D3DIM input starvation repair boundary, 2026-11-18

The repair is not implemented. The authored native 32-draw control again expires an unchanged 800 ms Escape before application. New per-record timings show that its long kick contains many short draws: 320 DRAW_UP records total 5,958.527 ms, with a maximum individual record of 35.279 ms. The longest full kick in that run is 769.720 ms. This measures the authored fixture, not an individual record from the previously closed NFS3 original.

Evidence: scratch/runs/20261117-d3dim-render-input-starvation-nfs3-repair/. The preceding original and its 2,916.589 ms kick remain documented in the supervision investigation. No new original was launched. NFS3 driving after repair, Pause, Continue and native Quit remain unproved. The old SIGTERM sender remains unknown.

Native controls

The controls reuse published source 013d916c64632243ff0aa768debbbc39870f04d5, module SHA-256 cf674e8d1780249498a3b9797142d5d9f01e9daedcf2089c8fc880f515c753c0, and the existing Device2 indexed RGB565+D16 fixtures. Source bodies stay in a sealed memfd; Git and the recorded patches remain the source archive. The published 976156b26 public observer is unchanged. A separate bounded passive wrapper measures executor records without additional native reads or yields.

There are three launched native controls: the first uses only the existing public observer, the second adds record timings to the 32-draw fixture, and the third adds record timings to the one-draw fixture. Both loaded controls cancel Escape before handler application. The timed loaded control observes physical Escape/arrow UP and an empty queue 959 ms after the intended release. That is late physical release, not proof of on-time release or native input consumption. A preceding invocation failed locally on a missing argument before creating a native process.

The one-draw control consumes Escape at the native USER return boundary and changes its visible caption to AUTHORED PAUSED Escape delivered. The selected whole PNG was inspected. Its grey caption fixture proves input handling and the authored Pause consequence, not game geometry or NFS3 playability.

Each launched control prearms an independent observable 65-second deadline and disk-floor watchdog. The launcher actually waits the supervisor and retains its stdout/stderr EOFs. The supervisor actually waits native exit 0 and watchdog exit 0, retaining both native pipes and both watchdog EOFs. Each independent release helper is separately waited with both EOFs. No watchdog fires. Loaded cancellation is terminal: no later query, capture, rebind or input is sent; only the already armed release and owned shutdown run. Native exit 0 is owned TERM cleanup, not native Quit.

Concrete scheduling blocker

GpuStream.drain() executes each complete record before publishing its read_pos, arena retirement, done_seq, and fence/readback slot completion. GpuStreamHost.Bridge.kick() calls that drain synchronously in cooperative mode. The native backend already offers prepare/step/release/cancel, but the software executor selects its synchronous draw() wrapper. Yielding between kicks therefore cannot interrupt the measured drain.

NFS3's previous native log identifies rendering on a cooperative guest thread. ThreadManager.runSlice() invokes its native run() synchronously, then reads registers, yield state and sleep state, with _runningThreadHandle held across that call. Synchronous message and DLL callbacks also enter guest execution. The CLI's main-owner facade deliberately remains local until its remaining startup, observation and yield paths stop accessing local guest registers. The existing worker import broker can park a worker, but the native CLI does not select that guest-main ownership route. The foreign guest-host, render-manager and legacy-adapter WIP were neither modified nor adopted into the sealed controls.

A small independently executed Wasm counterexample makes the missing seam concrete. On default Node v24.18.1, WebAssembly.Suspending and promising are absent; returning Promise.resolve(1) through a plain i32 import immediately returns 0 to Wasm. With --experimental-wasm-jspi, a promising export and suspending import correctly return 1, but an ordinary synchronous JS callback that re-enters Wasm before suspension throws SuspendError: trying to suspend JS frames. Adding a Promise to kick, or enabling JSPI only for the outer slice, is consequently insufficient. Neither experiment modifies the emulator or supplies a production repair.

The unresolved implementation requirement is an async-safe cooperative guest execution boundary, including synchronous callback entry and sibling scheduler exclusion, or a completed owning-guest-worker CLI route. It must preserve SYNC/readback return values, record/arena retirement, ordered presentation and close cancellation while permitting host input/release. No worker-busy wait, early successful SYNC, draw omission, weaker ownership predicate or clock change was introduced to obtain a passing result.

Validation limits

Decoder, ring and final passive-observer tests pass. The shared render-worker test passes its ordering, ownership/bounds, scoped-error and heap-retirement checks. The current render-worker lifecycle suite fails in its pre-existing VM fixture because require is undefined while loading browser-native-input; this is recorded, not reported as a pass. A canonical locked build was also attempted; its exact completion and any subsequent native GPS tests are in the result record. These controls do not qualify an absent repair.

No production files changed. This report records a concrete unresolved execution boundary; it does not claim that the task's repair or affected-original acceptance is complete. No boxes, additional workers, archives, source-tree snapshots, media copies or deployments were created. Box ID null, state NOT_CREATED, expiry null, stop receipt N/A. The task releases its file claims and stops after preserving evidence.