TR2 native progress and synchronous-clear observation repair — 2026-11-08
The one fresh original reaches visible gameplay after the corrected native-key preflight and reviewed Graphics default-OK Enter. Its 4096-record observation arms successfully, then the first new integration defect retires the source observer on the ordinary synchronous legacy route. No source/canonical/display row qualifies, so no Up, Space or Control is sent to the game. A narrow generic repair passes a new regression against the production dispatcher, actual WAT swap/page watches and canonical decode/composition. The repair is retained as an owned patch; production files and the original run were not changed or rerun. There is no second original.
Evidence: scratch/runs/20261108-tr2-qualified-native-progress/.
Exact inputs and reused prerequisites
Before allocation, all 596 input pins, JavaScript syntax checks and the
literal local helper dependency closure pass, including the corrected
native-preflight.cjs and browser-run.cjs from the d46f34034 run. The driver
preflight starts neither Chrome nor a game. The retained helpers are byte-exact
copies; only run-directory metadata is relocated. Runtime source is pinned to
published 098bda796b2ed168d500a2fd9f8e4e07abd8ec4e, with the accepted 558
tracked inputs matching 0e951a0be. The checkout SHA recorded before allocation
is 90c1fcfee91c4680877083e80498afcb73c9384d. Peer WIP is excluded.
The accepted real delayed blank-Chrome regression, five actual-Chrome native key cases, eight bootstrap controls, actualWorkers 4096 live-arm, queue-capacity, association, handoff, retirement/concurrent-arm and nine-process cost checks are reused without standalone reruns. This run adds only the new clear integration regression described below.
The prior 23 artifact references were independently reread locally, without an
old-box/runtime rerun. Their retained files and pins total 4,379,955 bytes,
matching the prior note. The coordinator separately reported 4,380,437
bytes. The 482-byte discrepancy remains unresolved; neither total is
silently substituted for the other. Paths, actual hashes and both totals are
retained in prior-inventory-discrepancy.json.
Git remains the source archive: source-sha.txt, runtime-pins.json,
relevant.diff, small owned new-file patches and synchronous-clear.patch
provide recovery. Media is referenced by its existing path, size and SHA-256
in media-refs.json; it is not copied into evidence. Unique runtime logs,
captures and the executing pinned module are retrieved without source/media
archives.
Native launch and first new defect
Fresh box bx_88dshc38 is identified before runtime. DevTools becomes ready in 3,681 ms. All six blank-page keys use 500 ms holds, retain individual receipts and produce twelve trusted, balanced, ordered browser events. Invalid durations produce no events. The existing gameplay physical bounds, independent release, late-ACK refusal, immutable deadlines and sticky UNKNOWN/no-replay behavior remain unchanged.
The reviewed native Tomb Raider II Graphics dialog shows Hardware 3D
Acceleration and the default focused OK button. Enter is admitted at capture
age 8,906 ms, within the unchanged 12,000 ms guard, held for 500 ms
and independently released. The resulting image visibly shows Lara in the
Venice passage. The one-way binding identifies the actual owning legacy WebGL
render context and exact CPU/render module SHA-256
6cf120d0d3b3db1823a8105eb26ef52a9093a7f871c37a46117475e213fda1c0.
The observed graphics strings identify WebGL 2.0 through ANGLE/Mesa Intel UHD
Graphics 620. The generic adapter wrapper's context field is null; the
separate instrumented context and binding are the ownership evidence.
The full-size 4096 live arm succeeds. Both armed and
qualified-before-input snapshots contain no source records, rows or discard
ledger. Guest slice and render command/fence counters advance between them;
these observations do not establish game or simulation FPS. A subsequent
read-only CDP observation of the owning render Worker identifies
active=false, reason=unsupported-synchronous-legacy-call, with the exact
owner/generation and journal capacity 4096. The page join's live=true alone
therefore does not establish source coverage.
The only production Encoder route emitting legacy-call is viewport clear
opcode 0x20004. The render dispatcher unconditionally retires observation
for that route. This identifies the integration defect from the actual
retirement status and exact executing source; the original message payload
was not separately logged. No private handler, export, guest memory write,
forced scene, configuration, clock or seed is used to progress the original.
One auxiliary read-only diagnostic initially fails at Node module parsing
before connecting to CDP; its error is retained and the corrected reader does
not replay a runtime command.
Four numbered commands receive exact successful ACKs: Graphics Enter, arm, passive capture and quit. There are no original gameplay movement/weapon commands or UNKNOWNs. An empty journal is not used to infer zero FPS or a qualified original count. Original raw=null, qualified=null, discarded=null (unobserved); no fabricated 0=0+0 partition is asserted.
Tested generic repair, confined to owned evidence
clear-observation-repair.cjs and synchronous-clear.patch permit only
viewport clear 0x20004 to continue an active observer. The existing shared
message wrapper still enters writer exclusion before the operation, invalidates
prior host versions and retains exclusion across an asynchronous adapter call.
Clear emits no source record. Only a later queued flip can seal a fresh source
proof. Other synchronous opcodes still retire the observer; adapter failures
still poison the epoch. No identity, lifetime, byte comparison or publication
gate is relaxed.
The new test-clear-observation.cjs checks the original and corrected production
dispatcher. It uses the pinned real WASM module, actual native DIB swap/page
watch exports, production journal/Join and real canonical decode/composition,
with a controlled asynchronous adapter. It is a focused integration test,
not original gameplay or actual-GPU throughput evidence.
The original path reproduces retirement. The corrected path preserves the
adapter result and exact ACK sequence, blocks canonical qualification throughout
the delayed clear, refuses the earlier source version, then qualifies the next
fresh source through canonical read, transfer and display publication. Cached
republish adds no source. Its observed partition is explicitly
raw 2 = qualified 1 + discarded 1. Separate unsupported-opcode and throwing
adapter cases retain their fail-closed behavior. Corrected-source syntax and
the focused regression pass. The full unchanged-main build passes under
wine-build.lock; this is not a claim that the candidate patch is integrated
into production. Ownership restricts changes here to the run directory and
this note. Integrating the patch into lib/d3d-render-worker.js requires the
next appropriately scoped worker and its production build/tests.
Custody and shutdown
The independent observable 25-minute watchdog and normal-stop delegate are armed before creation. Initial box expiry is 10:45:13.985 UTC; the independent watchdog deadline is approximately 10:45:13.799 UTC. The browser has its own immutable 10:30:40.175 UTC cap and command cleanup reserve. Local launch disk exceeds 5 GB. The local runtime lock is used only for the new local integration fixture and released on its exit; remote provisioning, supervision, retrieval and reasoning never hold it. The remote browser uses its separate lock.
After quit, the owned controller/Chrome PIDs are absent, both HTTP 39043 and DevTools 40223 ports are closed, profile and writer lock are absent, Chrome exits 0, no matching owned process remains and all 596 input pins remain unchanged. All 74 remote artifact references, 68 unique hashes, 7,814,479 bytes are retrieved and independently reread for size/SHA-256 before normal shutdown is requested. Normal stop receipt stop_1bc19d4e0dad724b completes at 10:23:39.266 UTC, forced=false, error=null. Final box state is archived, stopped 10:23:39.239 UTC, active expiry null. Both lifecycle delegates exit 0 and their supervisor has no remaining children.
The task ends with a tested substantive repair of the first new integration defect. Native movement/weapon consequence and qualified original joins still need a separately authorized fresh run after integration. Full27, parity, audio, game FPS and simulation FPS remain open. No old-box resume/delete, second original, public deployment or subworker occurs.