MCM renderer synchronization: native16 queries and Flip completion

One fresh local production dx_boids sample reproduces delayed live identity reads and separates two measured renderer waits. Native16 error queries are a synchronous observation point; Flip also waits for outstanding GPU work to complete. Removing error queries is not a demonstrated correction: their elapsed time can include preceding GPU work, and substantial Flip waits remain even in intervals with little error-query time. No production code changed. This is an observation prerequisite for the MCM investigation, not original gameplay, responsiveness, FPS, or parity acceptance.

Evidence: scratch/runs/20261107-mcm-renderer-sync/. The predecessor is mcm-worker-cdp-latency-20261107.md, terminal aae0f01e8. This worker used its executable preparer and driver in the fresh owned directory, then added run-only completion/result instrumentation. There was exactly one Chrome launch, with two captures through the unchanged live identity chain and no input. No original, box, accepted fixture matrix, held retry, private input callback, or guest/config/clock write occurred.

Executed identity and coverage

Source commit: df0bbae0c5af942debbcdd45f5537ea3be97fd13. source-sha.txt, empty initial-relevant.patch, and new-wininit-boot.patch record Git and relevant uncommitted source. The actual guest and owning renderer module recovered through CDP is SHA-256 e41511fe861363fbe1639d80f20cbcdd7240f46937c93a801eaa256ca094c467, 2,182,493 bytes, retained once under sample/. Chrome was 151.0.7922.108, V8 15.1.206.16, headless local WebGL with --enable-unsafe-swiftshader. The actual GL renderer/vendor string was not collected; that flag alone does not establish the executing GPU backend.

All 171 served source-response hashes and lengths matched existing files; 39 successful media-response references retain existing paths, hashes and lengths. Full HTTP records, including other responses, are in sample/served.jsonl. Source and media were not archived or copied into runs. Own new scripts and this note have individual binary patches.

python3 scratch/runs/20261107-mcm-renderer-sync/analyze.py verifies:

This is complete coverage of the instrumented calls within those windows, not every renderer function, browser startup, fast-call timelines, or work still in flight when the recorder was read. In particular, the sample did not wrap synchronous fence() or shared-message service as separate spans; their instrumented native16/GL calls still contribute counters and slow rows. Nested durations below must not be added as CPU costs. Date timestamps have millisecond precision; durations use each Worker's monotonic clock.

A delayed request through renderer and GPU completion

The identity request at 15:05:26.604Z replies at 15:05:31.202Z, latency 4597.86 ms. Guest slice35, requesting 100000 blocks, runs from 15:05:26.392Z to 15:05:31.200Z. The request overlaps 4593 ms of guest _waitFor, 1560 ms of retained renderer getError, and 2371 ms of Flip fence.

Event UTC, 2026-11-07 Evidence
Native16 getError inside replay batch48 15:05:26.699–28.152 1452.78 ms; returns NO_ERROR; exact _storeNative16 parent retained
Batch48 replay completes 15:05:28.313 Sequence48, eight commands
Guest wait for sequence49 returns 15:05:28.426 Identity request remains pending
Batch50 Flip fence 15:05:28.427–29.207 GPU poll returns CONDITION_SATISFIED at 29.205; guest wait51 returns 29.207
Guest wait54 returns 15:05:29.605 Identity request still pending
Batch55 Flip fence 15:05:29.607–31.198 GPU poll returns CONDITION_SATISFIED at 31.186; guest wait56 returns 31.199
Slice35 returns; identity reply received 15:05:31.200; 31.202 No reply at either earlier GPU completion or guest wait return

analysis.json retains all five requests, matching waits/batch sequences, completion results, and the slow native16 queries with their parent calls. Another read, 15:05:19.275Z, waits 4896.99 ms while overlapping only 39 ms of retained getError, but 4432 ms of Flip fences. Therefore this sample does not support treating error-query removal alone as a responsive identity-read fix.

Across the renderer window, 15,632 _storeNative16 calls make exactly 15,632 getError calls, all returning zero. Error queries total 4374.28 ms inside 4489.03 ms of native16 storage. Native GL draw/copy submission calls return quickly: 31,264 drawArrays calls total 13.23 ms and 15,632 copyTexSubImage2D calls total 6.60 ms. Those submission times are not GPU execution times.

Eight Flip fences total 13924.34 ms. Their 3359 polls comprise 3351 TIMEOUT_EXPIRED and eight CONDITION_SATISFIED results. Poll API calls total 45.01 ms; eight buffer collections total 79.28 ms. Most fence wall time lies between issue and GPU completion, including task scheduling. This is actual pending GPU completion evidence, not proof that all of that time is unavoidable GPU execution or a measurement of GPU utilization.

Correction boundary and executable next step

The source path is D3DIMGpu._submitImmediate → one native primitive draw → _storeNative16 → full-target color copy and conversion quad → gl.getError. This preserves native16 high bits after each primitive, including overlapping primitives inside a public draw. fenceFlip separately issues a PBO readback, polls a GPU sync, collects bytes and writes them back before replay publishes completion. The guest's synchronous _waitFor and uninterrupted ex.run retain the pending identity read until the slice returns.

The query synchronization point is measured. Its avoidable fraction is unknown. A zero result does not justify suppressing error handling, moving failure past later draws, or changing native16 storage order. Nor does the trace justify timeout, yield, budget or freshness changes. There is no failing-before/pass-after production correction here, so no source edit or new locked production build is claimed.

The prepared follow-up adds Chrome GPU-service tracing to the same sparse probe, plus explicit synchronous fence and shared-message service spans. It is syntax-checked and has mocked trace lifecycle/category controls, but has not launched a runtime. For a separately authorized next sample:

python3 scratch/runs/20261107-mcm-renderer-sync/prepare-gpu-service-probe.py
CASE=gpu-service-next SPANS=1 BROWSER_DEADLINE=$(($(date +%s%3N)+180000)) \
  node scratch/runs/20261107-mcm-renderer-sync/gpu-service-browser.cjs

It requires available GPU trace categories, records host/Worker clock mapping, retains the service trace with a 64 MiB cap, and stops without replay on failure. Correlate GPU process/service events during a long native16 query with its preceding copy/draw commands and later Flip completion. If the driver exposes no execution/completion detail, retain that limitation rather than subtracting query wall time as a predicted speedup. Any proposed checkpoint correction still needs an error-injection regression proving failure before completion, native16 primitive order and state restoration, then a locked build.

Checks and terminal state

The bound helper is unchanged, SHA-256 9df1455f51f91d76e17890ff4d90afe9264759d3e30e224a9a7569e29b1f8062. Offline controls pass for all 16 required identity fields, missing/replaced owners/modules, stale zero effects, object mismatch closure, UP sent before a pending DOWN acknowledgement, wrong-number ACK timeout, mismatched prior ACK ID, and sticky UNKNOWN after a late ACK. The actual sample sends no native input or numbered transport commands; it does not claim live input acceptance. Recorder controls verify one-for-one native calls, unchanged return/promise/ throw identity, scalar GPU results, native16 parent/batch association, explicit drops, uncapped aggregates, expiry bypass and restoration.

Focused existing Node regressions test/test-d3dim-async-readback.js and test/test-d3dim-gpu-batching.js pass without additional browser launches. ops/check-run-evidence.js reports the pre-existing Alice containment error; none of its errors/missing/invalid entries names this run. Own hashes, script syntax, correlation assertions and documentation whitespace pass.

Chrome exits code0; exact driver/Chrome PIDs are absent, server/debug ports closed, wrappers restored, profile and writer lock absent. The messageboard watcher is closed before release. Free disk is sampled above 3 GB; no continuous disk claim. No box was assigned: ID/state/expiry are not applicable and lifecycle ownership remains with the coordinator. This bounded worker releases its paths and stops after its explicit documentation commit and immediate pull/push.