Q3 durable release service

Evidence: scratch/runs/20261110-q3-durable-release-service-0106/. Prior cause: startup deadline attribution.

The Q3-only authenticated custody adapter now drains ready kernel stops before releasing queued tasks. One ordered durable journal prefix can authorize a cohort of already-consumed stops. The shared custody provider remains unchanged.

Every identity, lineage, consumed wait, retirement, release intent, cancellation and error record is required before the next actual CONT. Only results of already-authorized releases (ptrace-cont-returned, parked-child-release, parked-child-release-pending) may remain behind that durable frontier. They remain in the same bounded ordered journal, become durable with the next required prefix or final close, and are never dropped. Saturation still applies synchronous backpressure. A prior storage error always rejects release, even when a cohort was previously certified. New required records after saturation cannot inherit the preceding batch's certification. Gzip footer and final durability remain mandatory.

The ready-stop scan retains creator/terminal priority, PID/start checks, nonleader exec transfers, and actual wait obligations. Command-output service still yields to ready kernel stops and queued resumes. Runtime, cleanup, image and protocol deadlines are unchanged. No guest calls or writes were added.

Focused controls compare the prior adapter recovered from exact Git commit 2bc65ee74e5a5726a58a86ea1a390f6ec1e3d185 with the changed adapter. They observe successful real fsync completion separately from readable OS-cache bytes before every actual CONT, require the durable prefix through all required records, and require final journal bytes to equal every observed record in order. Real Node fork traffic, unlinked irreversible exit, nonleader exec transfer, pending-lineage cancellation and superseding exec are exercised. The generated consumer controls use real Node children and ptrace custody under slow storage; refusal, socket loss, synchronous send failure, withheld acknowledgement and storage failure retain actual parent/adoption waits and separate pipe EOFs. The storage-failure supervisor stays rejected even after its killed descendants are independently drained; those later waits do not make a failed journal valid.

The first full consumer suite incorrectly compared total wall time, including the unchanged 20-second cleanup grace, against a startup-performance threshold. The corrected suite measures startup through the deliberate pre-navigation refusal. Original failed logs are retained. An initial unlinked-exit probe expected gzip although the reused fixture deliberately writes plain NDJSON; the probe now handles both formats. Production failures and invariants were not waived.

Repair 7607cab84 was committed with six explicit paths and immediately published on main using strict SSH. The final compressed-output generated consumer suite passes all eight cases. With injected 40ms storage latency, startup through the deliberate pre-navigation refusal fell from 5.731641s to 3.542586s (38.2%); synchronizations fell from 65 to 48. Total wall time, including the unchanged cleanup grace, fell from 28.316s to 25.517s. The withheld reply still uses the unchanged 20-second protocol timeout and the existing peer's separate 35-second exit cap. No later protocol/input/capture, navigation or quit crossed a terminal boundary.

The final 20-child Node pressure control used 122 versus 87 synchronizations and 9.455s versus 6.921s under injected 25ms service plus shared-host load. The same worker's required-prefix, full final-byte retention, sticky storage, footer, saturation, actual terminal/lineage/exec, command-output preemption, generated-caller and retained-cause controls pass. Forty existing custody regressions also pass; their standalone subprocess fixtures still load the unchanged shared provider, while the new direct fixture controls explicitly bind the changed adapter in each subprocess. The final locked build passes.

Exact source recovery is Git plus the small owned patch, not a source archive. Native closure preparation independently reread all provider SHA/size pins at each of five phase staging locations. The selected module remains the existing build/wine-assembly.wasm, 2,186,068 bytes, SHA-256 3e491c0840bb7771328738a0d160b4bbaa3d7ab72074cbd38a7b41e5d7e8ec18.

The one changed native control passed on the new box. The accepted red-triangle fixture was individually model-reviewed before its single Enter input; the second reviewed image showed the expected 80-pixel translation. Planned owning cancellation then drained 240 owned tasks through 240 actual reaps with zero remaining tasks and separate command EOF. The generated, prelaunch and runtime parents have actual wait receipts. This exercised the changed published service and unchanged fixture acceptance mechanisms.

The conditional original reached a usable Q3 Arena main menu under the unchanged startup image allowance. Its exact 137,030-byte PNG was model-reviewed before one native Enter hold of 150ms. Native key-down/up acknowledgments and two wholly-held physical Enter samples were recorded. The next individually reviewed image still showed the same main-menu selection; logo animation alone changed. This first live nonresponse triggered immediate owning cancellation. No later game protocol, input, capture, retry, quit or deadline extension was issued. Arena movement and firing were therefore not attained. The result does not establish whether the guest ignored the key or had not yet handled it.

The original drained 170 owned tasks through 170 actual reaps, zero remaining, with separate command EOF and actual generated/prelaunch/runtime waits. The native service logged 503 synchronizations totaling 0.603s for the changed control and 342 totaling 0.567s for the original. These native figures describe this run; differing host conditions prevent attributing their entire difference from the prior run to the repair. The matched slow-storage generated-consumer comparison above supplies the causal measurement.

The input wire contains exactly the two successful input requests. General protocol observation is bounded and its transport recorder saturated; the separate execution-boundary stream is empty. Consequently it is not an exhaustive protocol transcript. The published cancellation-file guard executes before every native send regardless of recorder state. The owning command stream contains the one input and cancellation sentinel, and there is no later image or input. These distinct evidence scopes are retained without treating missing observations as observed events.

All 138 unique remote artifacts (157 regular-file references, 3,785,079 unique bytes) were recovered before stopping. A separate process independently checked every unique SHA-256 and size, all recovered gzip streams, and PNG chunk CRCs; a second remote inventory matched. Four actual images were individually model-reviewed. Existing media was referenced by path/hash, not duplicated.

The local outer allocation transport wrapper terminated with status 143 for an unknown reason and did not write its own wait receipt. Its partial gzip stdout is preserved as-is with its hash in allocation-transport-interruption.json. This limitation does not replace the independently retained successful actual watchdog wait (status 0, separate stdout/stderr EOF), the native parent waits, or the independently observed completed stop.

Only box bx_hsh4576p was created. Before creation, its independent immutable 25-minute watchdog was armed and observed for UUID 75868b71-c444-4f0a-85db-cf0dc699dfbb, deadline 2026-11-10T01:47:43.330530Z; initial API expiry was 2026-11-10T01:47:42.943Z. After independent unique-artifact verification, exactly one delegated normal unforced stop completed: stop_09d00d982cdc6042 at 2026-11-10T01:28:05.953Z. The independent final API receipt records state archived, expiry archiveAfter: null, and stoppedAt 2026-11-10T01:28:05.937Z. The watchdog acknowledged the confirmed normal stop, exited with an actual wait and separate EOF, and released its lifecycle lock. Prior stopped boxes were untouched. No box remains live for this task.

The causal service repair is shipped. Any further Q3 menu/input investigation requires a separate assignment; this single task ends without another attempt.