NFS3 command-consumer scheduling investigation, 2026-11-18

No emulator scheduling, clock, or admission defect was proved. No runtime repair was made. No fresh original was launched, and visible driving and Pause remain unverified. Preparation ended at a concrete blocker: the normal-browser render pressure control completed only 48 of the required 60 presents before its finite 70-second preparation cutoff. That scope closed UNKNOWN with no replay or later endpoint request; its observed counterpart was never started.

Evidence: scratch/runs/20261118-nfs3-command-consumer-scheduling-cause/. The final result.json records publication, checks, identities and cleanup. This report does not reopen the original in report 44210b200.

Offline predicates and scheduling

The previous original's Up returned EAX=1 from 0x4c4fe0 at 0x4c4a70 17 times. Its mode and player selector were zero, simulation cursor 328 and sequence 332. Disassembly resolves the mode-zero destination to 0x4c5126: the signed comparison (simulationCursor + lookahead) <= sequence0 returns 1; otherwise it returns 0. Mode one first tests entry EAX: nonzero returns 0; zero falls through to that comparison. Modes two through four use a different branch. Initialization sets the mode-zero lookahead to 4. Lead 4 is consistent with that rejection, but the internal comparison's operands were not dynamically observed. This is an offline predicate explanation, not a new measurement of the exact taken instruction. The retained disassemblies and prior-gate-analysis.json preserve the distinction.

The previous Right packets were admitted, extracted and consumed by the actual selected-car path. Escape command 2 was present at counters 6562/6563 while the consumer cursor was 6496; approximately ten seconds later that cursor was 6617 and the 32-entry ring was zero. The exact 0x4c33e5 lag-greater-than-16 discard explains loss of those samples, but does not establish why servicing stopped. The 28 observed Pause decisions returned EAX=0. Gear, binding and hold duration are not established explanations.

The old native block observer changes dispatch by disabling chaining/uops and records roughly 25.9 million callbacks. Its ten-second gap cannot be treated as normal guest-worker performance. The new observer wraps actual native run boundaries and existing get_ticks/gpu_stream_kick imports; it never enables debug dispatch, invokes an extra guest poll, reads a read-and-clear sleep flag, writes guest memory, or changes time. Rows are bounded and sampled at 100ms. Aggregate run time includes synchronous host-import waits, so it is not CPU time. Sampling cost excludes wrapper, aggregation and scheduling cost.

GetTickCount, timeGetTime and QPC share host.get_ticks; QPC scales milliseconds by 1000 with a 1MHz frequency. They are not independent clocks. Browser workers read an atomic tick published by the page's GuestThreadHost, normally every 4ms and subject to page scheduling. CLI threads publish per batch instead. The host clock also deliberately holds time during runnable-worker boot and adjusts its origin afterward; startup mapping can further transform elapsed time. Consequently, a wall/guest delta difference alone is not proof of a clock defect. Scheduler/host RPC/worker waits were inspected offline; no defective branch was isolated. --d3dim-render-owner explicitly rejects CLI --threads, so the cooperative CLI path cannot qualify the normal browser guestWorkers route.

Compiled controls

Real compiled i686 PE controls create a native thread, exercise real USER and DirectInput, produce/consume a 32-entry ring, read real guest clock APIs, and render through Device2. No fabricated guest poll or injected game state is used. The pressure variant repeats drawing eight times and performs bounded native arithmetic. Frozen sources are Git c9d0c2cf720cc13b962ea8b9c2565de0b480d319; actual native module SHA256 is 7d347a6a631ab77335a14935a533feb896be71aa2508fc5aa19b1e5df8f096ce. Other agents' shared runtime edits were excluded. Retained C sources, unique executables, invocation scripts and identities provide recovery without media copies or source archives.

Native control Observer Guest elapsed ms Produced/consumed Max producer/consumer guest gap ms Max lag
Light off 2910 59/59 184/444 1
Light on 3218 59/59 171/511 1
Pressure off 20215 180/177 451/1603 3
Pressure on 13031 180/177 318/807 3

All four finished 60 frames and exited zero. Pixels matched exactly. The three unconsumed pressure records were the matching pending tail when the fixture finished; no divergent ring result was measured. Guest QPC elapsed agrees with guest tick elapsed to one millisecond. On-minus-off elapsed was +10.6% for light and -35.5% for pressure. These sequential measurements on a busy shared host do not isolate observer overhead; their opposite signs prohibit a neutrality claim or attribution of the earlier original's ten-second gap. Native observed main-thread maximum wall gaps were 419ms light and 581ms pressure. CLI termination does not flush the auxiliary worker's observer report; the guest counters therefore do not supply independent auxiliary wall scheduling.

The normal-browser light pair used actual guest workers and software rendering. Both reached 60 completed presents; all three inventoried native instances in each scope matched module 7d347…. Whole-page screenshots matched. Total launch-to-checkpoint time was 38043ms off and 32065ms on (-15.7%). Main/auxiliary observed maximum between-run gaps were 741ms/200ms. Main tick elapsed was 11153ms over 12418ms wall; auxiliary tick elapsed was 11459ms over 11459ms wall. Startup worker clock holding was not separately instrumented, so this is not a proved clock failure. Debugger module inventory occurred after measurement.

The browser pressure off control then reached only 48 presents by the cutoff. Its UNKNOWN report records zero later endpoint requests. No observed pressure case, game-field-reader qualification, or original followed. Earlier native cooperative/incomplete controls, the incompatible CLI option preflight, an overweight pressure fixture, and a sealed-loader preflight rejection are retained as failures and excluded from the passing pairs above.

Scope and completion

The new observer and focused regression are preparation tools, not an emulator repair. The regression checks unchanged native return values, exactly-once imports, bounded samples, read-only guest memory, no debug/read-clear calls and actual WebAssembly instance identity. Focused checks, compiled controls and a locked full build are required before publication.

The browser's existing guards are 500ms confirmation /15s game freshness, whereas this task specified unchanged 800ms/45s and independent 800ms UP. A coordinator board note at 03:10:36 permits retaining the browser guards. No guard was changed and no original input was issued. Regardless of that clarification, failed pressure qualification independently prevented the authorized original. The original allowance and conditional repair allowance remain unspent. Native USER/packet consumption in authored controls does not prove visible NFS3 driving or Pause. Ordinary game Quit was not attempted.

The original 135 media pins, genuine chooser and module were rehashed unchanged. Unchanged 70062/9191/0888 and observer proofs are reused only for their prior routing, async-owner, loader and observer claims. Finite native supervision, actual waits, both EOFs, independent watchdog receipts and owned-process cleanup are retained. No box was created; ID, expiry and stop receipt are null. Only two selected named whole control PNGs are retained. The task ends with the concrete preparation blocker, ownership released and no held investigation.