NFS3 Pause: pre-handler delivery boundary, 2026-11-17

No production change or new original was made. The retained NFS3 Pause remains terminal UNKNOWN. Authored native Device2 fixtures reproduce expiry before handler application under synchronous D3DIM rendering, but the original lacks the ingress/render timestamps needed to establish that this caused its failure. There is no proved NFS3-specific repair to publish, so the conditional original allowance was not exercised. Driving/Pause/Continue/Quit verification remains open.

Evidence: scratch/runs/20261117-nfs3-pause-native-delivery-cause/. This follows the inherited-timing run, published as c9ff5505236c34357b59829e5185523948445ea6. Its released 0→3→4 MPH and steering observations remain intact. Its Escape deadline was 800 ms, release-bound, and returned CANCELLED, unconsumed, handlerApplied=false. There was no native Escape receipt. The old endpoint was never queried or replayed.

The retained original counters distinguish two render paths. Its legacy d3dCommands worker had zero submissions, waits and replay commands. The D3DIM stream had 587 kicks, 73,769 records and 567 native presents. Thus an explanation based on the legacy Encoder's one-second Atomics.wait is unsupported. The published host-imports.js routes D3DIM stream kicks through gpu-stream-host.js; the local bridge synchronously drains published records. test/run.js services HTTP on event-loop turns between guest batches, and the guard refuses an expired intention before invoking the physical handler.

The fixture executes an authored x86 PE, compiled from C using existing public Device2 fixture recipes: RGB565 primary/backbuffer, D16, indexed geometry, Flip, PeekMessageA, and ordinary Escape handling. It uses the original's exact published source 013d916c64632243ff0aa768debbbc39870f04d5 and module SHA-256 cf674e8d1780249498a3b9797142d5d9f01e9daedcf2089c8fc880f515c753c0. The sealed in-memory inventory hash matches the original across 4,681 files; foreign WIP was excluded. A local resolver fallback permits already pinned Git modules whose filenames have since disappeared from the shared tree. It changes no runtime source body. No source archive or copied media was created.

Passive host wrappers measure existing D3DIM calls without additional native calls, waits or yields. Existing transport and public-input diagnostics locate request dispatch, guard preparation, physical handler and native-return phases. The same 800 ms immutable deadline and prearmed independent release apply to all input cases. Guest clocks, runtime flags and configuration remain unchanged.

The 32-draw fixture authored Escape at 1794957637147 ms, expiring at 1794957637947. Two synchronous render spans covered the ingress interval: 1173.129 ms starting at 1794957637040, then 1072.099 ms at 1794957638236. HTTP dispatch occurred at 1794957639314, 2167 ms after authorship and already 1367 ms after expiry. Guard preparation immediately cancelled it without a physical handler or native DOWN. The independent release completed; no game query or replay followed terminal UNKNOWN. A second authored 32-draw case reproduced pre-handler expiry and separately observed Escape=0, Up=0 and an empty queue during the prearmed release, without an extra endpoint query.

A one-draw control consumed the exact native WM_KEYDOWN Escape through PeekMessageA within 71 ms of authorship. The guest's own handler changed its caption to AUTHORED PAUSED Escape delivered, inspected in the selected whole PNG. A separate passive release observation recorded Escape=0, Up=0 and an empty queue. Receipt, native application, visible caption and physical UP are separate evidence. The whole image is grey with the authored caption; it is not a raster-geometry oracle or evidence of game driving. This fixture proves its own Pause only, not original Pause/Continue/Quit.

Preparation failures are retained honestly: the first uncached source launch hit a desktop-helper startup timeout before input; restoring the inherited sealed memfd cache resolved preparation. A 256-draw fixture measured 11–14 s synchronous kicks but exceeded its readiness observation bound, so sent no input. Neither is an original-game failure or a successful input test.

Focused guard/native-boundary and main/worker event-scope regressions pass. There is no production edit, so no new canonical build is claimed; the exact previously built module was reused. All 135 retained media pins were rehashed unchanged. Native children, independent release helpers, supervisors and watchdogs were actually waited with both pipe EOFs, including expected fixture failure paths. Own shutdown was SIGTERM, not native Quit. No watchdog fired. No box, worker, deployment, source/media archive or frame series was created. Box ID null, state NOT_CREATED, expiry null, stop receipt N/A.

The concrete remaining boundary is original authorship→server ingress→guard application. Its retained receipt cannot divide that interval into transport, scheduling and rendering time. The authored reproduction establishes a causal mechanism on the affected public path, not its occurrence in the original. Changing timeouts or the guard would hide the boundary; a speculative scheduler or render change would not meet the task's proved-cause requirement. Stop with this evidence and keep the original allowance unused.