Redline bounded key delivery — offline diagnosis, 2026-11-09

The first-comma native failure from 0856 remains a concrete blocker. No safe delivery-latency repair is justified by the retained evidence. The implemented correction preserves the failed hold receipt and records send/response timing; it does not claim to make that native run pass.

The actual path is redline-native-fixture.cjs → production redline-browser-run.cjs → numbered transport → execute → identity admission → fixture baseline poll → the retained scratch/runs/20261108-redline-car-door-native/bound-input.cjs hold. That helper arms its independent 600 ms release before sending comma keydown through the production pending-CDP map. It awaits the response before sampling. The timer dispatches keyup independently. On a response after release it rejects with input delivery exceeded hold bound, attaches holdReceipt, and joins release. The transport records first failure; the caller closes admission and joins independent cleanup. No later input, capture or retry is allowed.

0856 admission was recorded at 08:59:41.553Z, first transport failure at 08:59:42.547Z, and caller failure at 08:59:42.548Z. The stack identifies this exact retained helper and production caller. There is no saved hold receipt, sample, individual CDP dispatch/response timestamp or native held-read proof for the failed command. Admission-to-failure elapsed time includes the pre-hold fixture poll and cannot be assigned wholly to keydown service. Custody evidence does not identify this CDP command's service latency. No specific Chrome, ptrace, fsync or guest scheduling cause is established.

The new redline-key-delivery.cjs observer delegates to that exact helper with the same options and send order. It records monotonic elapsed and wall times for each dispatch and response, including errors, and saves the helper's failure receipt in input-ID-delivery.json. It returns the original successful receipt or throws the original exception. An evidence-write error cannot replace the first input error. The production caller now invokes this observer and requires it in the verified startup-provider closure. Future staging must include this provider; an incomplete pin set fails closed. No timing constant, timer origin, sampling condition, input type or CDP timeout changed.

The offline test executes the actual fixture controller and production caller, real retained input/identity helpers, numbered file transport, real Node child waits and separate stream EOF. CDP/page/Worker/PE execution is inert. Controls:

All negative controls prohibit post-terminal guest reads/input/captures and retries, observe the actual inert child exit and both stream EOF events, and join the independent controller watchdog. No fabricated native held read is introduced. Restarting the timer after ACK, granting delivery grace, or treating pre-ACK observations as held proof would weaken the required bound or evidence; none is implemented. The historical failure therefore remains terminal.

Evidence is in scratch/runs/20261109-redline-key-delivery-repair-0912/. Accepted unchanged proof references and exact retained helper bytes were SHA/size checked and reused. Removing only the observation call and provider requirement reconstructs the pre-change caller exactly. Startup admission, fixture PE, module/instance/endpoint/slot checks, flags, freshness, native REVOKED and all-writers/edict-ABA UNKNOWN semantics remain unchanged. Recovery uses the recorded Git base and small binary patch; no source/media/module archive was made. Gzip logs begin compressed at first byte. The initial offline watcher compression error and initial stalled-control expectation error are retained and excluded; neither ran a browser or guest.

Validation: node tools/redline-key-delivery.test.cjs and flock /tmp/wine-build.lock bash tools/build.sh. No unchanged native run or separate qualification task was performed. No Chrome, box, guest, worker, agent, deployment, shared host/WAT/custody, Q3, Heroes II or Claude edits. This task's box ID, UUID, expiry and stop receipt are null; state NOT_CREATED. There is no allocation wait or physical-drain claim for this offline task.

Next, only under runtime authorization, collect the new bounded delivery receipt through a fresh production fixture before considering original input. Native key-delivery latency and gameplay consequence remain unverified. SingleTaskSTOP.