MCM trace recovery: transport works, production stream exceeds cap

Chrome's actual tracing start/end/stream lifecycle works on an idle page. The single subsequent production dx_boids sample received a completion notification, but its stream exceeded the 64 MiB cap before EOF. The retained production trace is incomplete. GPU execution versus scheduling remains unknown, and there is no demonstrated production repair or speedup. The sample was not repeated.

Evidence: scratch/runs/20261107-mcm-trace-recovery/. Predecessor: mcm-gpu-service-20261107.md, 0b95dee4a. Only this note is committed; all drivers, controls, logs, traces and patches are run evidence. Production module, automatic configuration, native16 copy/draw/query ordering, error checks and exact input guards remain unchanged. Zero input, original launches, box launches or nested workers.

Actual lifecycle and terminal failure

The retained next-trace-browser/control was adapted into the owned run. One Chrome process first traces about:blank, before navigating to any guest workload. Both phases select timeline/user timing and the required GPU, service, decoder and command-queue categories. There is one immutable browser deadline, 15:27:31.447Z, established before launch; completion waits are at most 60 seconds and stop 15 seconds before that deadline. Stream reads preserve a cleanup reserve. The cumulative stream write cap is 67,108,864 bytes. Notification logging is bounded to 1 MiB, with uncapped notification counters.

Check Idle lifecycle Production sample
Start/end accepted yes yes
Tracing.tracingComplete yes yes
Producer dataLossOccurred false false
Stream EOF / valid whole JSON yes / yes no / no
Retained stream bytes 135,044 66,846,720
Clock markers found idle marker both Worker markers in prefix

The idle marker, parsed nonempty trace and EOF pass the transport gate before the production navigation. Production tracing begins at 15:24:15.213Z. End is accepted at 15:24:57.070Z; completion arrives at 15:25:04.079Z. At 15:25:05.459Z, the next stream chunk would exceed the cap, so the first terminal failure is preserved, the partial stream is closed, and Chrome is closed. Total retained stream bytes are 66,981,764. The production JSON ends inside an event; no tail metadata or EOF was recovered. A false producer loss flag does not make this capped transport complete. The predecessor's missing completion is not explained by this later run's successful notification.

trace-notification-counters.json records 43 notifications, 6046 logged bytes, zero notification drops: two completion notifications and 41 buffer reports. No unbounded inline event logging is used.

Coverage and limits

python3 scratch/runs/20261107-mcm-trace-recovery/analyze.py validates recorder expiry, zero drops, paired retained spans, all instrumented guest/service/fence/ replay calls against uncapped counters, one error query per native16 store, and wrapper restoration. Guest has 204 rows; renderer has 2216. Five delayed identity reads, 3102–7852 ms, fall within both recorder coverage windows. All 13,678 recorded error queries return zero. These are observations under instrumentation, not comparable performance measurements or removable-query evidence. Fast calls are counted; sparse renderer rows retain slow calls, errors and successful polls. Coverage excludes startup, uninstrumented functions and calls pending at read.

The retained prefix contains 400,665 complete event lines. gpu and gpu,toplevel.flow account for 57,208,341 bytes of those lines. Names include Scheduler::RunTask, Scheduler::ScheduleTask, Scheduler::FinishTask and command-buffer service events. Prefix presence does not establish complete service execution coverage, matched scheduling spans or hardware execution time. No execution/scheduling attribution is made from these incomplete events.

Both unique Worker console.timeStamp markers match retained events and Worker sessions/threads. timeOrigin + now differs from Date.now() by less than 1 ms; the two local wall-minus-trace offsets differ by about 0.288 ms. This establishes marker correspondence in the prefix, without verifying clock coverage across a complete trace. Missing service/execution detail remains unknown.

Source SHA is c1b447b7c686b937a8b579c668a54fce7fb03c11. Relevant dirty hunks and new lib/wininit-boot.js are binary patches; owned run scripts and this note also have patches. Both actual Workers identify the production module e41511fe861363fbe1639d80f20cbcdd7240f46937c93a801eaa256ca094c467, 2,182,493 bytes, retained once. Chrome is 151.0.7922.108. The bound production statistics report ANGLE / Intel UHD Graphics 620 / Mesa 23.2.1; the separate early backend probe has empty targets and does not supply that identity. Served responses are verified against existing paths/hashes: 171 source responses and 39 media responses. Media is referenced, not copied. No source or media archives.

Exact next executable diagnostic

The unlaunched next-short-browser.cjs, next-short-control.cjs and next-short-spans.cjs retain the idle lifecycle gate, required GPU categories, markers, sparse probes, cap and deadline. They explicitly exclude unselected categories and reduce the production trace body from two bound captures to one bound chain.readLive(). Keeping gpu preserves the category that actually emitted the observed scheduler/service events. The shorter window is a transport diagnostic, with no promised size, service detail or speedup:

CASE=short SPANS=1 BROWSER_DEADLINE=$(($(date +%s%3N)+210000)) \
  node scratch/runs/20261107-mcm-trace-recovery/next-short-browser.cjs

This command was prepared and mock-tested, not executed here. A separately authorized run must obtain EOF, valid complete JSON, loss indicators, clock markers and complete relevant probe coverage before attributing execution or scheduling. If hardware execution detail is absent, it remains unknown. No timeout, yield, slice-budget, freshness, error-suppression or native16 ordering change is justified. Any later production repair still requires error injection before completion, primitive ordering/state restoration regressions, focused tests and the locked build.

Verification and handoff

Run-only controls pass trace lifecycle/EOF/marker checks, unavailable-category rejection, missing completion, producer-loss rejection, the cumulative cap and stream closure. Sparse-recorder controls pass native call/return/throw transparency, ordering, explicit drops with continuing counters, expiry bypass and restoration. The prepared successor passes syntax and controls proving one bound read, no capture/input and both recorder reads/disposal. Existing Node async-readback and GPU-batching tests pass. No accepted fixture matrix was repeated. No production source change or new locked build is claimed. The global evidence checker reports existing Alice and Tunnel/Twist containment errors and historical missing/invalid records; none names this run. Own artifact hashes and containment are verified separately in evidence-manifest.json.

Chrome exits zero. Driver/Chrome PIDs are absent, ports 41821/37739 closed, profile and writer absent, guest/renderer wrappers restored. The board watcher is stopped at release. Free disk remains above 2 GiB at recorded checks. No box was assigned: ID/state/expiry are not applicable; coordinator lifecycle ownership is unchanged. Evidence and the first failure are retained; stop after this one bounded task, own-file commit and immediate main pull/push.