Quake III original startup task boundaries, 2026-11-15
Evidence destination: scratch/runs/20261115-quake3-original-startup-task/.
Diagnostic published as 29a19f34b; one fresh original completed. No production repair is justified.
The originals documented in the operation-cause and snapshot-latency notes stay closed.
Measurement route
tools/quake3-startup-task-controller.cjs composes the released original controller
in memory, retaining its three-second connection, registered launch, capture,
queued-input, owner, visibility, finite independent UP and irreversible Quit rules.
The diagnostic composition adds only installation of timing bindings, this run's
qualification destination and an export of the original browser entry point.
Exact source anchors fail closed. The original helper is not edited.
The diagnostic server serves the actual production d3d render Worker with narrowly anchored timing statements around initial script execution, its ordinary serialized message envelope, native renderer initialization and ordinary frame publication. Declarations retain their original global lexical scope. Both original and served source hashes are recorded before server start; the transformation is recoverable from Git plus the diagnostic tool. No source bundle is retained. Actual Manager, GL adapter and current interpreter module remain the production implementations.
The pre-start Worker instrumentation measures public importScripts, WASM dispatch and native GL calls. It preserves arguments, receiver, return value, promise identity and thrown exception. It issues no GL call or synchronization on its own and attaches no promise settlement handler, preserving unhandled-rejection behavior. The server instruments the existing d3d instantiate awaits and host.js compile awaits in place, without adding awaits or promise reactions; their actual source hashes are separately pinned. An asynchronous envelope is never treated as continuous native occupancy. No guest export is called by the observer; native initialization is timed only where the production code already calls it.
The exact 47-byte snapshot query remains unchanged. Its wrapper records entry, original body start/completion and wrapper completion, with the body samples emitted after the call. Binding JSON overhead is outside the original body. Metadata contains sequence, stage, clocks, endpoint/owner integers when actually available, target and session. There are no GL arguments, GL results, pixels, command buffers, guest-memory, source or media payloads in timing records.
The first sixteen asynchronous/message/script envelopes announce entry before executing. Synchronous native/import/GL call samples are emitted after return, so binding overhead is outside their measured call interval; a native call that never returns consequently has missing completion evidence. The first sixteen calls of each kind and later calls lasting at least five milliseconds retain pairs; counts/maxima summarize quicker calls. A hard 30,000-event per-context limit bounds output; the last event explicitly announces overflow, with subsequent drop totals unknown outside the target. The snapshot wrapper always attempts its samples, subject to that same limit. Incomplete or ambiguous evidence cannot establish queue/body attribution. Worker monotonic durations are separate from host/Worker epoch differences, which retain the established two-millisecond sampling tolerance without clamping negative residuals.
CDP operation metadata identifies request send/response, Debugger source requests and scriptParsed receipt. Inspector dispatch itself is not measured by those receipts and remains null unless separately demonstrated. No Chrome task trace is currently collected. A message/instantiate envelope overlapping send is not proof that a native call blocked inspector dispatch for the full envelope.
Gates and scope
Light preparation follows Sims native and Klotski modal work. The coordinator's subsequent request to place Win4 lifecycle work before Q3 was acknowledged while Q3 was still light. Heavy work requires both actual explicit source/build/runtime releases and personally verified waits, both EOF and absent recorded PIDs. The then-current module must be qualified; neither f866 nor c775 bytes can stand in for a different module.
Before the one authorized fresh diagnostic original: changed-route queue/body, Debugger, error/identity and passive-late controls, the complete queued native key/Mouse1/immutable-capture/UNKNOWN suite, focused tests and full locked build must pass, and the changed route must be published on main. The original checks both the released helper qualification and this tool's exact dependency hashes. Complete media, source, autoimports, TLB, fonts, resolved executables and native dependencies are pinned before launch. No original is justified by unchanged synthetic timing fixtures alone.
An UNKNOWN permanently closes the route. Only already-arriving events/responses may be observed afterward, with no CDP, evaluation, source-fetch or cleanup request. Independent host process teardown remains separate from native Quit. Diagnostic success, actual menu/arena review, gameplay/frag and native Quit are separate outcomes. No menu, arena, frag or native Quit progress is claimed.
Qualification before publication
The current shared module is SHA-256
6872c84daf073f76a4c2b7f1a54e2b3f7051129c1bc38d12ed03c145889dcaa4
(2,203,094 bytes), distinct from the clean committed Win4 module. Git SHA plus
binary patches preserve the actual shared source; neither module is bundled.
Seven diagnostic unit controls and twelve existing observer/latency controls,
the released helper protocol controls and the full locked build passed.
chrome-controls/complete-pins/ passed the actual current-module, transformed
Worker, Debugger/source-burst, error/identity, queue/body and passive-late controls.
Forty repeated exact probes completed alongside 41 ordinary publications in
1,524 ms, with maximum query wall time 8.129 ms. The synthetic queued native wait
measured 3,350.645 ms before wrapper entry and 0.290 ms original body time;
the deliberately slow synthetic body measured 3,400.775 ms. These calibrate the
new route and do not explain any original. No ordinary three-second overrun was
reproduced by these fixtures. Eight 20,000-call no-op calibration pairs measured
0.000640–0.001191 ms incremental observer cost per call; this is fixture overhead,
not an upper bound on production perturbation. Binding overhead is recorded
separately, and the diagnostic changes scheduling cost without promising zero
perturbation.
chrome-qualified/complete-pins/ passed the complete released native key/Mouse1,
independent UP, owner/freshness/visibility, immutable capture and irreversible
UNKNOWN controls with the new instrumentation and production compile awaits.
The host consent image was personally reviewed within its bounded window.
Both successful qualifications have actual native/watchdog/supervisor waits,
both EOF, absent observed PIDs, complete prelaunch native dependency pins,
hash-verified served responses and clean postflight source/module pins.
Two rejected qualification attempts remain explicit: the first timing run found 22 dynamically loaded native libraries missing from its prelaunch pins; those were pinned before the fresh successful run, without altering the rejected pins. The first full input run expired awaiting personal host review and shut down. Neither attempt qualifies an original. No original has been launched by a qualification fixture.
A further admission attempt was rejected before Chrome, HTTP or Q3 launch:
the qualification manifest omitted the released latency helper. Its rejected
pins remain in admission-rejected/. The manifest was corrected, both full
qualifications repeated with all dependencies pinned, and a separate admission
precheck passed before the sole actual original. Measurement code remained the
published bytes; no guard was relaxed and no rejected evidence was overwritten.
Fresh original: real queued native shader work
The actual original loaded the pinned production module in both guest and render Workers, and the actual transformed render source matched its pre-server hash. There were 443 exact 47-byte snapshot requests, with complete wrapper/body pairs and no event-cap overflow. No snapshot exceeded its three-second deadline. Historical request 1574 remains unexplained; these are new original measurements.
The longest snapshot, request 1405, took 2,303.894 ms from send to response. Its measured pre-entry interval was 2,293.790 ms (two-millisecond epoch tolerance), original body 0.025 ms, wrapper 0.065 ms, and wrapper completion to response 10.145 ms. Body/wrapper durations use the Worker's monotonic clock. The host send epoch and Worker epoch are sampled separately; these independently sampled components are not forced to sum to the host monotonic response duration.
Exact connection 281534:1, request/response operation sequences 3299/3302,
render target C161A91C48699A51A3C3FCAA6C829AB0, session
8F15EBFEFABE81CE0759F1FE7C3BE959 identify this result. The real serialized
message span 1989, endpoint 1, ran from epoch 1794714597015.240 to
1794714599319.620 ms. Request send was 1794714597026 ms. Its native GL
getShaderParameter spans 2002 and 2006 lasted 1,129.080 and
1,152.380 ms, followed by getProgramParameter span 2011, 8.945 ms.
The first native query definitively enclosed send beyond the clock tolerance.
The snapshot wrapper entered at 1794714599319.790 ms, just after the message
completed. This directly demonstrates delayed entry alongside substantial native
shader-query occupancy on this actual render Worker, with a tiny snapshot body.
It does not measure Chrome's internal inspector dispatch, or establish that every
millisecond before entry was occupied by those native calls.
The independently observed startup owner was process 1000, slot 0, window
65543, endpoint/render endpoint 1, epoch
f966cfe5-6374-4f68-87ee-b941a4243411; its GL snapshot had context 1 with
owner pair [0,1]. Native call timing rows themselves have null endpoint/owner;
the endpoint association comes from their enclosing serialized message and the
separate owner receipt, not a fabricated native-call owner field. Page compile
rows have the actual page session but null raw target; page-target.json supplies
the session-to-target mapping. Initial Worker parse and inspector dispatch remain
unmeasured. No Chrome trace was collected. No general production bug or safe
repair follows from two legitimate native status-query delays in this run.
Visible outcome and terminal closure
The complete personally reviewed host-consent screenshot shows a stone-texture startup image, not a menu or arena. Ordinary GL draw/present/transfer production was observed. Only the finite host fullscreen consent click occurred, with its independent UP and actual release-child wait. There was no guest gameplay input, permanent game binding, frag, arena progression or native Quit claim.
Before output selection, the owner receipt already reported an intersecting
painted CANVAS#touch-cursor box. After consent, publication 90 remained
pending held/in-flight paint and ordinary damage; the existing startup guard then
threw startup owner/output not live and fully visible. The precise final DOM
box was not captured, so the earlier touch-cursor observation is not asserted to
be the final refusal's complete cause. The guard was preserved.
The helper's subsequent failure screenshot request 1465, operation sequence 3420, hit the fixed three-second deadline at epoch 1794714623347 ms. There was no passive screenshot response before host teardown and no pixels from that request were recovered. Its internal Chrome/compositor cause is missing. UNKNOWN permanently closed the route: zero later CDP requests, source fetches, evaluations or protocol cleanup requests. Passive task bindings continued until independent host termination. No second original, rebind or replay occurred.
Actual Chrome, native, watchdog, supervisor, outer and host-release waits retain
both EOF and absent recorded PIDs. The shared source/build/runtime slot was
explicitly released; subsequent work must not reuse these pins for changed bytes.
No box was created (ID/expiry null, state NOT_CREATED, stop not applicable).
An early ancillary board watcher was stopped after orphaning and has no actual
wait receipt; that limitation is explicit and distinct from runtime child waits.
The later tool-managed board watcher was interrupted and actually collected.