Q3 startup custody service investigation, 2045

The new discriminating evidence is a real 14,250.230 ms ptrace hold of Chrome child 6731 during the retained request-6 interval. Q3 now batches the custody journal at its existing durable executable-CONT barriers. Controlled slow-storage tests establish the resulting reduction in tracer service time. This does not establish that journal service caused native request 6's deadline, or that the change repairs it. Native validation was not reached.

Code commit on main: 02e95626d (full identity in the run's Git records). Evidence: scratch/runs/20261109-q3-startup-autoattach-cause-2045/. Git plus small patches is the source archive; no source/media archive was made.

Retained evidence and causal control

The independently SHA/size-checked 2035 custody journal registers child 6731 at 20:35:28.096 UTC, about 3.204 seconds after request 6. Its first returned CONT is at 20:35:42.347 UTC, about 17.455 seconds after that request. This is actual tracer evidence of withheld executable progress, distinct from Node's send, scheduler and public callback observations. The log does not identify the child's Chromium role or its dependency on request 6. Inter-record gaps do not directly measure fsync or browser service time.

Exact shared-source review found that Journal.write synchronously flushes gzip, flushes the raw stream and calls fsync. The published supervisor does this for multiple intermediate stop/registration records, while its existing executable-CONT path already flushes the complete ordered journal prefix.

tools/q3-startup-custody.py authenticates the exact published shared provider SHA before one assertion-anchored transform: intermediate records use the existing bounded pending journal. The executable-CONT flush, ownership/lineage, consumed waits, PID identity checks, cancellation, irreversible retirement, sticky service errors and finalization code remain byte identical. Saturation still applies synchronous backpressure. A bounded service-time observer records the tracer's own journal writes; it supplies no wire/browser timing.

Only Q3's generated native lifecycle imports this adapter. Preparation pins the adapter and the generated wrappers. The shared custody file and protected RF2045/Nox2020 providers were not edited. The generated Node consumer and Python command caller are unchanged, retaining c43c85981's real input-wire composition regression and 2b945315a's stdin repair.

The final controlled fork/exec test inserted 12 ms service delay per fsync:

Exact supervisor fsync calls Elapsed Owned/reaped
Published 115 1.731 s 13/13
Q3 batching adapter 51 0.777 s 13/13

This proves a causal storage-service reduction in the control, not the cause of the retained browser failure. The complete sealed generated native Node caller also runs under the adapted real ptrace supervisor, with actual native cause/input-wire hooks and a controlled peer. Request 6 resolves, Page.enable and Runtime.enable follow, then a deliberate protocol refusal stops before navigation. There are 19/19 actual reaps. The peer supplies no Chromium model. Separate native controls verify a complete durable ordered prefix before every executable CONT, bounded/saturated journals, sticky sync failure and failed finalization. The CDP deadline remains 20,000 ms; no failed request is replayed.

Sole remote attempt: preparation error, terminal refusal

The 944-provider candidate was sealed at manifest SHA a56667e2015d74fcd538be30520999eeed15a4fd3f5021ef1033183197e6af3d. All common providers match 2035 except tools/q3-current-module.cjs; all 327 media pins and both freshly built modules are unchanged.

I armed a bounded passive loopback packet observer too early, creating a directory that the remote read-only preparation check expected to create. Its FileExistsError is retained. Both checks then passed in fresh directories, but my corrected preparation omitted copying the manifest to the phase's prepared/runtime-pins.json.gz.

The single actual generated launch consequently refused in q3-native-prelaunch.py:136 with FileNotFoundError for that phase-local manifest. This is my staging error. The generated parent (7650) and prelaunch child (7651) actually joined with status 1 and separate stdout/stderr EOF. The runtime parent, kernel supervisor, Chrome and CDP were NOT_CREATED. There was no request 6, input, screenshot, GL sample or original in this run. The refusal was terminal: no repaired relaunch, retry or conditional original.

The initial and actual-attempt passive observers both joined 0 with separate stdout/stderr EOF. Neither saw a debug port or retained a packet. Thus this run adds no actual wire or browser-service evidence. Their fixed 90-second, 512-packet/256-KiB bounds are independent of the native protocol deadline.

Routine offline test errors are also retained: a reused model assumed an earlier synchronous append; the first local peer lacked its parent directory; an experimental real fsync-refusal test contaminated a subsequent supervisor with its still-retiring child. Final tests use the native durable-CONT control and the existing isolated sticky-service controls. No local Chrome ran.

Lifecycle and next step

All 19 unique retrieved artifacts were independently reread for SHA/size and gzip CRC before exactly one delegated normal stop. There were no PNGs. Box bx_hyqpv258, UUID 4d2a5f42-4fd3-4e76-9847-f02b0f0e5e39, is independently API-confirmed ARCHIVED, with unforced completed stop stop_7d720ee5ed345c95 at 20:54:52.254 UTC. Initial expiry was 21:16:26.784 UTC; final archiveAfter is null and expiresAt is not returned. The independent immutable watchdog was observed before creation, had fixed deadline 21:16:27.435 UTC, and actually joined 0 with separate stdout/stderr EOF. No second stop, resume or delete was issued. Free /home remained above 1.5 GB.

Passed: new causal/storage and whole-generated-consumer tests; actual generated prelaunch/custody paths for both phases; capture boundary and real input-wire hook regression; foreign closure with 27 refusal controls; local controller; and flock /tmp/wine-build.lock bash tools/build.sh.

The next separately assigned native run must first place and independently hash-check the sealed manifest at each phase-local prepared path, in addition to the check-only directory. That prerequisite was missing here. Native request-6 repair and alive arena interaction remain UNVERIFIED.