+
+        // Read it back through the texture, with the framebuffer that owns it unbound so a
+        // path that secretly read the framebuffer instead would answer from somewhere else.
+        glBindFramebuffer(GL_FRAMEBUFFER, 0);
+        glBindTexture(GL_TEXTURE_2D, source.depthTexture);
+        const size_t pixels = static_cast<size_t>(kWidth) * kHeight;
+        std::vector<unsigned int> packed(pixels, kPacked24_8Poison);
+        glGetTexImage(GL_TEXTURE_2D, 0, GL_DEPTH_STENCIL, GL_UNSIGNED_INT_24_8, packed.data());
+        EXPECT_EQ(FirstGLError(), 0u);
+        size_t bad = 0;
+        for (unsigned int value : packed) {
+            const float depth = static_cast<float>(value >> 8) / 16777215.0f;
+            if (std::fabs(depth - 0.375f) > 0.01f || (value & 0xFFu) != 5u) ++bad;
+        }
+        EXPECT_EQ(bad, 0u) << bad << " of " << pixels
+                           << " words from glGetTexImage(GL_DEPTH_STENCIL) are wrong (first word 0x" << std::hex
+                           << packed[0] << std::dec << ")";
+
+        glBindTexture(GL_TEXTURE_2D, 0);
+        DestroySource(source);
+        Gl().EndFrame();
+    }
+
+    // glCopyTexImage2D out of a depth attachment, then read the copy - verify_copy_tex_image.
+    TEST_F(DepthStencilReadbackMatrixScenario, CopyTexImageFromADepthAttachmentSurvivesAReadBack) {
+        if (!Ready()) return;
+        DepthSource source = MakeTextureSource(GL_DEPTH24_STENCIL8);
+        ASSERT_TRUE(SourceIsUsable());
+        FirstGLError();
+        ClearDepthStencil(GL_DEPTH24_STENCIL8, 0.75f, 6);
+        ASSERT_EQ(FirstGLError(), 0u);
+
+        GLuint copy = 0;
+        glGenTextures(1, &copy);
+        glBindTexture(GL_TEXTURE_2D, copy);
+        glTexImage2D(GL_TEXTURE_2D, 0, GL_DEPTH24_STENCIL8, kWidth, kHeight, 0, GL_DEPTH_STENCIL,
+                     GL_UNSIGNED_INT_24_8, nullptr);
+        glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
+        glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
+        glCopyTexImage2D(GL_TEXTURE_2D, 0, GL_DEPTH24_STENCIL8, 0, 0, kWidth, kHeight, 0);
+        EXPECT_EQ(FirstGLError(), 0u) << "glCopyTexImage2D from a depth/stencil attachment";
+
+        glBindFramebuffer(GL_FRAMEBUFFER, 0);
+        const size_t pixels = static_cast<size_t>(kWidth) * kHeight;
+        std::vector<unsigned int> packed(pixels, kPacked24_8Poison);
+        glGetTexImage(GL_TEXTURE_2D, 0, GL_DEPTH_STENCIL, GL_UNSIGNED_INT_24_8, packed.data());
+        EXPECT_EQ(FirstGLError(), 0u);
+        size_t bad = 0;
+        for (unsigned int value : packed) {
+            const float depth = static_cast<float>(value >> 8) / 16777215.0f;
+            if (std::fabs(depth - 0.75f) > 0.01f) ++bad;
+        }
+        EXPECT_EQ(bad, 0u) << bad << " of " << pixels << " copied depth values are wrong (first word 0x" << std::hex
+                           << packed[0] << std::dec << ")";
+
+        glBindTexture(GL_TEXTURE_2D, 0);
+        glDeleteTextures(1, &copy);
+        DestroySource(source);
+        Gl().EndFrame();
+    }
+
+    // The integer client widths, which are a separate conversion each.
+    TEST_F(DepthStencilReadbackMatrixScenario, DepthAndStencilConvertIntoEveryClientWidth) {
+        if (!Ready()) return;
+        DepthSource source = MakeTextureSource(GL_DEPTH24_STENCIL8);
+        ASSERT_TRUE(SourceIsUsable());
+        FirstGLError();
+        ClearDepthStencil(GL_DEPTH24_STENCIL8, 0.5f, 200);
+        ASSERT_EQ(FirstGLError(), 0u);
+
+        const size_t pixels = static_cast<size_t>(kWidth) * kHeight;
+
+        std::vector<unsigned int> depthUint(pixels, 0xDEADBEEFu);
+        glReadPixels(0, 0, kWidth, kHeight, GL_DEPTH_COMPONENT, GL_UNSIGNED_INT, depthUint.data());
+        EXPECT_EQ(FirstGLError(), 0u) << "glReadPixels(GL_DEPTH_COMPONENT, GL_UNSIGNED_INT)";
+        // 0.5 of the full 32-bit range, with room for the source's 24-bit quantisation.
+        EXPECT_NEAR(static_cast<double>(depthUint[0]) / 4294967295.0, 0.5, 0.01)
+            << "GL_UNSIGNED_INT depth came back as " << depthUint[0];
+
+        std::vector<unsigned short> depthUshort(pixels, 0xBEEFu);
+        glReadPixels(0, 0, kWidth, kHeight, GL_DEPTH_COMPONENT, GL_UNSIGNED_SHORT, depthUshort.data());
+        EXPECT_EQ(FirstGLError(), 0u) << "glReadPixels(GL_DEPTH_COMPONENT, GL_UNSIGNED_SHORT)";
+        EXPECT_NEAR(static_cast<double>(depthUshort[0]) / 65535.0, 0.5, 0.01)
+            << "GL_UNSIGNED_SHORT depth came back as " << depthUshort[0];
+
+        // A stencil index is written unconverted into whichever width was asked for, so 200
+        // must survive intact in all of them - it is also large enough that a signed byte
+        // would wrap, which is the point of choosing it.
+        std::vector<unsigned char> stencilByte(pixels, static_cast<unsigned char>(kStencilPoison));
+        glReadPixels(0, 0, kWidth, kHeight, GL_STENCIL_INDEX, GL_UNSIGNED_BYTE, stencilByte.data());
+        EXPECT_EQ(FirstGLError(), 0u) << "glReadPixels(GL_STENCIL_INDEX, GL_UNSIGNED_BYTE)";
+        EXPECT_EQ(static_cast<int>(stencilByte[0]), 200);
+
+        std::vector<int> stencilInt(pixels, kStencilPoison);
+        glReadPixels(0, 0, kWidth, kHeight, GL_STENCIL_INDEX, GL_INT, stencilInt.data());
+        EXPECT_EQ(FirstGLError(), 0u) << "glReadPixels(GL_STENCIL_INDEX, GL_INT)";
+        EXPECT_EQ(stencilInt[0], 200);
+
+        DestroySource(source);
+        glBindFramebuffer(GL_FRAMEBUFFER, 0);
+        Gl().EndFrame();
+    }
+
+    // The PACK pixel-store parameters apply to a depth read exactly as they do to a colour
+    // one, and the gap regions they create must be left alone.
+    TEST_F(DepthStencilReadbackMatrixScenario, DepthReadbackHonoursThePackPixelStoreParameters) {
+        if (!Ready()) return;
+        DepthSource source = MakeTextureSource(GL_DEPTH_COMPONENT24);
+        ASSERT_TRUE(SourceIsUsable());
+        FirstGLError();
+        ClearDepthStencil(GL_DEPTH_COMPONENT24, 0.5f, 0);
+        ASSERT_EQ(FirstGLError(), 0u);
+
+        const int rectWidth = 4;
+        const int rectHeight = 3;
+        const int rowLength = 8;
+        const int skipPixels = 2;
+        const int skipRows = 1;
+        constexpr float kGap = -7.0f;
+        std::vector<float> destination(static_cast<size_t>(rowLength) * (skipRows + rectHeight) + 16, kGap);
+
+        glPixelStorei(GL_PACK_ROW_LENGTH, rowLength);
+        glPixelStorei(GL_PACK_SKIP_PIXELS, skipPixels);
+        glPixelStorei(GL_PACK_SKIP_ROWS, skipRows);
+        glPixelStorei(GL_PACK_ALIGNMENT, 4);
+        glReadPixels(0, 0, rectWidth, rectHeight, GL_DEPTH_COMPONENT, GL_FLOAT, destination.data());
+        const unsigned int readError = FirstGLError();
+        glPixelStorei(GL_PACK_ROW_LENGTH, 0);
+        glPixelStorei(GL_PACK_SKIP_PIXELS, 0);
+        glPixelStorei(GL_PACK_SKIP_ROWS, 0);
+        glPixelStorei(GL_PACK_ALIGNMENT, 4);
+        EXPECT_EQ(readError, 0u);
+
+        size_t written = 0;
+        size_t gapsTouched = 0;
+        for (size_t index = 0; index < destination.size(); ++index) {
+            const long row = static_cast<long>(index) / rowLength - skipRows;
+            const long column = static_cast<long>(index) % rowLength - skipPixels;
+            const bool inRect = row >= 0 && row < rectHeight && column >= 0 && column < rectWidth;
+            if (inRect) {
+                if (std::fabs(destination[index] - 0.5f) <= 1.0f / 4096.0f) ++written;
+            } else if (destination[index] != kGap) {
+                ++gapsTouched;
+            }
+        }
+        EXPECT_EQ(written, static_cast<size_t>(rectWidth) * rectHeight)
+            << "only " << written << " of " << (rectWidth * rectHeight)
+            << " destination pixels landed where GL_PACK_ROW_LENGTH/SKIP_* put them";
+        EXPECT_EQ(gapsTouched, 0u) << gapsTouched << " bytes outside the packed rectangle were overwritten";
+
+        DestroySource(source);
+        glBindFramebuffer(GL_FRAMEBUFFER, 0);
+        Gl().EndFrame();
+    }
+
+    // The readback borrows the application's context for a full-screen pass. Everything it
+    // touches has to come back, or the next draw inherits it - which is how an emulation
+    // that "works" takes the rest of the renderer down with it.
+    TEST_F(DepthStencilReadbackMatrixScenario, ReadbackLeavesNoGLStateBehind) {
+        if (!Ready()) return;
+        DepthSource source = MakeTextureSource(GL_DEPTH24_STENCIL8);
+        ASSERT_TRUE(SourceIsUsable());
+        FirstGLError();
+        ClearDepthStencil(GL_DEPTH24_STENCIL8, 0.5f, 4);
+
+        // A deliberately awkward state: nothing here is what an emulation pass would want,
+        // so anything it forgets to put back shows up below.
+        GLuint scratchTexture = 0;
+        glGenTextures(1, &scratchTexture);
+        glBindTexture(GL_TEXTURE_2D, scratchTexture);
+        glActiveTexture(GL_TEXTURE3);
+        glBindTexture(GL_TEXTURE_2D, scratchTexture);
+        glEnable(GL_SCISSOR_TEST);
+        glScissor(3, 5, 7, 11);
+        glEnable(GL_CULL_FACE);
+        glEnable(GL_BLEND);
+        glEnable(GL_DEPTH_TEST);
+        glDepthFunc(GL_GEQUAL);
+        glDepthMask(GL_FALSE);
+        glEnable(GL_STENCIL_TEST);
+        glStencilFunc(GL_NOTEQUAL, 0x5, 0x0Fu);
+        glStencilOp(GL_INCR, GL_DECR, GL_INVERT);
+        glStencilMask(0x3Cu);
+        glColorMask(GL_FALSE, GL_TRUE, GL_FALSE, GL_TRUE);
+        glViewport(2, 3, 5, 7);
+        ASSERT_EQ(FirstGLError(), 0u);
+
+        const std::vector<float> depth = ReadDepthFloat(0, 0, kWidth, kHeight);
+        const std::vector<int> stencil = ReadStencilInt(0, 0, kWidth, kHeight);
+        EXPECT_EQ(FirstGLError(), 0u);
+        ExpectAllDepth(depth, 0.5f, "state-preservation case depth");
+        ExpectAllStencil(stencil, 4, "state-preservation case stencil");
+
+        GLint viewport[4] = {0, 0, 0, 0};
+        GLint scissorBox[4] = {0, 0, 0, 0};
+        GLboolean colorMask[4] = {GL_TRUE, GL_TRUE, GL_TRUE, GL_TRUE};
+        GLint depthFunc = 0;
+        GLboolean depthMask = GL_TRUE;
+        GLint stencilFunc = 0, stencilRef = 0, stencilValueMask = 0, stencilWriteMask = 0;
+        GLint stencilFail = 0, stencilPassDepthFail = 0, stencilPassDepthPass = 0;
+        GLint activeTexture = 0, boundTexture = 0;
+        glGetIntegerv(GL_VIEWPORT, viewport);
+        glGetIntegerv(GL_SCISSOR_BOX, scissorBox);
+        glGetBooleanv(GL_COLOR_WRITEMASK, colorMask);
+        glGetIntegerv(GL_DEPTH_FUNC, &depthFunc);
+        glGetBooleanv(GL_DEPTH_WRITEMASK, &depthMask);
+        glGetIntegerv(GL_STENCIL_FUNC, &stencilFunc);
+        glGetIntegerv(GL_STENCIL_REF, &stencilRef);
+        glGetIntegerv(GL_STENCIL_VALUE_MASK, &stencilValueMask);
+        glGetIntegerv(GL_STENCIL_WRITEMASK, &stencilWriteMask);
+        glGetIntegerv(GL_STENCIL_FAIL, &stencilFail);
+        glGetIntegerv(GL_STENCIL_PASS_DEPTH_FAIL, &stencilPassDepthFail);
+        glGetIntegerv(GL_STENCIL_PASS_DEPTH_PASS, &stencilPassDepthPass);
+        glGetIntegerv(GL_ACTIVE_TEXTURE, &activeTexture);
+        glGetIntegerv(GL_TEXTURE_BINDING_2D, &boundTexture);
+
+        EXPECT_EQ(viewport[0], 2);
+        EXPECT_EQ(viewport[1], 3);
+        EXPECT_EQ(viewport[2], 5);
+        EXPECT_EQ(viewport[3], 7);
+        EXPECT_EQ(scissorBox[0], 3);
+        EXPECT_EQ(scissorBox[1], 5);
+        EXPECT_EQ(scissorBox[2], 7);
+        EXPECT_EQ(scissorBox[3], 11);
+        EXPECT_EQ(glIsEnabled(GL_SCISSOR_TEST), GLboolean(GL_TRUE));
+        EXPECT_EQ(glIsEnabled(GL_CULL_FACE), GLboolean(GL_TRUE));
+        EXPECT_EQ(glIsEnabled(GL_BLEND), GLboolean(GL_TRUE));
+        EXPECT_EQ(glIsEnabled(GL_DEPTH_TEST), GLboolean(GL_TRUE));
+        EXPECT_EQ(glIsEnabled(GL_STENCIL_TEST), GLboolean(GL_TRUE));
+        EXPECT_EQ(colorMask[0], GLboolean(GL_FALSE));
+        EXPECT_EQ(colorMask[1], GLboolean(GL_TRUE));
+        EXPECT_EQ(colorMask[2], GLboolean(GL_FALSE));
+        EXPECT_EQ(colorMask[3], GLboolean(GL_TRUE));
+        EXPECT_EQ(depthFunc, GLint(GL_GEQUAL));
+        EXPECT_EQ(depthMask, GLboolean(GL_FALSE));
+        EXPECT_EQ(stencilFunc, GLint(GL_NOTEQUAL));
+        EXPECT_EQ(stencilRef, 0x5);
+        EXPECT_EQ(stencilValueMask, 0x0F);
+        EXPECT_EQ(stencilWriteMask, 0x3C);
+        EXPECT_EQ(stencilFail, GLint(GL_INCR));
+        EXPECT_EQ(stencilPassDepthFail, GLint(GL_DECR));
+        EXPECT_EQ(stencilPassDepthPass, GLint(GL_INVERT));
+        EXPECT_EQ(activeTexture, GLint(GL_TEXTURE3));
+        EXPECT_EQ(boundTexture, GLint(scratchTexture))
+            << "the readback left a scratch texture on the application's texture unit";
+        EXPECT_EQ(FirstGLError(), 0u);
+
+        // Put the awkward state back so the next scenario in this process starts clean.
+        glDisable(GL_SCISSOR_TEST);
+        glDisable(GL_CULL_FACE);
+        glDisable(GL_BLEND);
+        glDisable(GL_DEPTH_TEST);
+        glDisable(GL_STENCIL_TEST);
+        glDepthFunc(GL_LESS);
+        glDepthMask(GL_TRUE);
+        glStencilFunc(GL_ALWAYS, 0, 0xFFFFFFFFu);
+        glStencilOp(GL_KEEP, GL_KEEP, GL_KEEP);
+        glStencilMask(0xFFFFFFFFu);
+        glColorMask(GL_TRUE, GL_TRUE, GL_TRUE, GL_TRUE);
+        glBindTexture(GL_TEXTURE_2D, 0);
+        glActiveTexture(GL_TEXTURE0);
+        glDeleteTextures(1, &scratchTexture);
+        DestroySource(source);
+        glBindFramebuffer(GL_FRAMEBUFFER, 0);
+        glViewport(0, 0, Gl().Width(), Gl().Height());
+        Gl().EndFrame();
+    }
+
+} // 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:
+            // Both backends now answer these reads. DirectGLES has no native ES path for
+            // either aspect (GL_NV_read_depth / GL_NV_read_stencil are optional and absent on
+            // both the Adreno device and Mesa's ES), so it stages the attachment into a
+            // scratch depth texture and samples it into a colour target; the assertions below
+            // are the same either way, which is the point.
+            bool BackendReadsDepthStencil() const { return true; }
+
+            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/DoublePrecisionScenario.cpp

@@ -0,0 +1,720 @@
+// MobileGL - MobileGL/MG_IntegrationTest/Scenarios/DoublePrecisionScenario.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 DOUBLES, RUN AT SINGLE PRECISION.
+//
+// No mobile GPU has 64-bit floats. Adreno and Mali both report shaderFloat64 == VK_FALSE, so
+// Magma cannot build a module that declares the Float64 capability, and ESSL has no fp64 type
+// at all, so SPIRV-Cross refuses the module outright on Espryt ("FP64 not supported in ES
+// profile") and the program never reaches the driver. MobileGL therefore narrows every 64-bit
+// float in a shader to 32 bits (ShaderTranspiler::DemoteFloat64Pass) rather than declining the
+// shader: `double` compiles and runs everywhere, at float precision.
+//
+// The narrowing is only half a contract. The other half is the API side: the global UBO is
+// laid out by reflecting the DEMOTED module, so glUniform*d has to store a float where the
+// shader reads a float, glGetUniform*v has to read one back, and a dmat4's columns are now
+// std140-padded like any other matrix's. Every one of those is a byte offset that fails
+// silently - the uniform simply reads as something else - so the cases below set values
+// through the API and have the SHADER report what it saw.
+//
+// What is deliberately NOT asserted: that the values are exact to double precision. They are
+// not, and cannot be. Every expectation here is the float value of the double that was set,
+// which is the whole point.
+
+#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 {
+
+        // Doubles in every shape the demotion has to handle - a scalar, a vector, a matrix
+        // whose column stride changes, an array whose element stride changes - all reported
+        // through one float SSBO so a single readback says which one moved.
+        constexpr const char* kComputeSource = R"(#version 430 core
+layout(local_size_x = 1) in;
+uniform double uScalar;
+uniform dvec3  uVector;
+uniform dmat4  uMatrix;
+uniform double uArray[3];
+layout(std430, binding = 0) buffer Output {
+    float g_out[];
+};
+void main() {
+    g_out[0] = float(uScalar);
+    g_out[1] = float(uVector.x);
+    g_out[2] = float(uVector.y);
+    g_out[3] = float(uVector.z);
+    // Column-major [column][row]. Off-diagonal entries catch a column-stride mistake that a
+    // diagonal-only check reads straight past.
+    g_out[4] = float(uMatrix[0][0]);
+    g_out[5] = float(uMatrix[0][3]);
+    g_out[6] = float(uMatrix[3][0]);
+    g_out[7] = float(uMatrix[3][3]);
+    g_out[8] = float(uArray[0]);
+    g_out[9] = float(uArray[1]);
+    g_out[10] = float(uArray[2]);
+    // Arithmetic on doubles, including an implicit float->double conversion and a literal
+    // with the fp64 suffix: this is what an application actually writes, and it is the part
+    // that has to survive the conversion folding.
+    double accumulated = uScalar * 2.0lf + 1.5;
+    g_out[11] = float(accumulated);
+}
+)";
+
+        constexpr int kOutputSlots = 12;
+
+        class DoublePrecisionScenario : 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<float> zeroes(kOutputSlots, 0.0f);
+                glBufferData(GL_SHADER_STORAGE_BUFFER, kOutputSlots * sizeof(float), 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_shapeOutput != 0) glDeleteBuffers(1, &m_shapeOutput);
+                if (m_shapeProgram != 0) glDeleteProgram(m_shapeProgram);
+                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<float> Dispatch() {
+                glUseProgram(m_program);
+                glDispatchCompute(1, 1, 1);
+                glMemoryBarrier(GL_BUFFER_UPDATE_BARRIER_BIT);
+                std::vector<float> values(kOutputSlots, -1.0f);
+                glBindBuffer(GL_SHADER_STORAGE_BUFFER, m_output);
+                glGetBufferSubData(GL_SHADER_STORAGE_BUFFER, 0, kOutputSlots * sizeof(float), values.data());
+                glBindBuffer(GL_SHADER_STORAGE_BUFFER, 0);
+                glUseProgram(0);
+                return values;
+            }
+
+            unsigned int m_program = 0;
+            unsigned int m_output = 0;
+            unsigned int m_shapeProgram = 0;
+            unsigned int m_shapeOutput = 0;
+            std::string m_buildLog;
+        };
+
+        // Every double-typed uniform shape GLSL has, all thirteen of them, in one program - the
+        // shape of KHR-GL43.compute_shader.fp64-case2. The scalar and the square matrices are
+        // covered by the cases above; what only a set like this reaches is the NON-SQUARE
+        // matrices, whose column stride and total size both change when the demotion turns a
+        // 64-bit column into a 32-bit one, and whose members therefore move every uniform
+        // declared after them.
+        //
+        // The shader reports every component separately rather than one pass/fail flag, because
+        // "the readback is wrong" is not a diagnosis: a wrong column stride, a wrong member
+        // offset and a wrong narrowing all fail the same single comparison, and only the
+        // component map says which.
+        // No #version here on purpose: it is handed over as a separate source string, the way
+        // the CTS case hands it over.
+        constexpr const char* kAllDoubleShapesSource = R"(
+layout(local_size_x = 1) in;
+uniform double  g_0;
+uniform dvec2   g_1;
+uniform dvec3   g_2;
+uniform dvec4   g_3;
+uniform dmat2   g_4;
+uniform dmat2x3 g_5;
+uniform dmat2x4 g_6;
+uniform dmat3x2 g_7;
+uniform dmat3   g_8;
+uniform dmat3x4 g_9;
+uniform dmat4x2 g_10;
+uniform dmat4x3 g_11;
+uniform dmat4   g_12;
+layout(std430, binding = 0) buffer Output {
+    float g_out[];
+};
+void main() {
+    g_out[0] = float(g_0);
+    for (int i = 0; i < 2; ++i) g_out[1 + i] = float(g_1[i]);
+    for (int i = 0; i < 3; ++i) g_out[3 + i] = float(g_2[i]);
+    for (int i = 0; i < 4; ++i) g_out[6 + i] = float(g_3[i]);
+    for (int c = 0; c < 2; ++c) for (int r = 0; r < 2; ++r) g_out[10 + c * 2 + r] = float(g_4[c][r]);
+    for (int c = 0; c < 2; ++c) for (int r = 0; r < 3; ++r) g_out[14 + c * 3 + r] = float(g_5[c][r]);
+    for (int c = 0; c < 2; ++c) for (int r = 0; r < 4; ++r) g_out[20 + c * 4 + r] = float(g_6[c][r]);
+    for (int c = 0; c < 3; ++c) for (int r = 0; r < 2; ++r) g_out[28 + c * 2 + r] = float(g_7[c][r]);
+    for (int c = 0; c < 3; ++c) for (int r = 0; r < 3; ++r) g_out[34 + c * 3 + r] = float(g_8[c][r]);
+    for (int c = 0; c < 3; ++c) for (int r = 0; r < 4; ++r) g_out[43 + c * 4 + r] = float(g_9[c][r]);
+    for (int c = 0; c < 4; ++c) for (int r = 0; r < 2; ++r) g_out[55 + c * 2 + r] = float(g_10[c][r]);
+    for (int c = 0; c < 4; ++c) for (int r = 0; r < 3; ++r) g_out[63 + c * 3 + r] = float(g_11[c][r]);
+    for (int c = 0; c < 4; ++c) for (int r = 0; r < 4; ++r) g_out[75 + c * 4 + r] = float(g_12[c][r]);
+}
+)";
+
+        // The values the CTS case sets, spelled the way it spells them - column-major, and small
+        // enough that every one is exact in a float. Nothing here is a precision question; a
+        // component that comes back wrong came back from the wrong bytes.
+        constexpr double kG0 = 1.0;
+        constexpr double kG1[2] = {2.0, 3.0};
+        constexpr double kG2[3] = {4.0, 5.0, 6.0};
+        constexpr double kG3[4] = {7.0, 8.0, 9.0, 10.0};
+        constexpr double kG4[4] = {11.0, 12.0, 13.0, 14.0};
+        constexpr double kG5[6] = {15.0, 16.0, 17.0, 18.0, 19.0, 20.0};
+        constexpr double kG6[8] = {21.0, 22.0, 23.0, 24.0, 25.0, 26.0, 27.0, 28.0};
+        constexpr double kG7[6] = {29.0, 30.0, 31.0, 32.0, 33.0, 34.0};
+        constexpr double kG8[9] = {35.0, 36.0, 37.0, 38.0, 39.0, 40.0, 41.0, 42.0, 43.0};
+        constexpr double kG9[12] = {44.0, 45.0, 46.0, 47.0, 48.0, 49.0, 50.0, 51.0, 52.0, 53.0, 54.0, 55.0};
+        constexpr double kG10[8] = {56.0, 57.0, 58.0, 59.0, 60.0, 61.0, 62.0, 63.0};
+        constexpr double kG11[12] = {63.0, 64.0, 65.0, 66.0, 67.0, 68.0, 69.0, 70.0, 71.0, 27.0, 73.0, 74.0};
+        constexpr double kG12[16] = {75.0, 76.0, 77.0, 78.0, 79.0, 80.0, 81.0, 82.0,
+                                     83.0, 84.0, 85.0, 86.0, 87.0, 88.0, 89.0, 90.0};
+
+        struct DoubleShape {
+            const char* name;
+            int base;
+            int columns; // 1 for the scalar and the vectors
+            int rows;    // component count for the scalar and the vectors
+            const double* values;
+        };
+
+        constexpr DoubleShape kDoubleShapes[] = {
+            {"g_0 double", 0, 1, 1, &kG0},      {"g_1 dvec2", 1, 1, 2, kG1},
+            {"g_2 dvec3", 3, 1, 3, kG2},        {"g_3 dvec4", 6, 1, 4, kG3},
+            {"g_4 dmat2", 10, 2, 2, kG4},       {"g_5 dmat2x3", 14, 2, 3, kG5},
+            {"g_6 dmat2x4", 20, 2, 4, kG6},     {"g_7 dmat3x2", 28, 3, 2, kG7},
+            {"g_8 dmat3", 34, 3, 3, kG8},       {"g_9 dmat3x4", 43, 3, 4, kG9},
+            {"g_10 dmat4x2", 55, 4, 2, kG10},   {"g_11 dmat4x3", 63, 4, 3, kG11},
+            {"g_12 dmat4", 75, 4, 4, kG12},
+        };
+
+        constexpr int kAllShapeSlots = 91;
+
+        // The conformance case's own shader, kept verbatim down to the literal suffixes and the
+        // unnamed, unqualified storage block - except that each comparison sets its OWN bit
+        // instead of collapsing all thirteen into one flag. That single flag is the whole reason
+        // the case was unexplained for a wave: it says "something is wrong" and nothing else.
+        //
+        // Verbatim matters here. Reading the components out one at a time (the case above)
+        // passes; whatever fails does so through the shape the conformance case actually
+        // writes - whole-matrix comparison against a constructor, a storage block with no
+        // layout qualifier and no instance name, values reached with constant indices.
+        constexpr const char* kCtsShapedSource = R"(
+layout(local_size_x = 1) in;
+buffer Result {
+  int g_result;
+};
+uniform double g_0;
+uniform dvec2 g_1;
+uniform dvec3 g_2;
+uniform dvec4 g_3;
+uniform dmat2 g_4;
+uniform dmat2x3 g_5;
+uniform dmat2x4 g_6;
+uniform dmat3x2 g_7;
+uniform dmat3 g_8;
+uniform dmat3x4 g_9;
+uniform dmat4x2 g_10;
+uniform dmat4x3 g_11;
+uniform dmat4 g_12;
+
+void main() {
+  g_result = 0;
+
+  if (g_0 != 1.0LF) g_result |= 1;
+  if (g_1 != dvec2(2.0LF, 3.0LF)) g_result |= 2;
+  if (g_2 != dvec3(4.0LF, 5.0LF, 6.0LF)) g_result |= 4;
+  if (g_3 != dvec4(7.0LF, 8.0LF, 9.0LF, 10.0LF)) g_result |= 8;
+
+  if (g_4 != dmat2(11.0LF, 12.0LF, 13.0LF, 14.0LF)) g_result |= 16;
+  if (g_5 != dmat2x3(15.0LF, 16.0LF, 17.0LF, 18.0LF, 19.0LF, 20.0LF)) g_result |= 32;
+  if (g_6 != dmat2x4(21.0LF, 22.0LF, 23.0LF, 24.0LF, 25.0LF, 26.0LF, 27.0LF, 28.0LF)) g_result |= 64;
+
+  if (g_7 != dmat3x2(29.0LF, 30.0LF, 31.0LF, 32.0LF, 33.0LF, 34.0LF)) g_result |= 128;
+  if (g_8 != dmat3(35.0LF, 36.0LF, 37.0LF, 38.0LF, 39.0LF, 40.0LF, 41.0LF, 42.0LF, 43.0LF)) g_result |= 256;
+  if (g_9 != dmat3x4(44.0LF, 45.0LF, 46.0LF, 47.0LF, 48.0LF, 49.0LF, 50.0LF, 51.0LF, 52.0LF, 53.0LF, 54.0LF, 55.0LF)) g_result |= 512;
+
+  if (g_10 != dmat4x2(56.0, 57.0, 58.0, 59.0, 60.0, 61.0, 62.0, 63.0)) g_result |= 1024;
+  if (g_11 != dmat4x3(63.0, 64.0, 65.0, 66.0, 67.0, 68.0, 69.0, 70.0, 71.0, 27.0, 73, 74.0)) g_result |= 2048;
+  if (g_12 != dmat4(75.0, 76.0, 77.0, 78.0, 79.0, 80.0, 81.0, 82.0, 83.0, 84.0, 85.0, 86.0, 87.0, 88.0, 89.0, 90.0)) g_result |= 4096;
+}
+)";
+
+        TEST_F(DoublePrecisionScenario, ADoubleUniformReachesTheShaderAtFloatPrecision) {
+            if (!Ready()) return;
+            glUseProgram(m_program);
+            const GLint scalar = glGetUniformLocation(m_program, "uScalar");
+            ASSERT_GE(scalar, 0);
+            // 0.1 has no exact float (or double) representation, so this only passes if the
+            // value really travelled through the demoted slot rather than being read out of
+            // some other four bytes.
+            glUniform1d(scalar, 0.1);
+            glUseProgram(0);
+
+            const std::vector<float> values = Dispatch();
+            EXPECT_EQ(glGetError(), static_cast<GLenum>(GL_NO_ERROR));
+            EXPECT_FLOAT_EQ(values[0], static_cast<float>(0.1));
+            EXPECT_FLOAT_EQ(values[11], static_cast<float>(static_cast<float>(0.1) * 2.0f + 1.5f))
+                << "arithmetic on the demoted value, including the folded fp64 literal";
+        }
+
+        TEST_F(DoublePrecisionScenario, EveryDoubleShapeLandsInItsOwnSlot) {
+            if (!Ready()) return;
+            glUseProgram(m_program);
+            const GLint scalar = glGetUniformLocation(m_program, "uScalar");
+            const GLint vector = glGetUniformLocation(m_program, "uVector");
+            const GLint matrix = glGetUniformLocation(m_program, "uMatrix");
+            const GLint array0 = glGetUniformLocation(m_program, "uArray[0]");
+            const GLint array2 = glGetUniformLocation(m_program, "uArray[2]");
+            ASSERT_GE(scalar, 0);
+            ASSERT_GE(vector, 0);
+            ASSERT_GE(matrix, 0);
+            ASSERT_GE(array0, 0);
+            ASSERT_GE(array2, 0);
+
+            glUniform1d(scalar, 5.0);
+            const GLdouble vectorValue[3] = {11.0, 12.0, 13.0};
+            glUniform3dv(vector, 1, vectorValue);
+            // Column-major, and every entry distinct so a transposed or mis-strided write
+            // cannot land on a value that happens to match.
+            GLdouble matrixValue[16] = {};
+            for (int i = 0; i < 16; ++i) matrixValue[i] = 100.0 + i;
+            glUniformMatrix4dv(matrix, 1, GL_FALSE, matrixValue);
+            const GLdouble arrayValue[3] = {71.0, 72.0, 73.0};
+            glUniform1dv(array0, 3, arrayValue);
+            glUseProgram(0);
+
+            const std::vector<float> values = Dispatch();
+            EXPECT_EQ(glGetError(), static_cast<GLenum>(GL_NO_ERROR));
+            EXPECT_FLOAT_EQ(values[0], 5.0f) << "scalar double";
+            EXPECT_FLOAT_EQ(values[1], 11.0f) << "dvec3 .x";
+            EXPECT_FLOAT_EQ(values[2], 12.0f) << "dvec3 .y";
+            EXPECT_FLOAT_EQ(values[3], 13.0f) << "dvec3 .z";
+            EXPECT_FLOAT_EQ(values[4], 100.0f) << "dmat4 [0][0]";
+            EXPECT_FLOAT_EQ(values[5], 103.0f) << "dmat4 [0][3] - within the first column";
+            EXPECT_FLOAT_EQ(values[6], 112.0f) << "dmat4 [3][0] - column stride";
+            EXPECT_FLOAT_EQ(values[7], 115.0f) << "dmat4 [3][3]";
+            EXPECT_FLOAT_EQ(values[8], 71.0f) << "double array element 0";
+            EXPECT_FLOAT_EQ(values[9], 72.0f) << "double array element 1 - element stride";
+            EXPECT_FLOAT_EQ(values[10], 73.0f) << "double array element 2";
+        }
+
+        TEST_F(DoublePrecisionScenario, TheTransposeFlagStillTransposes) {
+            if (!Ready()) return;
+            glUseProgram(m_program);
+            const GLint matrix = glGetUniformLocation(m_program, "uMatrix");
+            ASSERT_GE(matrix, 0);
+            GLdouble matrixValue[16] = {};
+            for (int i = 0; i < 16; ++i) matrixValue[i] = 100.0 + i;
+            glUniformMatrix4dv(matrix, 1, GL_TRUE, matrixValue);
+            glUseProgram(0);
+
+            const std::vector<float> values = Dispatch();
+            EXPECT_EQ(glGetError(), static_cast<GLenum>(GL_NO_ERROR));
+            // Transposed, so [column][row] now reads the source's [row][column].
+            EXPECT_FLOAT_EQ(values[4], 100.0f) << "dmat4 [0][0] is on the diagonal either way";
+            EXPECT_FLOAT_EQ(values[5], 112.0f) << "dmat4 [0][3] after transpose";
+            EXPECT_FLOAT_EQ(values[6], 103.0f) << "dmat4 [3][0] after transpose";
+            EXPECT_FLOAT_EQ(values[7], 115.0f) << "dmat4 [3][3] is on the diagonal either way";
+        }
+
+        TEST_F(DoublePrecisionScenario, TheUniformIsStillReportedAsADouble) {
+            if (!Ready()) return;
+            // The demotion is an implementation detail of how the value is STORED. What the
+            // shader source declared is what the application asked about, so the reflection
+            // keeps answering GL_DOUBLE* - an application that switches on the type and calls
+            // glUniform*d has to keep working, and it is the glUniform*d path that is correct
+            // for these uniforms.
+            struct Expectation {
+                const char* name;
+                GLenum type;
+                GLint size;
+            };
+            const Expectation expectations[] = {
+                {"uScalar", GL_DOUBLE, 1},
+                {"uVector", GL_DOUBLE_VEC3, 1},
+                {"uMatrix", GL_DOUBLE_MAT4, 1},
+                {"uArray[0]", GL_DOUBLE, 3},
+            };
+
+            GLint activeUniforms = 0;
+            glGetProgramiv(m_program, GL_ACTIVE_UNIFORMS, &activeUniforms);
+            ASSERT_GT(activeUniforms, 0);
+
+            for (const Expectation& expectation : expectations) {
+                bool found = false;
+                for (GLint index = 0; index < activeUniforms; ++index) {
+                    char name[128] = {};
+                    GLsizei length = 0;
+                    GLint size = 0;
+                    GLenum type = 0;
+                    glGetActiveUniform(m_program, static_cast<GLuint>(index), sizeof(name) - 1, &length, &size,
+                                       &type, name);
+                    if (std::string(name, static_cast<size_t>(length)) != expectation.name) continue;
+                    found = true;
+                    EXPECT_EQ(type, expectation.type) << expectation.name;
+                    EXPECT_EQ(size, expectation.size) << expectation.name;
+                    break;
+                }
+                EXPECT_TRUE(found) << "glGetActiveUniform never reported " << expectation.name;
+            }
+            EXPECT_EQ(glGetError(), static_cast<GLenum>(GL_NO_ERROR));
+        }
+
+        TEST_F(DoublePrecisionScenario, GetUniformdvReadsBackWhatWasStored) {
+            if (!Ready()) return;
+            glUseProgram(m_program);
+            const GLint scalar = glGetUniformLocation(m_program, "uScalar");
+            const GLint vector = glGetUniformLocation(m_program, "uVector");
+            const GLint matrix = glGetUniformLocation(m_program, "uMatrix");
+            ASSERT_GE(scalar, 0);
+            ASSERT_GE(vector, 0);
+            ASSERT_GE(matrix, 0);
+            glUniform1d(scalar, 0.1);
+            const GLdouble vectorValue[3] = {11.5, 12.5, 13.5};
+            glUniform3dv(vector, 1, vectorValue);
+            GLdouble matrixValue[16] = {};
+            for (int i = 0; i < 16; ++i) matrixValue[i] = 100.0 + i;
+            glUniformMatrix4dv(matrix, 1, GL_FALSE, matrixValue);
+            glUseProgram(0);
+
+            // The readback has to undo exactly what the write did - the same std140 column
+            // padding, the same 4-byte components - or a dmat4 comes back with its columns
+            // shifted and nothing else in the API would say so.
+            GLdouble readScalar = 0.0;
+            glGetUniformdv(m_program, scalar, &readScalar);
+            EXPECT_DOUBLE_EQ(readScalar, static_cast<double>(static_cast<float>(0.1)))
+                << "the value is what a float can hold, not the double that was passed in";
+
+            GLdouble readVector[3] = {};
+            glGetUniformdv(m_program, vector, readVector);
+            EXPECT_DOUBLE_EQ(readVector[0], 11.5);
+            EXPECT_DOUBLE_EQ(readVector[1], 12.5);
+            EXPECT_DOUBLE_EQ(readVector[2], 13.5);
+
+            GLdouble readMatrix[16] = {};
+            glGetUniformdv(m_program, matrix, readMatrix);
+            for (int i = 0; i < 16; ++i) {
+                EXPECT_DOUBLE_EQ(readMatrix[i], 100.0 + i) << "dmat4 component " << i;
+            }
+
+            // The float query sees the same storage through the type it is actually stored as.
+            GLfloat readFloat = 0.0f;
+            glGetUniformfv(m_program, scalar, &readFloat);
+            EXPECT_FLOAT_EQ(readFloat, static_cast<float>(0.1));
+            EXPECT_EQ(glGetError(), static_cast<GLenum>(GL_NO_ERROR));
+        }
+
+        TEST_F(DoublePrecisionScenario, ADoubleUniformKeepsItsDeclaredInitializer) {
+            if (!Ready()) return;
+            // A declared initializer is seeded straight into the uniform shadow at link, and the
+            // seeding used to skip 64-bit floats outright ("no 32-bit shadow encoding") - which
+            // was true before the demotion and silently left every such uniform reading zero.
+            const char* source = R"(#version 430 core
+layout(local_size_x = 1) in;
+uniform double uSeeded = 2.5lf;
+uniform dvec3  uSeededVector = dvec3(4.0lf, 5.0lf, 6.0lf);
+layout(std430, binding = 0) buffer Output {
+    float g_out[];
+};
+void main() {
+    g_out[0] = float(uSeeded);
+    g_out[1] = float(uSeededVector.x);
+    g_out[2] = float(uSeededVector.y);
+    g_out[3] = float(uSeededVector.z);
+}
+)";
+            const GLuint program = CompileComputeProgram(source);
+            ASSERT_NE(program, 0u) << m_buildLog;
+
+            glUseProgram(program);
+            glDispatchCompute(1, 1, 1);
+            glMemoryBarrier(GL_BUFFER_UPDATE_BARRIER_BIT);
+            std::vector<float> values(4, -1.0f);
+            glBindBuffer(GL_SHADER_STORAGE_BUFFER, m_output);
+            glGetBufferSubData(GL_SHADER_STORAGE_BUFFER, 0, 4 * sizeof(float), values.data());
+            glBindBuffer(GL_SHADER_STORAGE_BUFFER, 0);
+            glUseProgram(0);
+            glDeleteProgram(program);
+
+            EXPECT_FLOAT_EQ(values[0], 2.5f) << "scalar double initializer";
+            EXPECT_FLOAT_EQ(values[1], 4.0f) << "dvec3 initializer .x";
+            EXPECT_FLOAT_EQ(values[2], 5.0f) << "dvec3 initializer .y";
+            EXPECT_FLOAT_EQ(values[3], 6.0f) << "dvec3 initializer .z";
+            EXPECT_EQ(glGetError(), static_cast<GLenum>(GL_NO_ERROR));
+        }
+
+        TEST_F(DoublePrecisionScenario, EveryDoubleUniformShapeArrivesWhereTheShaderReadsIt) {
+            if (!Ready()) return;
+            // Built the way the CTS case builds it, because every step of that build has been a
+            // bug here at least once: the source arrives as TWO strings (the version directive
+            // and the body), the shader is attached before it has a source and deleted while
+            // still attached, and the program is linked twice.
+            m_shapeProgram = glCreateProgram();
+            ASSERT_NE(m_shapeProgram, 0u);
+            {
+                const GLuint shader = glCreateShader(GL_COMPUTE_SHADER);
+                glAttachShader(m_shapeProgram, shader);
+                glDeleteShader(shader);
+                const char* const sources[2] = {"#version 430 core\n", kAllDoubleShapesSource};
+                glShaderSource(shader, 2, sources, 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);
+                    FAIL() << "compute shader did not compile: " << log;
+                }
+            }
+            glLinkProgram(m_shapeProgram);
+            {
+                GLint linkedOnce = 0;
+                glGetProgramiv(m_shapeProgram, GL_LINK_STATUS, &linkedOnce);
+                if (linkedOnce == GL_FALSE) {
+                    char log[2048] = {};
+                    glGetProgramInfoLog(m_shapeProgram, sizeof(log) - 1, nullptr, log);
+                    FAIL() << "compute program did not link: " << log;
+                }
+            }
+
+            glGenBuffers(1, &m_shapeOutput);
+            glBindBuffer(GL_SHADER_STORAGE_BUFFER, m_shapeOutput);
+            const std::vector<float> zeroes(kAllShapeSlots, 0.0f);
+            glBufferData(GL_SHADER_STORAGE_BUFFER, kAllShapeSlots * sizeof(float), zeroes.data(), GL_DYNAMIC_DRAW);
+            glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 0, m_shapeOutput);
+            glBindBuffer(GL_SHADER_STORAGE_BUFFER, 0);
+
+            const auto location = [&](const char* name) { return glGetUniformLocation(m_shapeProgram, name); };
+
+            // Pass one sets through glProgramUniform*, pass two through glUniform* after a
+            // re-link - the two entry-point families the CTS case exercises, and two different
+            // routes into the same uniform storage.
+            const auto setWithProgramUniform = [&]() {
+                glProgramUniform1d(m_shapeProgram, location("g_0"), kG0);
+                glProgramUniform2d(m_shapeProgram, location("g_1"), kG1[0], kG1[1]);
+                glProgramUniform3d(m_shapeProgram, location("g_2"), kG2[0], kG2[1], kG2[2]);
+                glProgramUniform4d(m_shapeProgram, location("g_3"), kG3[0], kG3[1], kG3[2], kG3[3]);
+                glProgramUniformMatrix2dv(m_shapeProgram, location("g_4"), 1, GL_FALSE, kG4);
+                glProgramUniformMatrix2x3dv(m_shapeProgram, location("g_5"), 1, GL_FALSE, kG5);
+                glProgramUniformMatrix2x4dv(m_shapeProgram, location("g_6"), 1, GL_FALSE, kG6);
+                glProgramUniformMatrix3x2dv(m_shapeProgram, location("g_7"), 1, GL_FALSE, kG7);
+                glProgramUniformMatrix3dv(m_shapeProgram, location("g_8"), 1, GL_FALSE, kG8);
+                glProgramUniformMatrix3x4dv(m_shapeProgram, location("g_9"), 1, GL_FALSE, kG9);
+                glProgramUniformMatrix4x2dv(m_shapeProgram, location("g_10"), 1, GL_FALSE, kG10);
+                glProgramUniformMatrix4x3dv(m_shapeProgram, location("g_11"), 1, GL_FALSE, kG11);
+                glProgramUniformMatrix4dv(m_shapeProgram, location("g_12"), 1, GL_FALSE, kG12);
+            };
+            // Deliberately does NOT re-issue glUseProgram: the CTS case leaves the program
+            // current across the re-link and writes into it from there, so this is the path
+            // where a re-link has to keep the current program's uniform storage addressable.
+            const auto setWithUniform = [&]() {
+                glUniform1d(location("g_0"), kG0);
+                glUniform2d(location("g_1"), kG1[0], kG1[1]);
+                glUniform3d(location("g_2"), kG2[0], kG2[1], kG2[2]);
+                glUniform4d(location("g_3"), kG3[0], kG3[1], kG3[2], kG3[3]);
+                glUniformMatrix2dv(location("g_4"), 1, GL_FALSE, kG4);
+                glUniformMatrix2x3dv(location("g_5"), 1, GL_FALSE, kG5);
+                glUniformMatrix2x4dv(location("g_6"), 1, GL_FALSE, kG6);
+                glUniformMatrix3x2dv(location("g_7"), 1, GL_FALSE, kG7);
+                glUniformMatrix3dv(location("g_8"), 1, GL_FALSE, kG8);
+                glUniformMatrix3x4dv(location("g_9"), 1, GL_FALSE, kG9);
+                glUniformMatrix4x2dv(location("g_10"), 1, GL_FALSE, kG10);
+                glUniformMatrix4x3dv(location("g_11"), 1, GL_FALSE, kG11);
+                glUniformMatrix4dv(location("g_12"), 1, GL_FALSE, kG12);
+            };
+
+            const auto dispatchAndRead = [&]() {
+                glUseProgram(m_shapeProgram);
+                glDispatchCompute(1, 1, 1);
+                glMemoryBarrier(GL_BUFFER_UPDATE_BARRIER_BIT);
+                std::vector<float> values(kAllShapeSlots, -1.0f);
+                glBindBuffer(GL_SHADER_STORAGE_BUFFER, m_shapeOutput);
+                glGetBufferSubData(GL_SHADER_STORAGE_BUFFER, 0, kAllShapeSlots * sizeof(float), values.data());
+                glBindBuffer(GL_SHADER_STORAGE_BUFFER, 0);
+                // The program stays current on purpose - see setWithUniform.
+                return values;
+            };
+
+            const auto expectEverything = [](const std::vector<float>& values, const char* pass) {
+                for (const DoubleShape& shape : kDoubleShapes) {
+                    for (int c = 0; c < shape.columns; ++c) {
+                        for (int r = 0; r < shape.rows; ++r) {
+                            const int component = c * shape.rows + r;
+                            EXPECT_FLOAT_EQ(values[shape.base + component],
+                                            static_cast<float>(shape.values[component]))
+                                << pass << ": " << shape.name << " column " << c << " row " << r;
+                        }
+                    }
+                }
+            };
+
+            setWithProgramUniform();
+            expectEverything(dispatchAndRead(), "glProgramUniform*");
+
+            // A re-link zeroes every uniform, so pass two proves its own writes rather than
+            // reading pass one's bytes back.
+            glLinkProgram(m_shapeProgram);
+            GLint linked = 0;
+            glGetProgramiv(m_shapeProgram, GL_LINK_STATUS, &linked);
+            ASSERT_EQ(linked, GL_TRUE);
+            glBindBuffer(GL_SHADER_STORAGE_BUFFER, m_shapeOutput);
+            glBufferSubData(GL_SHADER_STORAGE_BUFFER, 0, kAllShapeSlots * sizeof(float), zeroes.data());
+            glBindBuffer(GL_SHADER_STORAGE_BUFFER, 0);
+
+            setWithUniform();
+            expectEverything(dispatchAndRead(), "glUniform* after re-link");
+            EXPECT_EQ(glGetError(), static_cast<GLenum>(GL_NO_ERROR));
+        }
+
+        TEST_F(DoublePrecisionScenario, TheConformanceUniformShaderAgreesWithEveryValueItWasGiven) {
+            if (!Ready()) return;
+            m_shapeProgram = glCreateProgram();
+            ASSERT_NE(m_shapeProgram, 0u);
+            {
+                const GLuint shader = glCreateShader(GL_COMPUTE_SHADER);
+                glAttachShader(m_shapeProgram, shader);
+                glDeleteShader(shader);
+                const char* const sources[2] = {"#version 430 core\n", kCtsShapedSource};
+                glShaderSource(shader, 2, sources, 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);
+                    FAIL() << "compute shader did not compile: " << log;
+                }
+            }
+            glLinkProgram(m_shapeProgram);
+            GLint linked = 0;
+            glGetProgramiv(m_shapeProgram, GL_LINK_STATUS, &linked);
+            if (linked == GL_FALSE) {
+                char log[2048] = {};
+                glGetProgramInfoLog(m_shapeProgram, sizeof(log) - 1, nullptr, log);
+                FAIL() << "compute program did not link: " << log;
+            }
+
+            glGenBuffers(1, &m_shapeOutput);
+            const GLint seed = 123;
+            glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 0, m_shapeOutput);
+            glBufferData(GL_SHADER_STORAGE_BUFFER, sizeof(seed), &seed, GL_STATIC_DRAW);
+
+            const auto location = [&](const char* name) { return glGetUniformLocation(m_shapeProgram, name); };
+            glProgramUniform1d(m_shapeProgram, location("g_0"), kG0);
+            glProgramUniform2d(m_shapeProgram, location("g_1"), kG1[0], kG1[1]);
+            glProgramUniform3d(m_shapeProgram, location("g_2"), kG2[0], kG2[1], kG2[2]);
+            glProgramUniform4d(m_shapeProgram, location("g_3"), kG3[0], kG3[1], kG3[2], kG3[3]);
+            glProgramUniformMatrix2dv(m_shapeProgram, location("g_4"), 1, GL_FALSE, kG4);
+            glProgramUniformMatrix2x3dv(m_shapeProgram, location("g_5"), 1, GL_FALSE, kG5);
+            glProgramUniformMatrix2x4dv(m_shapeProgram, location("g_6"), 1, GL_FALSE, kG6);
+            glProgramUniformMatrix3x2dv(m_shapeProgram, location("g_7"), 1, GL_FALSE, kG7);
+            glProgramUniformMatrix3dv(m_shapeProgram, location("g_8"), 1, GL_FALSE, kG8);
+            glProgramUniformMatrix3x4dv(m_shapeProgram, location("g_9"), 1, GL_FALSE, kG9);
+            glProgramUniformMatrix4x2dv(m_shapeProgram, location("g_10"), 1, GL_FALSE, kG10);
+            glProgramUniformMatrix4x3dv(m_shapeProgram, location("g_11"), 1, GL_FALSE, kG11);
+            glProgramUniformMatrix4dv(m_shapeProgram, location("g_12"), 1, GL_FALSE, kG12);
+
+            glUseProgram(m_shapeProgram);
+            glDispatchCompute(1, 1, 1);
+            glMemoryBarrier(GL_BUFFER_UPDATE_BARRIER_BIT);
+
+            GLint disagreements = -1;
+            glGetBufferSubData(GL_SHADER_STORAGE_BUFFER, 0, sizeof(disagreements), &disagreements);
+            for (int bit = 0; bit < 13; ++bit) {
+                EXPECT_EQ(disagreements & (1 << bit), 0)
+                    << kDoubleShapes[bit].name << " did not compare equal to the value it was given";
+            }
+            EXPECT_EQ(glGetError(), static_cast<GLenum>(GL_NO_ERROR));
+        }
+
+        TEST_F(DoublePrecisionScenario, TheFp64ExtensionIsNotAdvertised) {
+            if (!Ready()) return;
+            // The shader above compiled, linked and ran without the extension string, which is
+            // the point: an application does not need GL_ARB_gpu_shader_fp64 advertised to USE
+            // doubles here. What the string additionally promises is 64-bit precision, and that
+            // is the one thing the demotion cannot deliver - so it stays off unless
+            // MOBILEGL_ADVERTISE_FP64 asks for it, and an application that branches on the
+            // string keeps taking its float path.
+            GLint extensionCount = 0;
+            glGetIntegerv(GL_NUM_EXTENSIONS, &extensionCount);
+            ASSERT_GT(extensionCount, 0);
+            bool advertised = false;
+            for (GLint i = 0; i < extensionCount; ++i) {
+                const char* name = reinterpret_cast<const char*>(glGetStringi(GL_EXTENSIONS, static_cast<GLuint>(i)));
+                if (name != nullptr && std::string(name) == "GL_ARB_gpu_shader_fp64") advertised = true;
+            }
+            EXPECT_FALSE(advertised);
+            EXPECT_EQ(glGetError(), static_cast<GLenum>(GL_NO_ERROR));
+        }
+
+        TEST_F(DoublePrecisionScenario, A64BitVertexFormatIsDeclinedOnEveryBackend) {
+            if (!Ready()) return;
+            // The demotion leaves no 64-bit shader input to feed, so there is nothing a 64-bit
+            // vertex FETCH could be fetched into - on either backend, and no longer only on the
+            // ones whose device lacks shaderFloat64. Declined loudly rather than accepted and
+            // drawn as garbage; the matching POST row says the same thing at startup.
+            GLuint vao = 0;
+            glGenVertexArrays(1, &vao);
+            glBindVertexArray(vao);
+            while (glGetError() != GL_NO_ERROR) {}
+
+            glVertexAttribLFormat(0, 3, GL_DOUBLE, 0);
+            EXPECT_EQ(glGetError(), static_cast<GLenum>(GL_INVALID_OPERATION));
+
+            glBindVertexArray(0);
+            glDeleteVertexArrays(1, &vao);
+            while (glGetError() != GL_NO_ERROR) {}
+        }
+
+    } // namespace
+} // namespace MGITest

MobileGL/MG_IntegrationTest/Scenarios/DrawParametersScenario.cpp

@@ -0,0 +1,348 @@
+// MobileGL - MobileGL/MG_IntegrationTest/Scenarios/DrawParametersScenario.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
+//
+// gl_BaseVertex / gl_BaseInstance / gl_DrawID (GL_ARB_shader_draw_parameters),
+// read straight out of the shader that a draw command produced.
+//
+// Neither backend has these builtins for free, and each is wrong in its own way
+// when nobody watches:
+//
+//   * DirectVulkan HAS a BaseVertex builtin, but Vulkan's carries the draw's
+//     firstVertex on a NON-INDEXED draw where GL's is defined to be zero ("the
+//     value passed to the baseVertex parameter, or zero for a command with no
+//     such parameter"). Only the indexed meaning of the two agrees. Every
+//     DrawArrays form therefore takes the ZeroBaseVertex program variant.
+//   * DirectGLES has no such builtins at all: ESSL knows none of them, so the
+//     transpiler demotes each one to a uniform the draw paths feed. A uniform
+//     nobody writes keeps whatever the previous draw left in it - which is what
+//     made gl_BaseVertex report a stale base vertex, and what made
+//     gl_BaseInstance read an unbound storage buffer on a plain glDrawArrays.
+//
+// The shader paints the three values, so a draw that carries the wrong ones
+// paints the wrong colour rather than merely disagreeing with an expectation
+// somewhere. The framebuffer is cleared to WHITE and no case expects 255 in any
+// channel, so "the draw did not happen" can never be mistaken for a pass.
+
+#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/glext.h>
+
+namespace MGITest {
+    namespace {
+
+        // #version 450: glslang only declares the ARB builtins from 440 up.
+        //
+        // Each value is painted at 8 units per count, not 1: the errors these builtins
+        // actually have are OFF BY ONE (a sub-draw that never got its own gl_DrawID reads
+        // the previous one's, a base vertex that arrives one command late), and at one unit
+        // per count no readback tolerance can tell those from rounding.
+        //
+        // And biased by two counts, so that ZERO is not the clamp floor. Five of these cases
+        // expect zero, and an unbiased encoding would let every negative value - the shape a
+        // sign or rebase mistake produces - clamp to the same black and pass.
+        constexpr const char* kVertexSource = R"(#version 450 core
+#extension GL_ARB_shader_draw_parameters : require
+layout(location = 0) in vec2 aPos;
+flat out vec3 vParams;
+void main() {
+    vParams = (vec3(gl_BaseVertexARB, gl_BaseInstanceARB, gl_DrawIDARB) * 8.0 + 16.0) / 255.0;
+    gl_Position = vec4(aPos, 0.0, 1.0);
+}
+)";
+
+        constexpr const char* kFragmentSource = R"(#version 450 core
+flat in vec3 vParams;
+out vec4 oColor;
+void main() {
+    oColor = vec4(vParams, 1.0);
+}
+)";
+
+        struct Vertex {
+            float x, y;
+        };
+
+        // 3 dummy vertices, then the left half of the viewport as two triangles,
+        // then the right half. Nothing here is symmetric by accident:
+        //
+        //   * the padding makes a draw that ignores `first` / baseVertex paint a
+        //     degenerate triangle (i.e. nothing) instead of the right picture;
+        //   * the two halves let one multi-draw show TWO different gl_DrawID
+        //     values in one readback.
+        //
+        // Indices 3..14 together cover the whole viewport, which is what the
+        // single-draw cases use.
+        constexpr int kPad = 3;
+        constexpr int kLeftFirst = kPad;      // 3
+        constexpr int kRightFirst = kPad + 6; // 9
+        constexpr int kHalfCount = 6;
+
+        std::vector<Vertex> SceneVertices() {
+            std::vector<Vertex> vertices(static_cast<std::size_t>(kPad), Vertex{0.0f, 0.0f});
+            const float bounds[2][2] = {{-1.0f, 0.0f}, {0.0f, 1.0f}};
+            for (const auto& half : bounds) {
+                const float x0 = half[0];
+                const float x1 = half[1];
+                vertices.push_back({x0, -1.0f});
+                vertices.push_back({x1, -1.0f});
+                vertices.push_back({x1, 1.0f});
+                vertices.push_back({x0, -1.0f});
+                vertices.push_back({x1, 1.0f});
+                vertices.push_back({x0, 1.0f});
+            }
+            return vertices;
+        }
+
+        // GL's DrawArraysIndirectCommand / DrawElementsIndirectCommand, spelled out
+        // so a test can write one without depending on a GL header's struct.
+        struct ArraysCommand {
+            std::uint32_t count, instanceCount, first, baseInstance;
+        };
+        struct ElementsCommand {
+            std::uint32_t count, instanceCount, firstIndex;
+            std::int32_t baseVertex;
+            std::uint32_t baseInstance;
+        };
+
+        class DrawParametersScenario : public ScenarioTest {
+        protected:
+            void SetUp() override {
+                ScenarioTest::SetUp();
+                if (!Ready()) return;
+                std::string error;
+                m_program = CompileProgram(kVertexSource, kFragmentSource, &error);
+                ASSERT_NE(m_program, 0u) << error;
+
+                const std::vector<Vertex> vertices = SceneVertices();
+                glGenVertexArrays(1, &m_vao);
+                glBindVertexArray(m_vao);
+                glGenBuffers(1, &m_vbo);
+                glBindBuffer(GL_ARRAY_BUFFER, m_vbo);
+                glBufferData(GL_ARRAY_BUFFER, static_cast<GLsizeiptr>(vertices.size() * sizeof(Vertex)),
+                             vertices.data(), GL_STATIC_DRAW);
+                glEnableVertexAttribArray(0);
+                glVertexAttribPointer(0, 2, GL_FLOAT, GL_FALSE, sizeof(Vertex), reinterpret_cast<const void*>(0));
+                ASSERT_EQ(FirstGLError(), GLenum(GL_NO_ERROR)) << "scene setup left a GL error behind";
+            }
+
+            void TearDown() override {
+                if (!Ready()) return;
+                for (GLuint* buffer : {&m_ebo, &m_indirect, &m_parameter, &m_vbo}) {
+                    if (*buffer != 0) glDeleteBuffers(1, buffer);
+                    *buffer = 0;
+                }
+                if (m_vao != 0) glDeleteVertexArrays(1, &m_vao);
+                if (m_program != 0) glDeleteProgram(m_program);
+            }
+
+            template <typename T>
+            void FillBuffer(GLuint& name, GLenum target, const std::vector<T>& data) {
+                if (name == 0) glGenBuffers(1, &name);
+                glBindBuffer(target, name);
+                glBufferData(target, static_cast<GLsizeiptr>(data.size() * sizeof(T)), data.data(), GL_STATIC_DRAW);
+            }
+
+            // Clears to white, runs `draw` and reads the frame back.
+            template <typename DrawFn>
+            Image Render(DrawFn&& draw) {
+                BindDefaultFramebuffer();
+                glViewport(0, 0, HeadlessGL::Get().Width(), HeadlessGL::Get().Height());
+                ClearTo(1.0f, 1.0f, 1.0f, 1.0f);
+                glUseProgram(m_program);
+                glBindVertexArray(m_vao);
+                draw();
+                return ReadPixels(HeadlessGL::Get().Width(), HeadlessGL::Get().Height());
+            }
+
+            // The three builtins as the shader saw them, at a point in one half of
+            // the viewport. `half` is 0 for the left half and 1 for the right.
+            struct DrawParams {
+                int baseVertex = -1, baseInstance = -1, drawId = -1;
+            };
+            // Decodes the biased 8-units-per-count encoding back to the integer the
+            // shader saw. Rounding to the nearest step absorbs any UNORM slop; adjacent
+            // values stay eight units apart, so an off-by-one still reads as one, and a
+            // negative value lands below the bias and decodes negative rather than
+            // clamping into a legitimate zero.
+            static DrawParams ParamsAt(const Image& image, int half) {
+                const int x = image.Width() * (1 + 2 * half) / 4;
+                const Rgba8 pixel = image.At(x, image.Height() / 2);
+                const auto decode = [](std::uint8_t channel) {
+                    return (static_cast<int>(channel) - 16 + 4) / 8;
+                };
+                return {decode(pixel.r), decode(pixel.g), decode(pixel.b)};
+            }
+
+            static void ExpectParams(const Image& image, int half, const DrawParams& expected,
+                                     const std::string& what) {
+                const DrawParams actual = ParamsAt(image, half);
+                EXPECT_EQ(actual.baseVertex, expected.baseVertex)
+                    << what << ": gl_BaseVertex (half " << half << ")";
+                EXPECT_EQ(actual.baseInstance, expected.baseInstance)
+                    << what << ": gl_BaseInstance (half " << half << ")";
+                EXPECT_EQ(actual.drawId, expected.drawId) << what << ": gl_DrawID (half " << half << ")";
+            }
+
+            GLuint m_program = 0;
+            GLuint m_vao = 0;
+            GLuint m_vbo = 0;
+            GLuint m_ebo = 0;
+            GLuint m_indirect = 0;
+            GLuint m_parameter = 0;
+        };
+
+        // ---- the non-indexed forms: gl_BaseVertex is zero, `first` or not ----
+
+        // Vulkan's BaseVertex would answer 3 here (the draw's firstVertex); GL's
+        // must answer 0, because glDrawArrays has no baseVertex parameter at all.
+        TEST_F(DrawParametersScenario, DrawArraysReportsAZeroBaseVertexDespiteItsFirst) {
+            if (!Ready()) return;
+            const Image image = Render([&] { glDrawArrays(GL_TRIANGLES, kLeftFirst, 2 * kHalfCount); });
+            EXPECT_EQ(FirstGLError(), GLenum(GL_NO_ERROR));
+            ExpectParams(image, 0, {0, 0, 0}, "glDrawArrays(first=3)");
+            ExpectParams(image, 1, {0, 0, 0}, "glDrawArrays(first=3)");
+        }
+
+        TEST_F(DrawParametersScenario, DrawArraysInstancedBaseInstanceReportsItsBaseInstance) {
+            if (!Ready()) return;
+            const Image image = Render([&] {
+                glDrawArraysInstancedBaseInstance(GL_TRIANGLES, kLeftFirst, 2 * kHalfCount, 1, 5);
+            });
+            EXPECT_EQ(FirstGLError(), GLenum(GL_NO_ERROR));
+            ExpectParams(image, 0, {0, 5, 0}, "glDrawArraysInstancedBaseInstance(baseInstance=5)");
+        }
+
+        // The base instance of one draw must not survive into the next one. This is
+        // the shape that broke on DirectGLES: the emulation uniform is per-program
+        // state, so a draw that never writes it inherits the last writer's value.
+        TEST_F(DrawParametersScenario, APlainDrawAfterABaseInstancedOneSeesZeroAgain) {
+            if (!Ready()) return;
+            const Image image = Render([&] {
+                glDrawArraysInstancedBaseInstance(GL_TRIANGLES, kLeftFirst, 2 * kHalfCount, 1, 7);
+                glDrawArrays(GL_TRIANGLES, kLeftFirst, 2 * kHalfCount);
+            });
+            EXPECT_EQ(FirstGLError(), GLenum(GL_NO_ERROR));
+            ExpectParams(image, 0, {0, 0, 0}, "plain glDrawArrays after a base-instanced draw");
+        }
+
+        // ---- the indexed forms: gl_BaseVertex IS the base vertex ----
+
+        TEST_F(DrawParametersScenario, DrawElementsBaseVertexReportsItsBaseVertex) {
+            if (!Ready()) return;
+            std::vector<std::uint32_t> indices;
+            for (std::uint32_t i = 0; i < 2 * kHalfCount; ++i) indices.push_back(i);
+            FillBuffer(m_ebo, GL_ELEMENT_ARRAY_BUFFER, indices);
+
+            const Image image = Render([&] {
+                glDrawElementsBaseVertex(GL_TRIANGLES, 2 * kHalfCount, GL_UNSIGNED_INT,
+                                         reinterpret_cast<const void*>(0), kLeftFirst);
+            });
+            EXPECT_EQ(FirstGLError(), GLenum(GL_NO_ERROR));
+            ExpectParams(image, 0, {kLeftFirst, 0, 0}, "glDrawElementsBaseVertex(basevertex=3)");
+            ExpectParams(image, 1, {kLeftFirst, 0, 0}, "glDrawElementsBaseVertex(basevertex=3)");
+        }
+
+        // ... and is zero again for the command that has none, including after one
+        // that did: the same leak the base instance has, on the other builtin. The
+        // preceding draw MUST carry a non-zero base vertex or this case proves nothing -
+        // one index run reaches the geometry through the base vertex, the second through
+        // its own indices, so the two draws paint the same picture with different
+        // gl_BaseVertex and only the second one's value survives in the framebuffer.
+        TEST_F(DrawParametersScenario, DrawElementsAfterABaseVertexDrawReportsZeroAgain) {
+            if (!Ready()) return;
+            std::vector<std::uint32_t> indices;
+            for (std::uint32_t i = 0; i < 2 * kHalfCount; ++i) indices.push_back(i);
+            for (std::uint32_t i = 0; i < 2 * kHalfCount; ++i) indices.push_back(i + kLeftFirst);
+            FillBuffer(m_ebo, GL_ELEMENT_ARRAY_BUFFER, indices);
+            const auto rebasedRun = reinterpret_cast<const void*>(2 * kHalfCount * sizeof(std::uint32_t));
+
+            const Image image = Render([&] {
+                glDrawElementsBaseVertex(GL_TRIANGLES, 2 * kHalfCount, GL_UNSIGNED_INT,
+                                         reinterpret_cast<const void*>(0), kLeftFirst);
+                glDrawElements(GL_TRIANGLES, 2 * kHalfCount, GL_UNSIGNED_INT, rebasedRun);
+            });
+            EXPECT_EQ(FirstGLError(), GLenum(GL_NO_ERROR));
+            ExpectParams(image, 0, {0, 0, 0}, "glDrawElements after a base-vertex draw");
+            ExpectParams(image, 1, {0, 0, 0}, "glDrawElements after a base-vertex draw");
+        }
+
+        // ---- the multi-draw forms: one gl_DrawID per sub-draw ----
+
+        TEST_F(DrawParametersScenario, MultiDrawArraysNumbersItsSubDraws) {
+            if (!Ready()) return;
+            const GLint firsts[2] = {kLeftFirst, kRightFirst};
+            const GLsizei counts[2] = {kHalfCount, kHalfCount};
+
+            const Image image = Render([&] { glMultiDrawArrays(GL_TRIANGLES, firsts, counts, 2); });
+            EXPECT_EQ(FirstGLError(), GLenum(GL_NO_ERROR));
+            ExpectParams(image, 0, {0, 0, 0}, "glMultiDrawArrays sub-draw 0");
+            ExpectParams(image, 1, {0, 0, 1}, "glMultiDrawArrays sub-draw 1");
+        }
+
+        // Every field of an indexed indirect command at once: its own gl_DrawID, the
+        // baseVertex word (which the CPU reads out of the command) and the
+        // baseInstance word (which DirectGLES reads through a storage-buffer view of
+        // the very same buffer).
+        TEST_F(DrawParametersScenario, MultiDrawElementsIndirectCarriesEveryCommandsParameters) {
+            if (!Ready()) return;
+            std::vector<std::uint32_t> indices;
+            for (std::uint32_t i = 0; i < kHalfCount; ++i) indices.push_back(i);
+            FillBuffer(m_ebo, GL_ELEMENT_ARRAY_BUFFER, indices);
+
+            const std::vector<ElementsCommand> commands = {
+                {kHalfCount, 1, 0, kLeftFirst, 0},
+                {kHalfCount, 1, 0, kRightFirst, 4},
+            };
+            FillBuffer(m_indirect, GL_DRAW_INDIRECT_BUFFER, commands);
+
+            const Image image = Render([&] {
+                glMultiDrawElementsIndirect(GL_TRIANGLES, GL_UNSIGNED_INT, reinterpret_cast<const void*>(0), 2,
+                                            sizeof(ElementsCommand));
+            });
+            EXPECT_EQ(FirstGLError(), GLenum(GL_NO_ERROR));
+            ExpectParams(image, 0, {kLeftFirst, 0, 0}, "indirect command 0");
+            ExpectParams(image, 1, {kRightFirst, 4, 1}, "indirect command 1");
+        }
+
+        // glMultiDrawArraysIndirectCount was missing from the DirectGLES backend
+        // table entirely, so the frontend answered INVALID_OPERATION for every call
+        // while GL_ARB_indirect_parameters was advertised. The parameter buffer here
+        // holds a count SMALLER than maxdrawcount, so a path that ignores it draws a
+        // third command over the top of the second and changes the right half.
+        TEST_F(DrawParametersScenario, MultiDrawArraysIndirectCountObeysItsParameterBuffer) {
+            if (!Ready()) return;
+            const std::vector<ArraysCommand> commands = {
+                {kHalfCount, 1, kLeftFirst, 0},
+                {kHalfCount, 1, kRightFirst, 6},
+                {kHalfCount, 1, kRightFirst, 9},
+            };
+            FillBuffer(m_indirect, GL_DRAW_INDIRECT_BUFFER, commands);
+            const std::vector<std::uint32_t> parameters = {2};
+            FillBuffer(m_parameter, GL_PARAMETER_BUFFER, parameters);
+
+            const Image image = Render([&] {
+                glMultiDrawArraysIndirectCount(GL_TRIANGLES, reinterpret_cast<const void*>(0), 0, 3,
+                                               sizeof(ArraysCommand));
+            });
+            EXPECT_EQ(FirstGLError(), GLenum(GL_NO_ERROR));
+            ExpectParams(image, 0, {0, 0, 0}, "counted indirect command 0");
+            ExpectParams(image, 1, {0, 6, 1}, "counted indirect command 1");
+        }
+
+    } // namespace
+} // 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