[CI, scripts, tools] (P7): the pull control strips ctest -V's line prefix, retrace-split pins lavapipe, and the retrace summary reads role logs

.github/workflows/test.yml

@@ -2542,7 +2542,7 @@ jobs:
           # silently is guarded here. trace_cases.py now also rejects an unknown manifest key, which
           # was the hole: `"splitt": true` loaded clean and emptied the subset (review N-4).
           if [ "$(printf '%s' "${SPLIT_MATRIX}" | python3 -c 'import json,sys; print(len(json.load(sys.stdin)["include"]))')" -lt 1 ]; then
-            echo "::error::the split retrace subset is EMPTY. `split` is an OPT-OUT as of P6 - a CI case is in this subset unless it carries \"split\": false - so an empty subset means either every case opted out or the CI matrix itself is empty. Either way retrace-split would be skipped with no red, and exit gate E2 names OpenRA."
+            echo "::error::the split retrace subset is EMPTY. 'split' is an OPT-OUT as of P6 - a CI case is in this subset unless it carries \"split\": false - so an empty subset means either every case opted out or the CI matrix itself is empty. Either way retrace-split would be skipped with no red, and exit gate E2 names OpenRA."
             exit 1
           fi
           echo "split-matrix=${SPLIT_MATRIX}" >> "$GITHUB_OUTPUT"
@@ -2979,6 +2979,14 @@ jobs:
     name: retrace split (${{ matrix.backend }} / ${{ matrix.transport }}, ${{ matrix.case }})
     runs-on: ubuntu-latest
     timeout-minutes: 240
+    # THE LAVAPIPE ICD PIN the integration lanes get through -DMOBILEGL_ITEST_VK_ICD. Here it
+    # can be job-wide: add_trace_replay_test's ENVIRONMENT names no VK_ICD_FILENAMES, so nothing
+    # overrides it. mesa-vulkan-drivers installs every Mesa ICD, and an unpinned loader probes all
+    # of them inside the spawned server's lazy Magma bring-up - the MakeCurrent that outlasted the
+    # client's reply bound in runs 35671704873 / 35706183230. ClientSession's cold-start budget
+    # is the fix; this pin keeps the bring-up on the one device a runner has.
+    env:
+      VK_ICD_FILENAMES: /usr/share/vulkan/icd.d/lvp_icd.json
     needs:
       - build-linux-monolith-control
       - build-linux-split

scripts/ci/control_smoke_test.sh

@@ -121,6 +121,10 @@ expect PASSED "run_trace_case.cmake's own sentence, wrapped" -- run_retrace retr
 # search. The control accepted only the marker sentence until B3's fix round 2, and was red on
 # every real pull-library run (measured on the B3 package tree).
 expect PASSED "the no-client-log sentence a pull library gets"  -- run_retrace retrace-evidence-nolog
+# Both sentences again as `ctest -V` really prints them, with "1: " on every line: folded
+# naively that is "never 1: reported resolving it", and the control used to red on it.
+expect PASSED "the marker sentence under ctest -V's N: prefix" -- run_retrace retrace-evidence-prefixed
+expect PASSED "the no-log sentence under ctest -V's N: prefix" -- run_retrace retrace-evidence-nolog-prefixed
 # The pull library replaying green is the failure this control exists to catch.
 expect FAILED "a pull library passed the split retrace"      -- run_retrace retrace-green
 

scripts/ci/retrace_pull_library_control.sh

@@ -155,7 +155,10 @@ fi
 # width: "never reported resolving it" arrives split over two lines with a two-space continuation
 # indent, and a line-oriented grep for the literal finds nothing. That is not hypothetical - it is
 # the shape the stub reproduces in scripts/ci/testdata/stub_ctest.sh.
-if ! tr -s '[:space:]' ' ' < "${out}" | grep -qF "${EVIDENCE}"; then
+# `ctest -V` ALSO PREFIXES EVERY OUTPUT LINE WITH "<test number>: ", continuation lines included,
+# so without stripping it first the folded text reads "never 1: reported resolving it" and the
+# control reds on the very sentence it is looking for (the *-prefixed stub modes).
+if ! sed -E 's/^[0-9]+: //' "${out}" | tr -s '[:space:]' ' ' | grep -qF "${EVIDENCE}"; then
   echo "::error::the split retrace went red (ctest exit ${control_rc}) with the pull library in place, but the failure never says the library did not resolve the transport - neither of run_trace_case.cmake's sentences (\"${EVIDENCE//$'\n'/\" / \"}\") is in the output. A loader failure, a missing fixture, a timeout or an SSIM drop all land here, and none of them establishes that the transport-identity assertion is what caught the pull library. Only 'non-zero ctest' used to be checked (ID-46 finding 8b)."
   exit 1
 fi

scripts/ci/testdata/stub_ctest.sh

@@ -32,6 +32,8 @@
 #                      one match; the run fails one check earlier, with the runner's "wrote no
 #                      <client log> ... no evidence the transport ever resolved" - the shape a
 #                      PULL library produces since P6's per-role log rename
+#   retrace-evidence-prefixed, retrace-evidence-nolog-prefixed
+#                      the two above with `ctest -V`'s "1: " prefix on every line
 #   retrace-green      one match; the run PASSES
 #
 # EXIT GATE E2's DRAW-DROP CONTROL (scripts/ci/retrace_drop_draw_control.sh). The library's own
@@ -307,6 +309,15 @@ case "${mode}" in
     echo "  counted as a split retrace."
     exit 8
     ;;
+  retrace-evidence-prefixed|retrace-evidence-nolog-prefixed)
+    # The two modes above as `ctest -V` really prints them: EVERY line of a test's output carries
+    # a "<test number>: " prefix, continuation lines included, so after whitespace folding the
+    # wrapped sentence reads "never 1: reported resolving it" / "no evidence the 1: transport ever
+    # resolved". A short case directory breaks the sentence at exactly those words; CI's long
+    # paths happened not to (retrace-split, run 35671704873: 5 false reds from this shape).
+    STUB_MODE="${mode%-prefixed}" bash "$0" "$@" | sed 's/^/1: /'
+    exit "${PIPESTATUS[0]}"
+    ;;
   retrace-green)
     echo "100% tests passed, 0 tests failed out of 1"
     exit 0

tools/trace_replay/render_retrace_summary.mjs

@@ -144,9 +144,15 @@ function parseGpuLabel(logText) {
 
 async function collectGroupLabels(files, defaultGroup) {
   const labels = new Map();
+  // A pull library writes one mobilegl.log. A split library writes one log per ROLE
+  // (mobilegl.client.log / mobilegl.server.log) and no mobilegl.log, so the Linux split lanes'
+  // summaries had no GPU label at all; their role logs are read wherever mobilegl.log is absent.
+  const present = new Set(files.map((file) => normalizeSlashes(file)));
   const logFiles = files.filter((file) => {
     const name = path.basename(file).toLowerCase();
-    return name === "mobilegl.log" || name === "logcat.txt";
+    if (name === "mobilegl.log" || name === "logcat.txt") return true;
+    if (name !== "mobilegl.client.log" && name !== "mobilegl.server.log") return false;
+    return !present.has(normalizeSlashes(path.join(path.dirname(file), "mobilegl.log")));
   });
   for (const logFile of logFiles) {
     const group = inferGroup(logFile, defaultGroup);