TR2 clear integration and transient publication accounting — 2026-11-08

The narrow synchronous viewport-clear observation repair is integrated into production and published as 0ac9414c6 before the one fresh original run. The original now produces source records instead of retiring at clear. It still has no qualified source/canonical/display join, so no Up, Space or Control is sent. The first new demonstrated integration defect is permanent discard of pending source versions on a transient publication writer. A substantive generic production repair and new real-WAT/canonical regression address that defect without a second original run. Qualified native gameplay, movement and weapon consequence remain open.

Evidence: scratch/runs/20261108-tr2-clear-integration-progress/.

Integrated clear repair and exact launch closure

The predecessor's synchronous-clear.patch, clear-observation-repair.cjs and test-clear-observation.cjs were read before editing. Board claims precede production/test edits. lib/d3d-render-worker.js permits only 0x20004 through the synchronous legacy-call retirement check. Shared-message writer exclusion and host-version invalidation still bracket the entire asynchronous adapter operation. Clear produces no source completion and cannot qualify an old proof; only the next fresh queued flip can establish a new source proof. Unsupported synchronous opcodes still retire observation and adapter throws still poison the epoch. No identity, byte, lifetime, publication or input gate is relaxed.

test/test-d3dim-clear-observation.js is a permanent discovered unit regression using the production dispatcher, actual WAT swaps and page watches, production journal/Join and actual canonical decode/composition. A reconstructed old retirement predicate reproduces the baseline. The integrated path yields raw 2 = qualified 1 + discarded 1, preserves exact adapter result and ACK, excludes readers through a delayed adapter, refuses the old proof and rejects cached republish as a new source. Unsupported/throw cases retain their prior behavior. The focused regression and locked full build pass before the explicit production commit and immediate authenticated pull/push.

Before allocation, all 596 pins, 558 tracked published blobs, 7 HTTP pins, JavaScript syntax and literal helper dependency closure pass. Exactly one runtime tracked file differs from the accepted source closure: lib/d3d-render-worker.js. Runtime is pinned to published 0ac9414c6fd62f6abb5a562973330b702d73601f; peer WIP and the subsequent accounting repair are absent from this original. CPU/render WASM remains SHA-256 6cf120d0d3b3db1823a8105eb26ef52a9093a7f871c37a46117475e213fda1c0. The driver preflight starts neither a browser nor a game.

Accepted native 500 ms preflight, delayed blank-Chrome readiness, full4096 actualWorker arm, queue/association/handoff/retirement and cost evidence are reused. The inherited prior-23-reference inventory discrepancy remains 4,379,955 actual bytes versus 4,380,437 coordinator-reported bytes (482 bytes); this run does not reinterpret either total. One unnecessary broader identity regression was started locally, then canceled during compilation with no result or acceptance claim; its empty log and cancellation record are retained. The focused clear and new accounting regressions are the verification for these changes.

One fresh native original

Fresh UUID-idempotent box bx_gmtfj5ce is identified idle before runtime. Its observable independent 25-minute watchdog and separate normal-stop delegate are armed before creation. Local launch disk is above 5 GB. Local wine-runtime.lock is held only for actual local integration fixtures, and released before all provisioning, remote supervision, retrieval and reasoning. Remote runtime owns its separate lock.

All six native blank-page keys use the accepted 500 ms preflight and pass. Chrome readiness takes 253 ms. The reviewed actual Graphics dialog has Hardware 3D Acceleration selected and default focused OK. Enter is admitted at 7,784 ms, within the unchanged 12,000 ms guard, and independently released after 504 ms. The resulting image visibly shows Lara in the Venice passage. Owning render Worker/endpoint and its separately instrumented WebGL 2 context are identified; CPU and render module bytecode hashes match the pinned module. ANGLE/Mesa Intel UHD Graphics 620 is observed. The generic adapter wrapper's null context field is not used as context ownership proof.

The full 4096-record live arm succeeds. The pre-input capture contains 14 raw, 0 qualified, 14 discarded records. One of these source proofs is sealed without versionRejected, but publication marks it permanently seen with uncovered-publication-writer. At normal retirement the observed terminal partition is raw 51 = qualified 0 + discarded 51: 46 source records reject an overlapping/poisoned writer and five have sealed source proofs. The discard ledger contains 42 uncovered-publication-writer and nine uncovered-scene-publication dispositions. These are observation counts, not FPS. Before observation, raw/qualified/discarded remain null/unobserved; no empty journal is used to infer a frame rate.

Four numbered commands have exact successful ACKs: Graphics Enter, arm, passive capture and quit. No original Up, Space, Control, UNKNOWN, replay, guest-memory write, private handler, forced scene/configuration/clock/seed or second original occurs.

First new defect and tested production repair

SurfacePublicationJoin.publication previously called account('uncovered-publication-writer') whenever its epoch was temporarily unavailable. That permanently moved every pending source into the discard ledger even if the writer touched unrelated watched memory and the source remained immutable. A later complete canonical read could never qualify that source because the key was already marked seen. The native pre-input ledger provides the concrete new integration symptom; the focused regression isolates this generic behavior from legitimate source-overlap rejections.

The scoped change in lib/render-worker.js keeps qualification blocked during the writer and calls ordinary account() to discard only concretely rejected, superseded or invalidated sources. An otherwise valid source stays pending. A later qualification still requires a fresh canonical read, exact owner and request, unchanged host/page proof, transfers and display publication. An unqualified cached decode cannot acquire a proof. Retirement still accounts for every remaining source. Clear's host-version invalidation and requirement for a next fresh flip are preserved by the clear regression after this repair.

test/test-d3dim-publication-transient-writer.js runs the production dispatcher, actual WAT swap/page watches and canonical compositor. An async observed writer performs a real WAT write outside the leased source pages. Publication during the writer refuses qualification. The reconstructed old behavior permanently loses the source; corrected production qualifies exactly one row only after a new full canonical read. Cached republish adds no row. Actual source-page write, host-version invalidation and poisoned-epoch cases all remain unqualified and close with raw 1 = qualified 0 + discarded 1. The positive case closes with raw 1 = qualified 1 + discarded 0.

The required callback-ownership census host-file digest/lifetime review is updated only for this file. Export/callback/nonlocal lifetime behavior is unchanged. The initial build's census rejection is recorded; the updated review and subsequent locked full build pass. No original qualification or movement is claimed for this repair. Remaining source-overlap/scene coverage still needs investigation in a separately authorized fresh run.

Evidence custody and shutdown

The initial immediate post-quit drain assertion sees the owned launcher and controller still exiting; this failure is preserved. A distinct later read-only custody check confirms both PIDs absent, Chrome exit 0, HTTP 35485 and DevTools 41961 closed, profile/writer absent, no matching owned process and all 596 pins unchanged. This is not a replay of a runtime command.

All 75 artifact references, 69 unique hashes, 8,060,809 bytes are retrieved and independently reread for size/SHA-256 before normal stop. Unique captures, logs and executing modules are retained. Initial expiry is 10:57:57.686 UTC; the independent immutable watchdog deadline is 10:57:57.518 UTC and the browser cap is 10:43:13.879 UTC. Normal stop receipt stop_b0ab3b6182a80cff completes at 10:34:56.919 UTC, forced=false, error=null. Final state is archived, stopped 10:34:56.894 UTC, active expiry null. Both lifecycle delegates exit 0 and the supervisor has no children. The box is not left billing idle.

Git is the source archive: exact SHA, relevant binary diffs, explicit owned new-file patches and media path/size/hash references provide recovery. No source/media archives, old-box resume/delete, deployment or subworker is used. Full27, parity, audio, game FPS and simulation FPS remain open.