TR2 observation coverage repair — 2026-11-08

Repaired the measured observer exhaustion offline. The published observer's lifetime allowance was spent on checks surrounding ordinary basic-block execution, rather than on the requested native hold duration. It checked before and after each run(1) and counted even unchanged PC/ESP checks. The actual original reached 4097 checks after 32.105 ms, with zero records, during a requested 2000 ms / recorded 2001 ms native hold. That establishes insufficient observation coverage, not absent guest polling or defective Control.

Evidence: scratch/runs/20261108-tr2-observer-coverage-1733/ (R). result.json records the exact starting Git SHA, relevant binary diff, tested paths, selected manifest and existing media hashes. Git holds source history; small owned patches and helper transition patches support recovery. No source archive, media copy, peer WIP, game execution, Chrome, remote allocation or replay was used.

Cause and bounded strategy

src/13-exports.wat:$run takes a block budget; src/04-cache.wat chains blocks while that budget remains. One block can contain many instructions. Folds can also retire multiple blocks on run(1), as explicitly reported by get_last_run_blocks. Internal WAT callbacks/reentry remain inside the call and are not exposed as JS observation boundaries. The existing JS adapter cannot claim instruction coverage or other-thread coverage.

The repaired adapter observes contiguous ordinary run(1) returns until its existing 2500 ms installation deadline, explicit removal, failure, or first complete poll/Action pair. It does not replace 4096 with a larger lifetime count. Its work is bounded by the original caller's remaining signed-i32 block budget, with a deadline check at each observed boundary, and at most 32 retained records. It subtracts actual retired blocks, including fold overshoot, rather than the old one-per-call debit. A zero-progress result removes the observer and forwards the unspent ordinary budget once. Zero/negative budgets and entry yield servicing are forwarded rather than suppressed. After removal/expiry it forwards the remaining ordinary budget and preserves aggregate retired-block reporting. No additional guest budget, guest setter, debug flag, translator/materialization, callback invocation or private guest write is introduced.

Installation still prearms the Worker expiry timer. The ordinary helper now supplies a one-use activation callback to its existing passive sample, after native keyDown, with the independent native release timer already armed. Before activation, ordinary calls pass through unsplit and consume no observation allowance. Activation does not reset or extend the installation deadline. The first complete paired poll/Action removes the facade immediately with status witnessed; this bounds work and avoids spending the record cap on repeated polls after the needed witness. A failed HRESULT, zero Control bytes or clear Action is still a raw observed outcome, not a fabricated input defect or firing success. Final correlation still requires a physical VK17-down sample wholly within the unchanged native interval and an in-interval poll/Action pair. Out-of-interval, unarmed or missing-physical results fail qualification without retry.

Owner, thread, endpoint, EXE and loaded-code checks, frozen direct-window reads, call-stack pairing, RPC-publication validation and Action successor checks remain. Expiry, record overflow, bad code/stack/read, replacement, trap and incomplete pairing fail closed. A stalled WAT/host call cannot run a Worker timer until it returns; the observer cannot preempt that call. Independent native keyUp and the enclosing runtime's external lifecycle deadline remain necessary. No Worker timer or observer response controls native hold duration.

Tested causal reproduction

node R/causal.cjs retrieves the published observer from Git 16919b487 in memory and runs it and the repair against the same authored harness on exact CPU SHA-256 d8facd0658b7f158c6e60bef2de0936471f438f8bf41e9240fd75635f1d319dc. A three-byte INC/JMP loop consumes 12,000 ordinary blocks over twelve caller invocations. The published observer restores at 4097 checks with zero rows. The repair remains active past that point and then witnesses both the pinned poll-return and paired Action instructions, removing qualified with EBX 0x40. Both retire the same busy-loop guest budget. The harness deliberately transfers to the immutable instruction windows with authored registers/stack and a tiny bridge; it does not load or run the PE. Its 110 ms clock is a deterministic fixture clock, not an original wall-time performance measurement. causal.json retains both full reports.

node tools/tr2-action-observer.test.cjs: 32 checks pass. These exercise the actual serialized observer, exact static EXE windows, RPC bitmap layout, real CPU coverage harness, fold/no-progress/zero-budget/yield-entry contracts, successful/failed/retry polling, zero DIKs, binding/conflict, Action set/clear, no-poll, bad caller/code/publication, owner replacement, overflow, expiry, deferred expiry, pending pair and removal. The first-pair stop does not append after removal. Actual unchanged native hold execution verifies in-interval/physical correlation and negatives. Both a stalled sample and a stalled observer-activation response resolve only after independent keyUp, within the original hold bound. Fixture checks establish no original-game or backend qualification.

node R/helper-integration.cjs executes the actual selected helper's extracted performHold expression against its unchanged native hold executor: activation occurs exactly once after keyDown and release remains independent. Normal helper resolver validates all 29 helpers and 615 provider targets. The actual selected driver's DRIVER_PREFLIGHT=1 preloader validates all 615 physical provider hashes and exits before Chromium. Production static-server resolver passes all nine HTTP pins; 13 unchanged helper-closure checks pass. Direct and RPC production input-path tests each pass all nine checks. flock /tmp/wine-build.lock bash tools/build.sh passes, preserving exact d8fac CPU bytes.

Selected closure and remaining scope

R/prepare.py uses normal runtime_helper_closure.relocate from scratch/runs/20261108-tr2-action-observer-1652/runtime-pins.json (SHA-256 1b931b9a4f9b1266651575a9d79ec6e76cd09c44bdbf5d9c7655595194f8d00b). It relocates the same 29 small physical helpers, changes the browser helper only to call the activation callback from its existing sample, and explicitly repins the repaired observer provider. All other 585 non-helper providers remain selected bytes. runtime-pins.json:selectedGeneration retains each transition. The historical 614 generation and predecessor 615 manifest are unchanged. Existing media is referenced by path/hash, including TOMB2.EXE 0811efa84bfab8c458b9e29976088152db4c14601a99ca4910c75a2d009233aa.

Actual CPU/owning WebGL2 authority, bindings, native inputs and their policy, guest clocks and seeds remain unchanged. No weapon, aim, backend, after-release consequence, full27, parity, audio or FPS qualification is claimed. Internal WAT reentry and other threads remain explicitly unobserved; missing rows cannot establish absence of execution. Next is a separately authorized original using this explicit selected generation and its normal independent release/lifecycle bounds. No original is authorized by this preparation result.

No task box exists. Box ID/state/expiry/stop receipt: N/A. No allocation, resume, delete, public deployment or subworker. This single-task worker stops after tested publication.