NFS3 real Worker timer observer qualified; original and control reach race
The production-Worker observer preflight now passes in real Chrome. The guarded original and its same-box timer-unobserved control both reach a rendered race. The earlier INITTIMER abort did not recur, so its cause remains UNKNOWN. There is no proved runtime defect, production change or timer-fix claim.
Evidence roots:
scratch/runs/20261112-nfs3-real-worker-observer-repair-1725/ and
scratch/runs/20261112-nfs3-real-worker-observer-repair-1725-original/.
Source anchor: main 9c8c8196e42e60a180401bb5b433e79e48974d77, plus the
relevant binary diff, new-source patches and run-helper patches. Git is the
source archive; no source-tree or media archive was made.
Endpoint diagnosis and helper repairs
The previous preflight stopped after
five seconds. This run compared the exact successful
Forsaken bootstrap, its emitted attachment
helper, full-bootstrap controls and all 26 Chrome-parent receipts. Unchanged
controls are referenced by path and SHA-256 in reused-controls.json.
The new direct-ELF control recorded the actual PID, command line, process state,
wait channel and profile throughout a bounded 45-second endpoint window. Chrome
remained alive, initially with pipe_read and an empty profile, then published
its complete DevTools endpoint 15.195 seconds after spawn. This disproves
permanent Chrome failure inferred from the old five-second timeout. It does not
identify the inner reason for that delay. A later Page.navigate became UNKNOWN;
that accepted operation was not replayed. Chrome ultimately exited 0, with both
output EOFs and an actual parent wait.
The working /usr/bin/google-chrome launcher was retained. Its actual emitted
process uses /opt/google/chrome/chrome with software-rendering, password-store,
language, fingerprint and timezone flags. actual-guest-worker-lifecycle/ records
the published endpoint, actual /proc executable/arguments, allowlisted
environment observation and profile contents. There is no unmodified-Chrome or
uninstrumented-timing claim. Every new local Chrome had an independently observed
90-second watchdog armed before spawn, plus actual Chrome/watchdog parent
waits, stdout/stderr EOFs and profile removal.
Fixable preflight errors were corrected and rerun on the synthetic fixture:
missing fixture broker handlers for process/thread sound lifecycle notifications
and exit, then a real observer defect: TextDecoder.decode rejected its view of
shared memory. The observer now copies API-name bytes before decoding. It uses a
private, pinned copy of the production region map and pure-JavaScript translator
for guest reads, including sparse translation, without loading duplicate lexical
globals or invoking guest read/translation exports from imported hooks. The
proven beforeScriptExecution pause, call-frame evaluation and acknowledged
resume path installs it before actual Worker instantiation. No early paused
Runtime.evaluate workaround or forced original routine was used.
qualification.json, all-parents.json and the qualified observed/baseline
results prove actual GuestWorkerLink, production Worker imports and current04
CPU-module instantiation; paired Sleep(0) and ExitProcess(0) arguments/returns;
observation of fixture stores to 0x562ae8 and 0x79a1a4; no observer error/drop;
and native fixture exit 0. Observed and timer-unobserved fixtures match the full
60 MiB guest-window SHA-256, known stores, exit, and six forwarded hook messages.
A deliberately throwing observer is rejected without resuming its script; the
failure and cleanup controls pass. This PE is only a fixture, not NFS3. Fixture
passivity does not establish identical whole-game scheduling.
Natural original initializer witnesses
Only after qualification, the final caller's actual data/import branch and two
missing-import negative controls passed against all 133 installed files, exact
original EXE SHA-256
0defab3eeb22ee4b6e0007a4d5b26a99d868008ba77e2b9bd3ef770e924548ad
and executing CPU module SHA-256
04c534192ee3fe39e388536c2ce710c12c47a5cd2ca4d95c19f9d00b1e8f0a17.
The independent two-hour box watchdog was observed before allocation. Remote
provisioning was actually joined with exit 0 and both EOFs. All 604 final inputs
and the actual data/import branch passed again remotely before either original.
The incremental helper revision adds an explicit timer-unobserved case and
Chrome EOF receipts; its final manifest is runtime-pins-v2.json. All 604 files
also pass the final reread after both browsers close.
Both fresh profiles used the same native catalogue installer, first setup exit,
reload, native renderer chooser, Direct3D Device 1 selection, original unsupported
card warning, setup OK, reload and registered original launch. Native setup alone
wrote its settings. Both page routes request Software and retain
mmTimerThread=true, asyncMultimediaTimer=false. No guest-state/configuration,
clock, RNG, AI, readiness or physics override was supplied.
The observed main Worker is target 58C693A76318FAA87226E0682BD85384, endpoint 1,
slot 0. Its complete first drain has 1,059 rows and zero drops/errors. It records
caller 0x4f6e2e → initializer 0x48dad0 with native EAX=0, then a later
natural initializer call with EAX=128. The retained paired API witnesses are:
| Native action | Return |
|---|---|
| First timeGetDevCaps / timeBeginPeriod(1) | 0 / 0 |
| First timeSetEvent(9, 1, 0x48d810, 0, 0) | timer ID 1 |
| Natural reinitialization: timeKillEvent(109) / timeEndPeriod(1) | 0 / 0 |
| Second timeGetDevCaps / timeBeginPeriod(1) | 0 / 0 |
| Second timeSetEvent(6, 1, 0x48d810, 0, 0) | timer ID 110 |
At 0x48dc7b, callback counters are respectively 2 and 1, compared with
saved EBX=0; execution reaches successful branch 0x48dc9f. Ready/active are
100 and 128 respectively, retries stay zero, and the failed-counter branch
0x48dc84 is absent from the complete main-thread trace. The clearing routine
runs naturally during the second initialization before the new timer succeeds;
this is not an abort-path clearing witness. Static disassembly and
timer-analysis.json bind the API return addresses and branch interpretation.
The observed follow-up frame shows Single Race, lap 1/1, 2 MPH, with an owning
Software render endpoint 1/generation 1 and completed draws. The fresh
timer-unobserved control also reaches a rendered race/countdown, naturally
choosing Hot Pursuit. Its CPU target is 4832487B0C60BCF00C9F91C5E1651271; module
bytes, policy and Software render owner are verified. It retains debugger/module
capture and the launcher flags, so it is not a debugger-free timing control.
Different native random race modes are not normalized or overridden. These are
two planned matched cases, not replay of an UNKNOWN operation.
There are 174,607 retained observer rows overall. Later long drain intervals cap two secondary-thread buffers: their maximum cumulative dropped-row counts are 161,787 and 60,262. The main Worker has 1,217 retained rows and zero drops/errors throughout. The first startup drains have zero drops; subsequent secondary callback histories are incomplete. There is no exhaustive callback/timing claim.
Recovery, disposal and remaining causal gap
evidence-check.json verifies 203 recovered run paths, 90 provider logs, 26 whole
PNGs with chunk CRC/full IDAT inflation, and 26 command ACK/review chains. Every
retrieved remote path was independently fetched twice and hash-checked. Executing
module bytes and large logs have exact gzip recovery maps. No original command
became UNKNOWN and no accepted operation was replayed.
There were zero driving inputs. One balanced Escape pulse opened the native
paused menu; its helper had finally release but no independent release witness,
so it is not promoted into qualified driving evidence. A native Quit mouse click
did not exit the game. Both browsers were then normally closed; there is no
native guest ExitProcess claim for the originals. Chrome, controller and owner
PIDs are independently absent, profiles removed and server ports closed. Both
Chrome parent waits record exit 0 and both EOFs; both controller owners, the local
manager and independent watchdog have actual successful waits and EOFs.
After hash-verified recovery, delegated normal stop completed at
2026-11-12 17:35:10.209 UTC, forced=false, error=null. Independent final state:
box bx_prdn62ye ARCHIVED, expiry null, receipt
stop_5f65f1af023db95d. No preexisting box was resumed, stopped or deleted.
Actual /home samples remained above 1.5 billion free bytes; no RAM gate was used.
The earlier original abort remains valid failure evidence. Its causal blocker is now specific: a failing natural initializer's API returns, counter comparison and actual clearing/abort branch are still missing. This run supplies successful whole-original startup witnesses and a qualified observer, not a retroactive cause for the earlier failure. No runtime fix, build, repaired-original retest, purposeful corner/lap, FPS, audio, parity, full27 or broad crash-free claim follows. No workers or public deployment were launched.