Redline startup service investigation
Evidence: scratch/runs/20261109-redline-final-startup-service-repair-2145/.
The retained final generation has 6,002 custody rows and 4,096 observer records. Its last SYSCALL release precedes Chrome's spawn; there are zero syscall-stepped releases during the 10.038-second, 100 × 100 ms port gate. Chrome main PID 7267 has a 5.716837-second consumed-stop lower bound and 4.321163 seconds with execution permitted between completed resume and next consumed stop. Execution permitted includes kernel waits and descheduling. Neither interval attributes storage, DBus or kernel cause. The 306-row startup consumer diagnostic has zero drops.
The Redline adapter now measures fsync, ptrace service and priority wait scans, with monotonic wall and thread CPU clocks. At consumed waits it records bounded comm/executable/command-line snapshots, verifies their start-time binding and retains actual wait sequence references. These observations do not establish ownership or lineage. The separate diagnostic has fixed record/byte bounds and writes only at finalization. No new ptrace stop, clock extension, admission or shared-provider edit is introduced. The existing every-CONT full-prefix barrier, first-unlinked-EXIT handling, sticky failure and final-sync checks remain intact.
The composed fixture controller now waits for separate exact-image review. Its failure path suppresses Browser.close after the first failure. The conditional original caller uses the same terminal composition.
Offline causal controls ran the actual generated supervisor and shipped Node port consumer with a small real publisher (no local Chrome). An injected 8 ms storage service changed measured fsync wall time from 0.096 to 0.761 seconds; an injected 4 ms ptrace service changed measured ptrace time from 0.006 to 0.411 seconds. Raising the syscall observer cap from 2 to 4,096 changed measured ptrace calls from 92 to 1,446. All four publishers passed the unchanged port consumer. This distinguishes instrumentation pressure and operation service in controls; it does not establish the native failure's cause.
The final candidates were independently sealed and all 904 providers reread
before allocating one box. Native manifest:
35f149ab0bc430fead0e950d82015af73b223ad4b69bdd08fa377ffce373aec8.
The adapter was 2cc5b9b6231c6b3f1ffcc6307e8ad11815168a7e47bc464702f33d644d9101b6.
Exact Git base and small binary diffs are retained; no source or media archive
was made. Shared providers and Nox/Q3 files were not edited.
One instrumented native fixture passed: Chrome PID 7857 published in 4,520 ms, 45 polls of the unchanged 100 × 100 ms gate. The synthetic fixture's startup image was separately reviewed before comma input; its comma image was also reviewed. It completed its supported comma-only path and ordinary quit ACK. Custody reaped all 175 owned tasks; parent exit 0 has actual wait and separate stdout/stderr EOF.
This supplies a new measured service boundary. In that gate, measured durable journal fsync calls overlap 4.304231 seconds; measured ptrace service overlaps 0.003447 seconds and priority scans 0.008201 seconds. Roles include the real Chrome launcher and its readlink/dirname/mkdir/touch/hash/curl/jq helpers, Chrome main, Chrome_IOThread and Chrome_DevTools. The bounded diagnostic retained 29,846 rows with zero drops. These are call-service wall measurements including descheduling, not evidence of a filesystem, DBus, kernel or CPU-time cause. There is no native A/B causal improvement claim and no proven startup service repair: the earlier timeout's underlying cause remains unknown. This fixture was justified by the missing task/service identity, and supplies that boundary. The durable release barrier was preserved rather than weakened.
The one conditional original then published in 2,708 ms / 27 polls. Separate image reviews showed the Win32 launcher, title, Start Up menu and level grid. The first runtime refusal occurred at 21:59:16.456Z: the level grid had acquired pointer lock, so the ordinary absolute Stadium City click was rejected before input. I incorrectly relied on the preceding menu's unlocked state. This is an input-admission refusal, not proof of guest failure. No input, capture, retry or quit followed. No side-door entry or driving was reached. Original custody reaped all 354 tasks; parent exit 1 has actual wait and separate stdout/stderr EOF. Independent socket checks found both lanes' server and DevTools ports closed and no profile processes; remaining profile files were removed only after that physical drain.
Meaningful checks passed: actual generated supervisor plus shipped startup
consumer under independent storage/ptrace/stepping controls; fixed diagnostic
bounds and failure propagation; actual composed terminal-close and external
image-review gates; durable-prefix/first-unlinked-EXIT/sticky/saturation/final
sync tests; bounded custody creation/cancellation controls; startup/boundary
consumer tests; full generated parent/controller/caller offline preflight
(10 tasks reaped, expected missing native ACK); conditional original preflight;
and flock /tmp/wine-build.lock bash tools/build.sh.
Joining the actual consumed-stop and service rows sharpens that boundary: Chrome main's fixture consumed-stop lower bound is 2.597204 seconds, of which 2.557263 seconds overlap measured journal fsync service in the single supervisor. 1.337116 seconds match that main task's exact consumed wait sequence; the rest serviced other task releases while main remained stopped. Main's 1.922796-second execution-permitted interval is not CPU time. The original independently has 1.481210 seconds stopped with 1.450417 seconds overlapping journal fsync service (0.745583 seconds matching main's wait sequence). Both gates have zero SYSCALL releases. Thus the newly identified serial service is durable release journaling, not active SYSCALL stepping. No attribution is retroactively assigned to the retained failed gate. Original diagnostic retention later saturated at 32,768 rows with 235,717 explicit drops; aggregate totals survived. Startup's 88-row consumer diagnostic had zero drops. Analysis outputs retain per-task roles, wait sequences, intervals and the service intersection calculation.
Before shutdown, 223 remote artifacts / 204 distinct hashes / 13,768,449 bytes were downloaded and separately reread against independent size/SHA inventories; PNG CRCs and gzip CRC/EOF were checked. One original Chrome stderr gzip is truncated and retained explicitly as failed cleanup evidence. It is not claimed as browser stream EOF. The original's shutdown-failure preserves the first pointer-lock refusal; custody/parent EOF and physical drain are separate proofs.
The sole box was bx_fb2ugk2y, UUID
df08822e-66d9-471d-a83b-7d6192521aa0. Initial expiry was
2026-11-09T22:14:47.669Z; the independently observed immutable watchdog deadline
was 2026-11-09T22:14:47.928Z. Exactly one delegated normal unforced stop completed
at 22:01:43.883Z, receipt stop_5ed6ce24ec1fbe4b. An independent API reread reports
archived (final archiveAfter null). The watchdog was actually waited, exit 0,
with separate stdout/stderr EOF and unchanged configuration hash. It is absent
from /proc. No box remains running for this task.
Next, only under a new assignment: test a durability-preserving lane-specific release-service repair against this measured boundary before any fresh replay. For gameplay, the visible level menu requires a current pointer-lock observation and supported reviewed relative/cursor-only input. This task stops here.