Q3 checkpoint observation: durable owning-caller adapter

Delivered new opt-in tools/q3-checkpoint-observer.py and executable offline integration tests. Existing Q3/Q2 providers, manifests and active pin closure are unchanged. Q21495 retains sole NEXT runtime. This does not authorize a native retry or qualify cancellation.

Retained1475 does not establish a causal browser defect. The actual observer command was submitted at Unix1794232273.634765; remote command started 2026-11-09T13:51:13.704Z and finished13:52:17.472Z with the existing checkpoint assertion. First local terminal was1794232337.4970028. The raw browser budget record began1794232303949ms with startupDeadline1794232423949ms and overall 1794232754460ms. Kernel armed supervisor7077/start34224, command7078 and Chrome7507 are retained separately. No endpoint or admission had been published. The browser startup deadline remained open. Empty Chrome logs, absent receipts, and kernel reaps establish neither browser EOF nor historical cause; both remain UNKNOWN. retained-review.json preserves raw records, exact hashes, sizes, command, custody timestamps and comparisons against the initial Git base.

The original caller polls only checkpoint existence, then reads it. Executing that exact retained command offline demonstrates false acceptance of partial JSON, and of a published checkpoint alongside a prior browser failure. These are concrete observer defects, but are not asserted to explain1475's missing checkpoint. New tests reject both. No historical logs or CDP receipts are created.

The new adapter composes the actual retained task launch branch through existing separate-stream cancellation composition, refusing unsupported source shapes. bind_remote(task, session) replaces only the generated owning parent's Popen and drain adapter in memory. A durable exclusive spawn journal records intent before the actual Popen and a distinct spawn failure. The launch journal records actual child PID/start, bounded publication snapshots and final actual wait with independent stdout/stderr EOF. Snapshots label readiness and browserEOF UNKNOWN; they do not authenticate endpoint/CDP or convert publication into admission. The final post-drain snapshot closes a fast-exit observation race found in tests.

observe_checkpoint(task, session, remote) replaces1475's checkpoint remote poll, using the same60-second observation and70-second transport caps, with monotonic elapsed time. It records observer start, publication changes and missing/partial/damaged/oversize/nonregular state, deadline, cancellation, prior failure, invalid checkpoint or checkpoint observation. Owning Session checks bracket the operation. Prior failures and revocation take precedence over checkpoint presence; physical and cancellation-qualification gates still follow. No timeout, readiness waiver or first-terminal replacement is introduced.

Reads are nonblocking/no-follow and limited to64KiB compressed and expanded. Journals are fsynced NDJSON with128-record and1MiB bounds, exclusive creation, sequence numbers, wall/monotonic timestamps and raw publication hashes. Damage or cap exhaustion revokes input via the existing sticky cancellation channel; the existing identity-bound TERM, sole drain and actual wait finish before the diagnostic failure is raised. A separate failure-wait record preserves those actual waits. This helper adds no wait consumer or custody process and never claims browser EOF from kernel reaping.

Tests execute the actual generated caller and actual prelaunch with inert Python children: normal, missing/partial publication cancellation, damaged journal, record-cap cancellation, spawn failure; actual remote observer adapter success and sticky cancellation; original command false-ready controls; deadline, prior failure, shape, FIFO, symlink, size and gzip damage negatives. Independent observable75-second sibling watchdogs are armed before tested children, outside any custody wait domain. All owned children are actually waited; runtime and prelaunch waits and stdout/stderr EOF are separate. No browser, game, CDP, network endpoint, Box or native custody is invoked by these tests.

Focused existing caller controls pass, including EOF-before-exit and inherited writer cleanup. Existing separate-stream regression initially refused its stale offline fixture manifest (missing current-module provider). Its exact assertions pass with only that required test input added in memory; original failure is retained. New integration tests supply the current mandatory provider directly. Locked flock /tmp/wine-build.lock bash tools/build.sh passes. During development a fast-exit diagnostic assertion exposed the final-snapshot race, now repaired.

Evidence is scratch/runs/20261109-q3-checkpoint-observation-repair-1525/: exact Git base, small owned binary patch, raw retained review, test logs/waits, watchdog observations, locked-build output and a review-only future caller recipe. Git remains the source archive; no module/media/source archives are copied. Free/home remains above1.5GB. No shared custody, CPU/WAT/host, Q2, Redline, HeroesII or CLAUDE edits; no agents/workers, deployment or native qualification.

Next: a separately authorized future closure must explicitly stage and pin this new provider and use both adapter entry points; do not mutate existing versioned manifests or invalidate Q21495's active closure. No runtime attempt is authorized here. No Box was created or contacted: ID/state/expiry/stop receipt NOT_CREATED/NA. Inherited1475 bx_5bnxexae remains offline-uncontacted; retained final receipt is ARCHIVED/PHYSICAL_DRAINED with completed unforced stop_54e758ada497a4ee at 2026-11-09T13:53:20.194Z, initial expiry14:15:28.321Z, final expiry NOT_RETURNED.