TR2 current native control and presentation-rate attempt — 2026-11-08

The attempt establishes no current WebGL controls or presentation rates. The initial click was rejected locally for its coordinate format, and the final idle image lacks the sheet. A coordinator raw-pixel/hash review corrects the original worker assessment: startup, fresh-graphics and refresh-graphics all show the Graphics sheet; only native-after-idle shows the desktop. Persistent absence and an emulator composition defect are unproved. Metadata still lists the windows visible. Original mistaken review records remain preserved.

Evidence: scratch/runs/20261108-tr2-current-control-presentation-rate/. Only this note is committed. The previous TR2 gameplay evidence establishes bounded displacement, heading and weapons on its earlier module; it is not current-module control acceptance. The latest startup sweep ends at native Graphics and has no actual graphics backend.

Original route and blocker

One original used the unchanged registered TR2 executable/media and production index.html?debug&threads&app=tomb_raider_2_demo&gl-renderer=webgl&d3d-renderer=webgl. There were no guest/config/clock/seed edits, map changes or private input handlers. The browser runs at ordinary clocks, without stepping or freezing.

The original has four exact complete ACKs: three input-free captures and normal quit. There is no original mouse/key admission, UNKNOWN, partial action, replay, second launch, Venice scene or actual render endpoint. Browser preflight on about:blank qualifies five trusted balanced native key pairs and invalid-hold zero-effects checks. Local chain controls pass one-way native-to-render binding, owner replacement rejection, stale zero effects and independent key release; atomic transport controls pass exact ACK and sticky UNKNOWN handling. These controls do not replace actual gameplay qualification.

Counter qualification gap

The production pipeline and existing tools/measure-gameplay-presentations.js were inspected. A queue total, rAF cadence or screenshot count is not game FPS. The available stages have different meanings:

Stage Production boundary and remaining distinction
Queued submissions D3D stream flip records and batch sequences precede owner completion; batch sequences also cover non-frame work.
Completed GPU readback/flip lib/d3d-render-worker.js awaits adapter.fenceFlip() before d3dim_worker_flip; its completed batch sequence is written after replay. The flip has a record ordinal within that batch. A readback total alone does not identify its later page publication.
CPU/DIB uploads GPU _upload and page surface upload are distinct operations. Neither counts unique logical game updates.
Guest-present notifications lib/host-imports.js calls _noteGuestFrame/WinePerf.guestFrame; host.js also forwards DirectDraw dirty/present notifications. These need route-specific duplicate/completion qualification.
Page compositions Renderer observePresentationPublication(display, sequence) runs after ordinary composition. Its sequence identifies a publication, without an attached owning guest flip identity.
Content changes A publication can repeat the same content. No full-content comparison was qualified here; screenshot or sampled-pixel cadence would not settle this.
Logical guest updates No independently verified TR2 logical update marker was established.

The existing sampler validates monotonic telemetry, module/runtime identity and visibility, then differences presentsTotal and retains pacing timestamps. It does not join submission (endpoint, batch, ordinal) to completed readback, guest present, upload and page composition. Its local unit tests pass. counter-gap.cjs additionally exercises the actual released measureWindow function with synthetic data: two notifications for potentially one frame count as two, and a notification delayed into the next interval counts there. These are counterexamples to interpreting that tool as a unique-frame collector, not observed TR2 duplicate/delayed completions or game rates.

The required live collector qualification cannot proceed without an actual owning graphics endpoint. Completion duplication/delay, owning-frame clock association, content-change discrimination and observer overhead remain unqualified. No collector was armed, no three warmed 10-second gameplay windows were collected, and no interval median/p95 or rate is reported. Queued, completed, composed, content-change, logical FPS and physical scanout rates remain unknown. The optional matched software D3DIM run was not attempted.

Runtime identity and evidence checks

Staged published-main source is 0a38cf48438096482d059c5d8f27c048474c83b9; relevant worktree diff is empty. Runtime pins retain individual Git blob identities and existing untracked-source/helper patches. Actual guest Worker B5D5871013212A1FF7DFB33873FD3089, endpoint 1, slot 0, yields Wasm SHA-256 7ec49cbbae4cb177909fd8fd9d3599713ee5f2a5ffb7f7e8c890848a540052bf, 2,182,452 bytes, matching served bytes at all four checkpoints. This existing module was reused; no fresh build is claimed.

Chrome is 151.0.7922.108, V8 15.1.206.16. Native screenshot dimensions are 945×889; guest canvas 640×731, CSS rectangle 605×691 at (10,160). All 596 staged input pins remain unchanged in the cleanup checks; 173 successful HTTP reads and four recovered host-source reads match pinned bytes. Four non-200 HTTP responses and Chrome's environment diagnostics remain raw. All four original PNGs pass chunk CRC and IDAT inflation. Media are referenced by existing paths and SHA-256, including the unchanged registered three-path TR2 closure; no source tree or media archive was made.

Owning render module/context/backend are not-created. Adapter errors, fallbacks and GPU completion counts are unavailable, rather than zero. Physical GPU is unknown; the requested WebGL selector establishes no physical hardware identity. No exact page-to-OS-renderer PID mapping is claimed.

Retrieval and delegated stop

A persisted fresh UUID Idempotency-Key created only bx_arw2e4cn. Independent normal/watchdog delegates were observably armed before allocation, with the unchanged 25-minute deadline 06:12:55.394 UTC; initial box expiry was 06:12:55.578 UTC. No old-box resume, delete, nested worker or public deployment occurred. Sampled local free disk stayed above five billion bytes; no continuous minimum or RAM gate is claimed.

Chrome exits normally at 05:49:47.411 UTC, code0/signal null. Both cleanup checks show HTTP/CDP ports refusing and profile/writer absent, but the controller PID is still present. Full process drain before stop is not proven. Both failed assertions remain in evidence; no software launch followed them.

All 63 remote artifact references, 56 unique payload hashes, were retrieved and independently reread for size/SHA-256 at 05:50:00.856 UTC, before normal stop. Unique captures, logs and executed modules are retained. The independent normal-stop delegate then requests shutdown. Final state is archived, active expiry null, stopped 05:50:02.610 UTC. Receipt stop_c17945f526512ba2 completes 05:50:02.638 UTC, forced=false, error=null. Both delegates exit0; supervisor children are empty. The box is not left running.

Next work uses fresh original-resolution/hash-aware review and a correctly qualified normalized native OK action (or native default-button key), then ordinary Venice controls. One missing idle image does not justify assuming a persistent composition defect. A collector must prove frame association and overhead before three ordinary-clock windows can answer the owner rate request. Full27, backend parity, audio, full-level completion and logical FPS remain open.