Forsaken capture composition diagnostic, 8 November 2026
The instrumented fresh original did not reproduce the command deadline failure. It reached a personally reviewed initial Volcano world/HUD using the unchanged fixed budgets. There is no validated causal repair. The previous failed Start remains failed, and its exact pending CDP operation remains unknown. The concrete blocker to claiming a repair is the missing old per-await trace plus non-reproduction in the one authorized diagnostic original.
Evidence: scratch/runs/20261108-forsaken-capture-composition-cause/.
The prior 20261108-forsaken-stable-startup-gameplay/ run was read only, never
replayed or relabeled. Its handoff matches commit 24afe9620 exactly.
No production emulator, renderer, input transport budgets or guest scheduling
changed. Only this note is committed; instrumented helpers and evidence are
run-local.
Actual await instrumentation and controls
The existing browser-run.cjs, numbered transport.cjs, bound-input chain,
production Worker fixture and native key helpers were reused. The trace records
entry/exit/error, parent phase, timestamp and elapsed time. CDP awaits include
method, session, target, script ID and expression kind/hash. State, outstanding
Worker-source promises, capture, before/screenshot/after owner checks,
admission, independent release, hold sampling and waits have phase boundaries.
The command deadline snapshots all active awaits without changing WORK=20 s,
ACK or immutable browser/cleanup deadlines. No guest exports, memory, clock or
scheduling were modified for instrumentation.
Local controls exercised the actual executed composition:
execute -> admit -> independently released hold -> settle -> state -> chain.capture -> before/screenshot/after through real production guest/render
Workers and the shipped f965 module, rather than a lightweight ACK substitute.
local-composition/: native Enter 750 ms + wait 1 s + complete capture succeeded. Controlled 3.5-second host receipt delays on actualDebugger.getScriptSourceresponses then exhausted the unchanged 21-second command work window. The deadline identified the post-screenshot renderer source await by exact session/target/script. There was no success ACK; follow-up input was rejected by sticky UNKNOWN. Both Enter releases occurred at about 750 ms. Already-started local work was allowed to drain and is not relabeled as success.local-owner-before/andlocal-owner-after/: two additional actual guest Workers and controlled 4.5-second delays on unrelated-thread identity reads reproduced a different composition problem. The old scan expired at 21 s; selecting the already bound target completed in 2.981 s, with the same state collection, module-byte proofs, renderer checks and screenshot checks. The candidate additionally checks the page's exact bound references before reading that target. It was qualified locally, not assumed causal.local-identity/: the candidate passed the reused missing-renderer, prebinding process replacement, endpoint/renderer/process-owner replacement, every reviewed identity token, stale-scene and trusted DOWN/UP controls. Negative controls dispatched zero input.release-control.jsonalso retains slow-sampler and injected-UNKNOWN release controls. The generated caption and white canvas were reviewed as fixture evidence only.local-final-preflight/: the final instrumented driver, including its disk observer, passed native key/type/combo checks without launching the original.
Two metadata-only page-target assertions are retained separately. The initial
trace labels session-less browser commands page, because both the requested
session and the not-yet-assigned page session were undefined. The exact page
session-to-target mapping was subsequently recorded in
retrieved/original/page-session-target.json. Worker target/session IDs were
exact throughout. These metadata assertions did not dispatch guest input,
change the command stream or produce an input UNKNOWN.
Original measurements and causal limits
The optional bound-target selection marker was never staged or enabled in the original. Its terminal audit proves this. The actual original measurements reject unrelated-thread scanning as the explanation for the observed expensive startup capture: those reads were near immediate. The dominant measured waits were on the owning guest Worker, not the renderer Worker or module transfer.
original-analysis.json pairs actual CDP entry/exit records. Startup capture
lasted 16.472 s:
| Actual phase | Duration | Dominant owning-guest await |
|---|---|---|
state() |
4.194 s | Runtime.evaluate CPU/GPU flags: 4.031 s |
| capture binding before screenshot | 8.192 s | Runtime.evaluate instance identity: 7.991 s |
| capture binding after screenshot | 3.762 s | Runtime.evaluate instance identity: 3.531 s |
The owning guest target is 453AA715701F9F69B765702AD332BC28, session
B242103A54A3E1272717CB4B42671A1D. Across this diagnostic, actual
Debugger.getScriptSource calls took at most 142 ms; screenshot calls at
most 1.233 s. Unrelated guest identity reads took at most 53 ms.
The owner identity evaluations took up to 7.991 s, and owner state reads up
to 5.288 s. This locates the expensive observed awaits; it does not establish
why the guest's inspector service was delayed, a renderer hang, or the exact
await outstanding in the previous uninstrumented failure.
The closest successful input composition was select-singleplayer at
17.849 s against wait 1 s + WORK 20 s. Its capture alone cost 15.464 s.
The diagnostic start used Enter 750 ms and wait 1 s, completed in 10.241 s,
and received exact ACK 7. Its independent UP began 750 ms after hold start;
the passive sampler finished later, as the release design permits. The
following wait-8-second passive capture completed in 15.702 s against its
28-second work window. No command or CDP deadline fired. All 10 numbered
commands, including ordinary quit, have exact success ACKs and no outstanding
UNKNOWN. None of this retroactively qualifies the previous failed Start.
A possible next investigation is coalescing redundant owning-guest observations within state/binding composition while retaining every identity and renderer proof. No such repair was implemented, bypassed, hot-patched into the live stream or claimed validated here. A second original was not authorized and was not allocated.
Personally reviewed native observations
capture-review.json binds each review to the image hash. Command labels are
not accepted as proof of their intended transition:
startup.pngandcopyright-continue.png: ordinary native main menu with Single Player highlighted. Elapsed review time after startup allowed the attract sequence to begin before the first command's resulting capture.singleplayer.png,dismiss-attract.png,menu-return.pngandintro-continue.png: automatic Acclaim/Probe-logo tunnels without player HUD. They prove neither player movement nor successful submenu/menu-return input.escape-long.png: original copyright/demo notice after bounded native Escape.select-singleplayer.png: Single Player Game, Start and Level Volcano panel.start.png: native Volcano briefing and partial loading bar.loading-passive.png: readable initial textured Volcano world, Lives 5, shield/hull 128, POWER 1, PULS 2000 and MUG 5. This follows native menus and loading, rather than an attract-logo frame. It is initial world/HUD evidence; no thrust, look, fire, ammunition consequence or enemy interaction was tested.
The diagnostic was closed after this passive observation because a causal repair had not been validated. Full 27, backend parity, FPS and audio remain open, as do movement/combat verification and the old deadline failure.
Identity, preservation and shutdown
Staged source is main efe297deaf4a3daae1d8fbd6beb05258babc5d58 plus
relevant-source.patch, new-lib_wininit-boot.js.patch and the owned input
patches. Git is the source archive. runtime-pins.json identifies all 1,487
staged files, unchanged after execution; 743 successful served responses
match their pinned bytes. All 873 existing original media references match,
including executable SHA-256
366ec5d0943daca839836296f9fb52724e94dc155630c6b0c30d7181106b5e24
and manifest SHA-256
587ce145ab73a3fed2c7136425bb6a7bcf28def312d430c8bf7d1be802b2260b.
No source/media archive was created.
The actual route was
index.html?candidates&no-log&threads&app=forsaken_demo&d3d-renderer=webgl,
Chrome 151.0.7922.108. Actual guest/render module SHA-256 is
f9650607181301b5d2a9990420f437bd9f1ca10d40ac73067339f07499d53121.
All ten state checkpoints retain owning legacy WebGL endpoint 1, actual module
and instance/context bindings, zero renderer error and zero GPU fallback.
The actual observed owning-Worker context is 640x480 WebGL 2.0, ANGLE / Intel
UHD Graphics 620 / Mesa 23.2.1. The wrapper adapter's direct .gl field is null;
context proof comes from the delegated native context creation and matching
owning-Worker/context token, not that absent wrapper property.
All 143 remote references / 132 unique contents, 12,492,635 reference bytes, were retrieved and independently size/SHA-256 verified before normal stop. Ten original PNGs pass chunk CRC and decoded-row checks. Eight holds have completed independent-release receipts (750–752 ms for the seven 750-ms holds, 1,999 ms for the 2-second Escape). Chrome exited 0. Controller and browser PIDs, HTTP/CDP ports, profile and writer lock are absent. All five local fixture/preflight browsers are also closed. Sampled free disk stayed above 5 GB; no RAM admission gate was used.
The observable detached 25-minute watchdog was armed before UUID-idempotent
allocation. Box bx_v5jx78db is archived, current expiry null,
original expiry 2026-11-08T03:05:23.351Z. Normal stop receipt
stop_28b68056fa285937 completed 2026-11-08T02:49:23.340Z,
forced:false, error null. Both lifecycle delegates exited zero and the
supervisor has no children. All 36 run-helper syntax checks and this run's containment/schema checks pass;
the repository-wide evidence checker retains seven unrelated existing errors.
No box was resumed or deleted, no second original
or software retry ran, and there was no public deployment. This worker stops
with the causal repair unresolved.