Quake III actual snapshot causal latency, 2026-11-15
The predecessor stopped before native qualification at a foreign census mismatch.
Its retained evidence is scratch/runs/20261115-quake3-native-snapshot-latency/.
The continuation uses
scratch/runs/20261115-quake3-latency-after-census-repair/, referencing the
predecessor artifacts without copying their bundles. Current qualification and
original outcomes are recorded in that run's result after closure.
The original published with ac8fb9b43 / report 6a5983b78 stays permanently
closed. Its exact operations file hashes to
054638201db54bd9bd09128bbe0666295a51f1ed7871ebc1417939a12389462d.
Request 808 is the unchanged 47-byte snapshot expression and expired at the
fixed three-second deadline. Its valid object response arrived passively after
3605.703374 ms. There were zero postterminal requests, consent images or inputs.
Queue, body, inspector scheduling and native GL occupancy remain UNKNOWN.
Existing work and the missing evidence
Read the startup snapshot latency, fullscreen visibility, original public
visibility, original startup task, operation cause and source-inventory dedup
notes before designing this task.
quake3-startup-latency-controller.cjs already provides snapshot wrapper/body
timestamps and synthetic delayed queue/body/notification controls. Those earlier
controls and shader timings cannot supply missing measurements for request 808.
The retained task controller already wraps native/import/GL calls, with its
first sixteen and slow completed calls logged after return. Its note explicitly
acknowledges that a call that has not returned lacks completion evidence. The new
VM prototype reuses that exact instrumentation, announcing native entry before
every observed call up to the existing hard event limit, and separately records
the actual native start after the binding emission. Twenty consecutive fake
shader-status calls prove the entry remains available inside call twenty;
the old first-sixteen behavior cannot silently lose that pending interval.
Native duration excludes binding emission. This is a diagnostic design fixture,
not current Chrome or original evidence. More binding notifications have real
scheduling cost. Actual entry-emission intervals will be reported separately
before deciding whether to launch; they cannot bound all inspector perturbation.
The current observer's snapshot body itself only reads JavaScript counters,
context/owner maps and canvas dimensions; it contains no shader-status call.
Source reuse remains verified inventory equivalence, separate from actual
production script and WASM byte receipts.
The proposed addition must connect helper invocation/send/return boundaries, actual Worker snapshot entry/body/exit and ordinary synchronous GL calls under the same live target/session/unique-context identity. Native timing must wrap calls the renderer already makes, preserving receiver, arguments, return and exception; it must never add a GL query. Entry records must survive a hung call without depending on a later snapshot or post-UNKNOWN request. Sink delivery time must stay separate from the Worker timestamp. An entry without completion proves only an open observed call, not its ultimate duration or driver root cause.
The prepared composition preserves the actual production renderer and WASM bytes. It evaluates the retained task observer once before each paused Worker starts and installs no transformed renderer or host source. The exact 47-byte snapshot expression and the verified inventory implementation remain unchanged. Its helper wrapper records function invocation, immediately-before-send and send-return samples on one host monotonic clock, while preserving wire bytes. A passive listener records wire arrival before the existing helper handles the response. The helper's original three-second timer remains the timer of record. Broken diagnostic sinks do not issue requests or change returns.
Bindings retain live target/session/numeric-context/unique-context identities. Exact snapshot requests and snapshot entries are correlated by their per-context serial order; overlapping requests and composite expressions that could invoke snapshot repeatedly make correlation unsafe. Replacement, missing/duplicate samples, overflow and invalid native pairs cannot qualify. The first observable Worker JavaScript entry is the snapshot wrapper's entry, not an internal V8 inspector dispatch timestamp. Inspector-internal duration remains null.
The prepared native controls use the actual production render Worker and GL command transport with a finite synthetic clear/present and triangle. They add no game and no evaluation polling loop. Separate cases schedule a bounded synthetic native wait, place a bounded wait inside a fixture snapshot body, delay the helper's response handling after wire arrival, and briefly pause/resume the actual Worker inspector. These are independent calibration causes, never claims that the same causes occurred in an original. All intentionally late cases must remain terminal and send zero later requests. The inspector control records whether snapshot entry occurs before or after resume. It does not assume a paused Worker prevents inspector evaluation.
Independent controls must distinguish equal total delays before entry, inside the body and after exit; helper logging delay is a fourth interval. Missing, duplicate, out-of-order, replaced-context and incomplete records must refuse attribution. A returned heartbeat is insufficient. Inspector pause/resume and ordinary real Chrome GL controls must be measured separately from artificial Worker waits. Cross-process clock comparison retains explicit uncertainty; within-Worker body/native durations use its monotonic clock.
Predecessor scope and admission
Heavy work began only after both explicit TIMunsent426317 and FT467402 source/build/runtime releases. TIM's 96 named SHA receipts and 97 PIDs/groups, and FT's 148 named SHA receipts with 74 native/watchdog/supervisor-child closures, 297 PIDs and 223 recorded groups were personally checked. All direct actual waits have both EOF; all recorded PIDs and matching groups were absent. Both locks were personally observed free. Historical detached outer/ancillary gaps remain UNKNOWN. Only a new diagnostic tool, new test and this note are owned; existing renderer/helper/host and foreign WIP are preserved. One fresh original is conditional on materially useful diagnostics being qualified and published, with fixed 1800/1820 deadlines, 180-second Quit reserve, registered arguments/realticks/cooperative10000, immutable capture and independent finite UP unchanged. No original has been launched by this task.
Fifteen offline/VM checks currently pass, including an independent interval oracle for equal-total queue/body/delivery delays, pending call twenty, event overflow, result/exception/receiver preservation, missing and replaced identities, helper wire preservation and a broken sink. The proposed tool was retained only as an unapplied new-file patch throughout the predecessor's live work, then installed after release. The test now imports the tool directly and has no dependency on ignored patch files. Overlapping returned native intervals use their union, rather than adding durations twice. Duplicate/orphaned native pairs, negative duration and changed clock origin refuse attribution; a pending announcement alone has no duration.
Predecessor outcome: build blocker before native qualification
After the explicit releases and direct closure checks, the focused suite passed 57 tests with one expected native opt-in skip. The helper regressions passed, including finite independent key/Mouse1 UP, immutable capture, cancellation and irreversible UNKNOWN behavior. These remain fixture results.
The required unmodified flock /tmp/wine-build.lock bash tools/build.sh failed
at the callback producer census. The restored incoming foreign census contains
lib/ne-guest-bridge.js hash a66ad01d…, whereas the current shared file is
26c6b4f0…; lib/ne-lifecycle.js contains 377c6911…, whereas the file is
a99d2108…. This mismatch predates the new diagnostic tool. Updating those
foreign rows would violate this assignment's exact foreign-WIP preservation
constraint; no census or production file was edited to bypass the failure.
At predecessor closure the tool/test/note remained uncommitted and unqualified. No Chrome, native causal control or original was launched. There was no published diagnostic commit, current Chrome attribution, general render correction or gameplay claim. The one fresh original authorization remains unused. The next step was reconciliation of the two foreign census producer rows, followed by a passing shared build and the prepared native qualifications.
Continuation after the exact census repair
The coordinator refreshed only the two working-tree census hashes and released the passing unmodified full build at 07:51:51 UTC. Foreign code and all other metadata were preserved; mixed census metadata remains uncommitted. The new worker personally verified that repair, the predecessor's 92 artifact hashes, 19 actual waits with both EOF and 18 absent PIDs/groups, then waited for TIM and Sims to release source/build/runtime. TIM's direct closure and named waits, and Sims' 205 named actual waits with both EOF and 216 absent PIDs/groups, were personally verified before heavy work. Historical ancillary gaps remain UNKNOWN.
Two additional independent controls found that the analyzer accepted reordered
snapshot body samples and native end-before-begin records. Both controls failed
against the retained analyzer. The minimal correction requires the received
snapshot phase order and native begin/end order, preserving null attribution
for incomplete or ambiguous evidence. All seventeen offline controls pass with
the initial correction. Actual Chrome then exposed the retained wrapper's
redundant snapshot begin marker, which the hand-built protocol fixture had
omitted. That run was rejected. The analyzer now validates the complete emitted
five-phase order, and an additional positive VM control feeds actual
instrumentation records into the analyzer. This is a diagnostic evidence correction, not a reproduced
renderer defect or an explanation of request 808.
The continuation keeps the prepared real Chrome queue/body/delivery, actual WebGL and inspector-pause controls, actual source/WASM identity, finite native key/Mouse1 UP, immutable capture, visibility and irreversible UNKNOWN checks. It uses the existing qualification guards as a separate prelaunch precheck. Only qualified, published, materially changed diagnostics can admit the single unused original authorization. Fixed 1800/1820 deadlines, the 180-second Quit reserve, three-second operations and independent UP remain unchanged.
Qualified publication and the single original
Feature dfd7a8935 was published on main with strict pull/push before original
admission. The final focused suite passed 60 tests with one expected native
opt-in skip; helper regressions and the final unmodified locked full build
passed. Current real Chrome source qualification passed 16 checks, including
actual WASM bytes and verified same-context source reuse. Native qualification
passed finite key and Mouse1 gestures with separate acknowledged independent UP,
immutable captures, cancellation, fullscreen/viewport/ancestor visibility,
owner replacement refusal and irreversible UNKNOWN with zero later requests.
The host-only qualification consent image was personally reviewed.
All five causal Chrome controls passed with current prelaunch host/source pins.
The queue control measured 3393.921 ms before entry and 0.175 ms body; slow body
measured 3400.605 ms body; delayed delivery measured 3403.375 ms after completion.
Each intentional timeout remained closed. The actual idle Worker inspector
control required submitting its single snapshot before waiting for the paused
event: Debugger.pause pauses an idle Worker on its next task. After a real
pause/resume, entry followed resume, with 159.679 ms before entry. Internal
inspector dispatch remains UNKNOWN. Ordinary native WebGL control also passed.
Rejected runs retain the omitted-marker, idle-pause and newly discovered host
mapping failures. Five newly observed graphics libraries were pre-pinned before
the successful rerun; no renderer or WASM implementation was changed.
The one fresh original was admitted with exact current qualification pins, registered arguments/realticks/cooperative 10000, fixed 1800/1820 deadlines, three-second operations and the 180-second Quit reserve. The complete host consent image showed actual Quake III textured loading output and was personally reviewed. Its finite host gesture received independent UP acknowledgement.
The original's 325 exact 47-byte snapshot requests returned before the first
native-event overflow marker. Snapshot 728 returned in 2383.853 ms. The unchanged
qualified analyzer, applied offline to the immutable trace prefix strictly
before that marker, measures 2361.780 ms before Worker entry, 0.025 ms snapshot
body, and 20.710 ms after wrapper completion. Observed returned synchronous GL
intervals overlap 2347.189 ms of the enqueue-to-entry interval. The largest calls
were getShaderParameter at 1276.815 and 1090.590 ms. This is current-run
evidence of waiting behind observed synchronous GL work, with the analyzer's
two-millisecond cross-clock tolerance. It does not establish GPU execution,
driver-internal cause, a general renderer defect, or the cause of old request 808.
The full-run analyzer conservatively refuses native attribution after the
30,000-event limit and therefore reports all 325 as unattributed when given the
entire trace. The separately named pre-overflow analysis preserves the narrower
valid evidence without removing the marker or reopening the route. Later
native occupancy remains UNKNOWN. Actual source inventory retained one snapshot
source fetch plus 324 verified reuses, and both actual Worker WASM receipts
matched 62605b62… (2,217,109 bytes).
The original then closed on a different boundary: page-session
Runtime.evaluate 1113, expression SHA-256 80d3a42e85dd4d08558f618014f815a071d35cf954627708124e5fb085d5662b,
590,905 bytes, exceeded the fixed three-second deadline during passive startup
readiness. The last observation reported held/in-flight paint and ordinary
damage. No late response or cause for this page evaluation was established.
There were zero requests after terminal, no permanent game binding or game DOWN,
and no menu, arena, enemy consequence or native Quit claim. No retry occurred;
the single authorization is consumed.
At runtime release all 1026 source/input pins and 1394 host pins matched, 176 successful original HTTP responses matched their pins, 85 named actual waits had both EOF, 277 recorded PIDs and matching groups were absent, both locks were free, and the board watcher was stopped. Thirteen incoming foreign code/census paths were unchanged. Direct original native/watchdog/supervisor-child/launcher exits were 1/0/0/0. One superseded source-Chrome ancillary launcher receipt stays UNKNOWN after a filename collision; its direct waits were retained, and the successful repeated qualification has complete direct and launcher closure. Historical ancillary gaps remain UNKNOWN. No box was created.
Source/build/runtime and tool/test paths were explicitly released for the next worker; only this report remained for doc-only publication. The next concrete boundary is the page readiness evaluation and bounded diagnostic saturation. Further original work requires fresh authorization; request 808 remains closed.