NFS3 supervision and original input timing, 2026-11-17

The fresh original survived supervision and produced new released driving evidence: 0 MPH initially, 1 MPH during acceleration, a turned wheel during Up+Left, then 16 MPH with the car off the road. Pause expired before handler application. Its measured blocking span is now known: one synchronous gpu_stream_kick lasted 2,916.589 ms across the entire 800 ms intention. Pause, Continue and native Quit remain open. No production change or second original was made.

Evidence: scratch/runs/20261117-nfs3-supervision-signal-source-driving/. This follows the passive observer handoff. The earlier 35.5-second SIGTERM sender remains UNKNOWN. This run does not reinterpret that termination as normal shutdown or invalidate the retained c9ff55052 driving evidence.

Public supervision controls

The inherited launcher forwards TERM/INT to its supervisor; the supervisor signals its native process group and waits it while draining both pipes. The independent watchdog checks its supervisor identity, finite deadline and free disk. Public process/signal-only strace records exec/parent/wait events and kernel si_code, sender PID and UID. It does not trace request bodies, memory, render records or proprietary instructions. Source bodies remain in the inherited sealed memfd, with no source archive on disk.

The first orderly native control ran about 40 seconds. The final authored Device2 RGB565+D16 control ran 53.3 seconds through the same launcher, native host, final observer, interactive guard and custody chain. Both retain the 390-second host limit and prearmed 420-second watchdog. Escape was consumed at the native return boundary with its unchanged 800 ms intention, the authored PAUSED caption was reviewed, independent release observed physical UP, and orderly close completed native/supervisor/watchdog/launcher waits and both EOFs. This grey caption fixture proves its own handler, not game rendering geometry.

An absent authored-fixture path exercises the host error return (native host exit 1, watchdog 0, complete parent waits/EOF). A deliberate controller-side TERM exercises native cleanup with known signal provenance: controller 2004663 → launcher 2004677 → supervisor 2004729 → native 2004731, UID 1000. Killing supervisor 2004075 is a separate parent-death control. Watchdog 2004076 detects the lost parent and sends native 2004078 TERM; the enclosing subreaper actually waits the adopted native (0) and watchdog (2), retaining both read ends and observing each pipe EOF. The launcher actually waits its killed supervisor (-9); tracer exit 247 is preserved. PID absence is only a separate postflight observation.

No unexpected TERM or 35-second timeout reproduced. The observed bootstrap timeouts precede the native chain; the inherited runtime deadlines are 390/420 seconds. There is no proved accidental lifecycle timeout to correct. Two early local preparation errors occurred before supervisor/native launch and are recorded separately. The initial board watcher separately returned 143.

One authorized original

Runtime source remains 013d916c64632243ff0aa768debbbc39870f04d5, module SHA-256 cf674e8d1780249498a3b9797142d5d9f01e9daedcf2089c8fc880f515c753c0. All 4,681 sealed source pins matched. The only observation overlay is final published 976156b26b5e5e916ab032a111cbfabe5babfd32, SHA-256 affbb745adecbaa558f120f8fc5cbab8427662c013c663f57df609ad0568a18c. It adds no ownership snapshot, input, timer or yield. Its focused on/off, snapshot-count and bounded-ring checks pass again. The prior canonical locked build is reused; this documentation-only task makes no new build claim.

The genuine accepted chooser export remains SHA-256 c08b656b3718682a3a1f4324a08ab689c9444c1daf868c40f4a991cb9f1b9108. All 135 media pins matched before and after. Foreign WIP was excluded by the sealed inventory. No guest clock, configuration, save, media or state changed. The original naturally reached its rainy Hot Pursuit cockpit. The first gas hold showed no speed increase; the next held image shows 1 MPH. The steering hold shows a turned wheel, and the subsequent whole image shows 16 MPH and trees directly ahead after leaving the road. Transmission remains unverified; this is driving progress, not sustained speed or a completed race.

Action/release functions match the tested predecessor by AST. Driving retains 6,000 ms; Escape retains 800 ms. Each release helper was independently armed before DOWN. There are four consumed arrow DOWNs (three Up and one Left), followed by one unconsumed Escape. The table separates request authorship from server entry and the native receipt.

Intention Intent → HTTP authorship HTTP authorship → entry Native consumption
First Up 74.003 ms 890.603 ms consumed, 7 ms after application
Race Up 73.654 ms 841.103 ms consumed, 11 ms after application
Steering Up 65.292 ms 22.867 ms consumed, 4 ms after application
Steering Left 100.923 ms 3.236 ms consumed, 649 ms after application
Race Escape 259.798 ms 2,463.302 ms none; handler not applied

Escape intention began at 1794959320873.900 ms and expired at 1794959321673.901 ms. The synchronous render span began at 1794959320599 ms, lasting 2,916.589 ms. HTTP entry was 1794959323597 ms; preparation began a millisecond later and took 0.432 ms. There was no application span and no native input receipt for Escape. This directly identifies the synchronous render span blocking this original request; it is not an inference from the unused legacy D3D worker's waits.

Physical UP is separate evidence, not assumed from a release ACK. The observer records every arrow/Escape zero and an empty queue at each existing release. Observed release lateness was 139.606, 51.758, 495.159 and 1,938.099 ms, respectively. In particular, Pause's physical release was late because the same host was blocked; no strict on-deadline UP claim is made. The observer retained 669 of 2,048 available samples with no drops. Its 601 synchronous render calls totaled 115,817.805 ms; the 2,916.589 ms Pause overlap was the maximum.

The Escape cancellation permanently closed the endpoint. No later query, capture, rebind, input or replay followed; only the already armed release ran. Owned shutdown then records kernel sender 2010293 → launcher 2005571 → supervisor 2005647 → native 2005650, UID 1000. Native exit 0, watchdog exit 0, and supervisor/launcher waits have both pipe EOF receipts. The outer tracer has actual tool completion 0; the tool merges its output, so separate outer-tracer stdout/stderr EOF records are unavailable. The original ran about 191 seconds, with no watchdog firing or spontaneous TERM. This is owned shutdown after UNKNOWN, not native Quit.

Concrete repair boundary

The failure now has an original public timing cause, but no tested scheduling repair. The measured interval is a complete synchronous GPU drain, not an individual draw/record measurement. The published GPU-stream contract requires SYNC/CLOSE kicks to finish all published records, with producer checks and pixel-readback fences. Returning early or inserting a batch-loop yield cannot preempt this already-running import safely. Moving execution or making that boundary resumable needs an authored native counterexample and negatives for drain/fence/readback/close ordering before it can be a focused repair. No such repair was established here, so the conditional affected-original allowance was not used. A further original requires new authorization or a proved, tested and published repair under that conditional allowance.

Exact Git identities, relevant binary diffs, small run-input patches, selected named whole PNGs and artifact hashes preserve the evidence. No workers, boxes, archives, source snapshots, media copies, frame series or deployments were created. Home free space stayed above 1.5 GB; no RAM gate was used. Box ID null, state NOT_CREATED, expiry null, stop receipt N/A. Foreign WIP and other owners' work were preserved. This task releases and stops with the sender of the old TERM still unknown and the synchronous original rendering boundary documented.