NFS3 normal-Worker control qualification, 2026-11-18
The fresh browser pressure off/on pair passed. A separate authored Device2 control proved that the frozen public input guard required a GL endpoint and refused a healthy legacy Direct3D owner. The additive public-binder correction now binds the real legacy bridge and exact manager/port/module/memory/window identity. It preserves the 500ms confirmation, 15-second gameplay freshness, 250ms native scene fence and independent physical UP. No NFS3 emulator scheduling cause or repair has been established.
The same pressure fixture changed from native bind refusal to successful
binding with only this JS correction. Its first input preparation then
cancelled at a 332ms-old scene, before DOWN; the prearmed UP completed. A fresh
light authored control subsequently produced a genuine owning-Worker native
USER receipt and independent UP. Preparation was published as 994aa3a20
before one fresh registered NFS3
original. That original closed UNKNOWN during actual Worker module inventory,
before a whole-frame review or any physical input. The concrete remaining
blocker is the harness handling of retired Worker sessions.
Evidence is in scratch/runs/20261118-nfs3-normal-worker-scheduling-gameplay/;
result.json is written last. This task does not reopen any predecessor's
closed runtime scope.
Completed pressure qualification
Both fresh scopes used the same retained pressure PE
3b28b26d2fc14e97d66c4f8fca4c4d6c568600dec8879a0ffd2e3737304ce74a,
the actual normal guestWorkers route, software rendering, frozen source
c9d0c2cf720cc13b962ea8b9c2565de0b480d319 and native module
7d347a6a631ab77335a14935a533feb896be71aa2508fc5aa19b1e5df8f096ce.
Each new control had a finite 120-second observation allowance. The original's
500ms confirmation, 15-second game freshness and independent input/identity
requirements were not changed. The predecessor's 48/60 at 70 seconds was a
closed failure, not an endpoint or scope resumed here.
| Fresh browser pressure control | Completed presents | Launch-to-report ms |
|---|---|---|
| Observer off | 60 | 54465 |
| Observer on | 60 | 46772 |
Each scope recovered exactly three native modules: two actual guest Workers
and one render Worker, all matching the pinned module. Module inventory used
the debugger only after the measured interval. The 1280x960 whole-page PNGs
and decoded RGBA pixels matched exactly, SHA256
2542b8560f3c313664235302d608184480f0a03929cb8a0caa509421e2ec063a.
Both whole frames were personally reviewed: the authored blue Device2 output
occupies the upper-left game region, with the normal full-page host strip and
white remainder. This is a control image, not an NFS3 gameplay image.
The observer-on main/auxiliary sampling totals were 6.745ms/3.050ms for
65/62 retained samples. These costs exclude wrapper, aggregation, scheduling
and report serialization. Main-thread aggregate native time includes
24.419 seconds inside synchronous gpu_stream_kick imports; it is not CPU
time. Observed maximum main/auxiliary gaps between runs were approximately
950ms/545ms. The sequential shared-host elapsed difference of about -14.1%
does not isolate observer overhead or establish neutrality, a speedup, or
the cause of the old ten-second game servicing gap.
The four published native controls and browser light off/on pair from
753391a64/report 6a6d13714 are reused for their unchanged claims; they
were not rerun or relabeled as new measurements. Their source and retained
executable/module identities remain distinct from current shared-tree WIP.
Passive field-reader qualification
tools/nfs3-normal-worker-control.cjs serializes the existing published
observer with its unchanged 100ms cadence, 4096-row capacity and 16-address
field set. The serialized bootstrap was checked byte-for-byte against the
bootstrap actually executed in the passing browser pressure control.
The focused test compiles a 477-byte native Wasm control with distinct
little-endian data initializers at every original field address. It exercises
both actual WebAssembly.Instance and WebAssembly.instantiate, samples real
run-entry/run-exit and existing tick/GPU imports, verifies all 16 translated
values, and compares memory before/after. Native return values and exactly-once
import counts match. Debug and read-and-clear exports trap if called; neither
was called. The native module, WAT input and actual output are retained in the
run. The real frozen PE/browser control also sampled all configured addresses
without mapping errors. Its zero values are fixture data, not game operands.
This qualifies bounded passive reads without debug dispatch or extra guest polls. It does not turn asynchronous boundary samples into witnesses of an internal NFS3 instruction or decision. The earlier mode-zero/lead-four offline predicate explanation and Escape ring-lag catch-up remain as described in the predecessor report. The old 25.9-million-callback observer's perturbation and the emulator cause remain unresolved.
Diagnosed and repaired public-binder preparation
A third, fresh authored Device2 control enabled the existing browser native guard and completed 60 presents, with three recovered native modules matching the same frozen module. No physical input was injected. At the checkpoint:
- The native owner was healthy, unbound and nonterminal.
- Its main/focus HWND was 65537, process 1000 and native thread 1; no modal or capture was present. The valid scene observation was 46ms old.
- The legacy Direct3D bridge existed and had completed 60 presents.
- The GL bridge and GL endpoint were absent.
snapshot().readywas false andbind()refused withbrowser native owner unavailable or already bound.
The frozen lib/browser-native-input.js identity reads
wine.hostCtx.getGlGpsBridge()?.endpoint; ready() requires that endpoint.
An absent GL endpoint is therefore sufficient to prevent binding even with
the observed fresh healthy native scene. The actual refusal and source hashes
are retained in guard-readiness.json and guard-source-proof.json in the run.
This demonstrates failure of the available unchanged guard preparation on a
legacy Device2 owner; it is not an observation of NFS3's producer/consumer
servicing and is not a scheduler defect diagnosis.
The correction adds the actual legacy bridge to identity selection, without creating a GL bridge. It requires the real d3dim-gps endpoint in the owning render manager's port map, that manager's Worker/module/memory, the worker device and the native window selected by the bridge presenter. These references and the epoch stay pinned after binding. Ambiguous GL/legacy owners, replaced ports/devices/Workers/modules/memory/windows/epochs, closure and async faults are rejected. Focused negative tests also retain the timing and USER-actor fences. Existing GL guard regressions pass.
The matching real pressure control binds with the correction. A new light control completed 60 presents and then consumed Up through native PeekMessage/remove on thread 1, slot 0, HWND 65537. Physical application was 127ms before the native receipt; the separate prearmed release process acknowledged UP. This qualifies the authored route, not NFS3 driving. The first pressure input cancellation remains retained and closed, not retried.
The genuine chooser and all 135 original media pins were rehashed unchanged. The additive binder JS file is the only runtime overlay on frozen c9 and module 7d347. The sampled observer remains unchanged.
Fresh original: closed Worker-inventory failure
After strict pull/push of 994aa3a20, the fresh original imported the unchanged
genuine chooser and launched nfs3_chosen_renderer through the public browser
launcher, normal guest Workers and software rendering. Its served graph used
frozen c9, module 7d347 and all 135 media pins, with only the published binder
JS overlay and the qualified observer prefix. The watchdog was armed before
launch. This was a new registered original, not the predecessor runtime.
The code passed its boot fence of more than one completed present (this
checkpoint was not separately persisted),
then attempted actual module recovery before accepting input. The harness had
retained every created CDP session and applied the authored control inventory
loop to them. Unlike the two-guest-Worker authored fixture, the game created
Worker slots 0 through 7; its console records slot 2 returning normally before
its native thread ID was reused in slot 5. Module inventory encountered a
closed session and its finally request to Debugger.disable raised
TargetCloseError: Session closed. The finally failure can mask the earlier
inventory error, so the exact first failed request and target are not known.
The later, unreached assertion also incorrectly expected the authored
fixture’s three modules instead of a dynamic game Worker inventory. Neither
assumption was repaired or used to authorize another original.
This is a failed harness lifecycle assumption, not evidence of an emulator Worker defect. The original did not produce a completed module inventory, reviewed whole frame, retained game-field report or physical input. Neither exact original executor identity nor producer/consumer/gate operands were qualified. No visible driving, steering or Pause was obtained. Native supervision exited 1 with both EOFs; Chrome exited 0 with both EOFs; the independent watchdog exited 0, with no leaked tracked process. The scope closed UNKNOWN and no later runtime endpoint request or replay followed. The conditional extra original was not eligible: no actual emulator defect was proved and repaired. It remains unspent, and no runtime is held.
Validation and closure
The new compiled field-reader test, existing scheduling-observer regression,
existing browser-native-input regression, fresh pressure pair and guarded
Device2 control passed their stated assertions. An initial locked build
stopped on another owner's active $run lifetime-census WIP. That owner
subsequently posted build-ready; the final locked build and publication
receipts are recorded in the run.
Every browser scope had a disk/time watchdog armed before native launch,
finite native supervision and actual native/watchdog/Chrome waits with both
EOFs and leaked:false. The cancelled pressure-input fixture and the closed
original each returned native status 1; the passing controls returned 0.
Final cleanup independently
checks all 158 known owned PIDs and local listeners, and absence of temporary
Chrome profiles. Free disk stayed above 1.5GB; no RAM gate was used.
Only selected named control whole PNGs are retained; Git plus own patches and
input/module recovery recipes replace source archives and media copies.
Box ID, expiry and stop receipt are null; state is NOT_CREATED.
The public-binder correction and tests are published in 994aa3a20; this
report records the same-task original failure and final cleanup. Ownership is
released after evidence sealing. No scheduling-cause or gameplay success is
claimed.