MCM GPU service: bounded trace retrieval failure

One local production dx_boids sample reproduced delayed live identity reads and measured synchronous fence/shared-message service explicitly. GPU execution versus scheduling remains unknown: Chrome accepted trace start/end, but did not deliver Tracing.tracingComplete within the retained preparer's 10-second flush wait. No stream was recovered, and no runtime was repeated. There is no demonstrated production correction or avoidable-query speedup.

Evidence: scratch/runs/20261107-mcm-gpu-service/; predecessor mcm-renderer-sync-20261107.md, b3463ba3a. The retained preparer and dependencies were copied into this owned directory. Run-only adaptations add explicit native16 copy/conversion draw timestamps and Worker console.timeStamp clock markers. Production module, automatic config, input guards, native16 order and error handling were unchanged. Exactly one Chrome launch, two bound captures, zero input, zero original/box launches.

Identity and coverage

Source SHA: ee103ca0662c47fcbba6e5f95b50b3cddd161062 (source-sha.txt). initial-relevant.patch and new-wininit-boot.patch preserve relevant dirty source. Both actual Worker modules have SHA-256 e41511fe861363fbe1639d80f20cbcdd7240f46937c93a801eaa256ca094c467, 2,182,493 bytes, retained once. Chrome is 151.0.7922.108, V8 15.1.206.16; the launch uses headless WebGL and --enable-unsafe-swiftshader. The early backend read returned an empty target list, so the actual driver is unknown. 172 successful served source responses match local lengths/hashes; 38 media responses reference existing paths and SHA-256 values. No source/media archive.

python3 scratch/runs/20261107-mcm-gpu-service/analyze.py passes: guest 212 rows and renderer 2576 rows, dropped=0, paired begin/end spans, recorder reads before their immutable 60-second expiry. All six delayed identity reads fall within both coverage windows. Every instrumented guest call, replay batch, shared message, synchronous fence and Flip has complete recorded coverage. Other renderer calls retain slow rows (at least 2 ms), errors and successful sync polls, with uncapped counters for completed calls. Fast native16 stores are counted rather than individually retained. Coverage does not include startup, every renderer function, or calls still pending at recorder read. Nested durations are not additive CPU costs.

Measured causal chain

The identity request at 15:18:07.366Z replies at 15:18:12.910Z, 5543.65 ms later. Guest slice47 runs 15:18:07.300–12.907, requesting 100000 blocks. The request overlaps 5530 ms of guest _waitFor, 5528 ms of shared-message service, 1733 ms of synchronous fence, and 4721 ms of retained getError calls. The categories overlap and must not be summed.

At 15:18:07.368, native16 store70083 copies the target and submits its conversion drawArrays before getError70086; both submissions return within the same integer wall millisecond. That query returns NO_ERROR at 07.526, after 158.405 ms. Its retained parent records monotonic copy/draw/query order and replay context. analysis.json retains the other slow parents and queries. This proves a synchronous observation point after command submission, not the amount of avoidable query overhead or preceding GPU execution time.

Later Flip fences run 15:18:09.673–09.688 and 12.877–12.902. Sync polls return CONDITION_SATISFIED at 09.683 and 12.900. The identity reply remains pending through the first completion and arrives 3 integer milliseconds after slice47 returns. Actual GPU completion evidence therefore coexists with delayed Worker evaluation; neither proves a safe change to slice budgeting or error checks.

Across the renderer window: 18,386 native16 stores and exactly 18,386 error queries, all zero, take 26,421.00 and 26,185.89 ms respectively. There are 36,772 drawArrays submissions (25.14 ms), 18,386 copies (15.89 ms), 19 synchronous fences (6279.93 ms), 106 shared-message services (29149.09 ms), 87 replay batches (22861.53 ms), and ten Flip fences (2163.14 ms). Sync polls comprise 518 TIMEOUT_EXPIRED and ten CONDITION_SATISFIED results. Instrumentation and trace overhead are unmeasured; these are sample observations, not comparable performance measurements against the predecessor.

Trace blocker and executable next step

trace-categories.json verifies gpu, GPU service/decoder/device/debug, command queue and other GPU categories were available. The driver selected all available GPU categories plus timeline/user timing. It requested ReturnAsStream, with a 64 MiB write cap checked before writing each chunk. No stream arrived, so retained trace bytes are zero. Worker clocks satisfy timeOrigin + now versus Date.now() to millisecond rounding; marker requests are retained in trace-clock-markers.json. Trace-to-Worker clock mapping is unverified, because the trace containing the markers is missing. Category availability and accepted start/end do not establish trace coverage, trace drop0, service execution or scheduling. The first terminal failure is preserved; there is no assumption about why Chrome did not finish flushing.

The unlaunched next-trace-browser.cjs / next-trace-control.cjs preserves the same probe and identity chain, selects service/decoder/command-queue categories, records all trace notifications and buffer usage, and allows a bounded 60-second completion wait with 15 seconds reserved before the immutable browser deadline. It retains the 64 MiB cap and stops on failure without replay:

CASE=trace-recovery SPANS=1 BROWSER_DEADLINE=$(($(date +%s%3N)+180000)) \
  node scratch/runs/20261107-mcm-gpu-service/next-trace-browser.cjs

This is a separately authorized next sample, not executed here. Require a complete stream, trace loss checks, matching clock markers and complete probe coverage before classifying service scheduling versus execution. If the driver does not expose execution detail, retain unknown. No timeout/yield/budget, freshness, error suppression or native16 ordering correction is justified. Any later correction still requires injected errors before completion, primitive ordering/state restoration regressions, focused tests and locked build.

Checks and terminal state

Run-only mocked controls pass for category availability, trace lifecycle, small stream recovery, one-for-one native calls, return/throw identity, copy/draw/query order, scalar results, native16/replay association, explicit drops with continuing counters, expiry bypass and wrapper restoration. Bound guard controls pass all 16 identity fields, owner/module replacement, stale zero effects, exact ACK, independent UP and sticky UNKNOWN cases. No accepted fixture matrix was repeated. Existing Node async-readback and GPU-batching tests pass. No production code edit or new production build is claimed. Evidence checker reports unrelated Alice and Tunnel/Twist containment errors and historical missing/invalid records; none names this run.

Chrome exit0; exact driver/Chrome PIDs absent, server/debug ports closed, profile absent, guest/renderer wrappers restored. Watcher closed on release. Free /home sampled above 3 GB. No box assigned: ID/state/expiry not applicable; no lifecycle action, coordinator retains lifecycle ownership. This worker releases its paths and stops after the explicit documentation commit and immediate pull/push. No original steering, FPS or parity claim.