Redline first reply delay, 1250 — native cause unresolved offline
No production fix is justified. The 1235 hold guard deterministically rejects a keyDown reply that arrives after its independent release has started; it did so correctly. The cause of that native reply delay remains UNKNOWN. No old-failing/new-passing defect in the actual current consumer was found. SingleTaskSTOP; native REVOKED; allwriters UNKNOWN; edictABA UNKNOWN.
Evidence: scratch/runs/20261109-redline-first-reply-cause-1250/. Read the 1235 and successful-fixture 1150 notes and retained artifacts. All experiments here are offline, with inert browser/guest services. No Box API contact, browser, guest, local Chrome, agent, worker, or deployment occurred. Historical evidence was read, never rewritten. Only this note and fresh task evidence paths were owned.
Comparison and causal boundary
comparison.json and pin-delta.json establish:
- All 836 provider paths shared by 1150 and 1235 have identical SHA-256 pins; 1235 adds 49 paths, removes none. The generated caller is byte-identical, SHA-256
326b2a6fa946c3a5c4a3650285278f742c4ae44e5e1e5190fcedbf633ff8edbc. Bound input, ordinary keys, transport and observer references are identical. The eight directly checked production providers, including host, renderer/input handlers, fixture and Wasm, match the retained pins and current bytes. Both native runs report Chrome151.0.7922.108 / JS15.1.206.16. - 1150 comma600: DOWN dispatch0.174ms, reply8.312ms; UP dispatch600.017ms, reply600.753ms, six held samples and native high/release-zero proof. 1235: DOWN dispatch0.162ms, reply1147.669ms; UP dispatch601.110ms, reply1148.158ms, samples empty. Both 1235 replies succeeded. Neither successful CDP replies nor their close arrival times prove physical input consumption or locate the delay.
- 1235 fixture baseline completed eight COM setup results at zero, endpoint1/slot0 and zero comma/period counters. Its admitted startup capture authenticated the page/worker owner before dispatch. Fixture preparation calls the production
browserInput.wireCanvasInput, focuses the canvas and uses the ordinary renderer; there is no fixture key handler substitute in the native route. This establishes setup, not the handler execution time at the failed keyDown.
Actual path: generated execute -> admitted fixture baseline poll -> observeHold -> retained bound.hold -> production send -> existing browser WebSocket -> flattened page session -> native DOM window-capture key handler -> production renderer physical state/worker-owned queue. send installs its pending entry before socket send. The actual receiver resolves that entry by request ID, without waiting for asynchronous target-event work. A flattened browser-socket request ID is allocated globally by this consumer; page and worker sends do not create duplicate IDs. Capture/admission work completes before the hold; the hold does not start sampling until keyDown resolves. The independent 600ms release therefore runs while DOWN is unresolved; the guard then refuses the late DOWN, retains the failed receipt and samples, and prevents a successful capture ACK or subsequent period command.
Retained 1235 artifacts do not include raw CDP request/reply IDs and socket receive times, native key EventDispatch entry/exit, or the relevant page task timeline. Console records stop after fixture DLL setup; the existing mouse-event and guest-consumption traces cannot supply missing keyboard callback timing. A sampler that never started is not evidence of a stalled sampler. No evidence selects wrong routing, slow handler/audio/wake work, native dispatch queueing, or reply transport delay as the cause.
Offline checks
diagnose.cjs executes the actual generated send and exact receiver assignment: independent request order, reverse reply arrival, unknown IDs, protocol refusal, existing20s timeout/late-reply discard, and terminal no-send pass. Its initial DOM run retained the helper's inert character callback; handler-recheck.log / recheck-production-handlers.json repeat all eight cases with actual production character handling as well as actual DOWN/UP handling. Trusted comma/period yield physical high32768 then zero through the worker-owned queue route; exclusive-input, page-dialog and unowned-UP refusal controls pass. This is an inert DOM/service test, not native trust or scheduling proof.
generated-suite.cjs reuses the existing whole production-fixture test in memory with the exact 1235 generated caller. All six cases pass: positive comma/period capture ACK0/1 and ordinary joined quit ACK2, close without protocol ACK, wrong live module refusal, delayed DOWN response, delayed DOWN service and release failure. Actual caller/watchdog waits and separate pipe EOF are checked. The delayed-response control leaves the inert key released; delayed service can apply DOWN after UP and leave it high until cleanup. Both reject the hold with empty samples and no later input/capture. These authored alternatives show why the native receipt cannot identify a cause; neither is a reproduction of a proved native defect.
The existing key-delivery suite also passes all six cases, including historical missing-receipt failure and stalled-DOWN work-deadline refusal. All test output paths were explicitly fresh and task-owned. flock /tmp/wine-build.lock bash tools/build.sh passed. Negative-case stack traces in the logs are expected refusal controls. No timer, request order, routing, hold/CDP/caller bounds, independent UP, sticky first failure, capture ACK, quit or trust policy changed.
Exact smallest missing-boundary experiment — specified, not authorized or run
If separately authorized, run one fixture-only first comma600 command through this same generated caller/setup, with the same caps and independent UP. Retain actual raw request/reply IDs, session IDs and monotonic send/receive times at the existing CDP send/receive boundary. Enable Chrome's built-in input-latency and renderer EventDispatch/task tracing at browser startup, with a task-owned output file and finite10s trace duration (--trace-startup, --trace-startup-duration=10, --trace-startup-file=..., input/latencyInfo/devtools.timeline categories). Retrieve that file only through existing no-input cleanup; no post-terminal Tracing.end or relaxation of the terminal send guard. Do not inject a second DOM input handler, extra sampler, heartbeat, polling loop, or timing/routing change. Retain the trace clock mapping and EventDispatch keydown/keyup timestamps/durations alongside the existing delivery receipt. Verify the trace actually flushed and contains those events; absent events leave the cause UNKNOWN.
This is the minimum missing causal comparison: late EventDispatch entry versus long EventDispatch/task duration versus prompt completion with late socket receipt. Raw frames alone distinguish consumer pending-map handling from receive delay but cannot distinguish renderer queueing from a slow handler. Existing input-profile marks begin after renderer DOWN handling for printable keys, so enabling those alone does not fill the entry boundary. A trace localizing a long callback still needs its stack attribution before naming a handler defect. No production fix should follow merely from reproducing latency.
On first failure preserve sticky terminal state, permit only the already independent release and no-input shutdown/evidence collection; no later guest input/capture, original, retry or guessed workaround. If the command succeeds, retain its ordinary capture ACK and use ordinary joined quit; that success does not explain 1235. An unchanged native retry is not the next action. This experiment is only a specification; current task authorization is OFFLINEONLY and ends here.
git-base.txt, zero relevant production.diff and owned-inputs.patch record recovery; Git remains the source archive. No archive, media or module copy was created. Free/home remained above1.5GB. No new box exists. Prior bx_qh4wn3gk, UUID 88f7621f-cd2c-4a06-b9c2-49bf6c513897, remains UNCONTACTED: retained 1235 receipt says archived/physical drained, unforced stop_204911fc8ca471a8 completed2026-11-09T11:54:35.927Z, initial archiveAfter2026-11-09T12:17:45.374Z, final archiveAfter null, expiresAt not returned. No new stop receipt applies.