TR2 native full-facade follow-up — 2026-11-08

The one fresh original run of published 4003f6502 still produced zero qualified canonical/DX display joins. It exposed a new integration branch: the native decoder prepared an unused GDI backing canvas, whose valid native client rectangle caused host scratch writes before decoding the source. That generic branch is now repaired and regression-tested. No original run of this new repair, movement, weapon consequence or FPS result is claimed.

Evidence: scratch/runs/20261108-tr2-full-facade-native/.

Exact original and newly observed cause

Runtime source was published main 0cfbac6573533c93bb24660dc359909a6e69a928, excluding peer WIP. The 615-file closure included byte-identical published lib/render-worker.js from 4003f6502, menu metadata WAT from f4fc82774, host queries from 9c93a0f13, and ancestor 2483eec51 post-fence repair. Actual CDP-retrieved CPU and owning render-worker modules matched d8facd0658b7f158c6e60bef2de0936471f438f8bf41e9240fd75635f1d319dc. The owning endpoint used actual WebGL2, reporting Intel UHD 620 through ANGLE. This original validates that module's execution, not successful publication.

The blank-page native preflight delivered six ordinary 500 ms holds with independent UP and trusted balanced events. Personally reviewed Graphics Hardware/Full Screen/640x480 High Color/OK admission was 8443 ms old; Enter released independently after 500 ms. The actual world was visible. Observation armed the existing bounded 4096-record journal; pre-arm raw remains null. Final retirement accounted for 534 raw = 0 qualified + 534 discarded: 241 source-version-overwritten, 292 coalesced-before-publication, one normal shutdown retirement. Every source sealed under a sole stable unpoisoned writer from beforeSwap through sealExit. No GDI substitute was counted.

native-all-writers.json retains passive production logpoint stacks: _presentNow -> _presentDxIfDirty -> presentBestDxOffscreen -> _presentDxSurfaceToMainWindow -> getWindowCanvas -> _computeClientRect -> wnd_client_screen_x/y -> get_window_rect. These observed native geometry exports and nested real host callbacks advance host versions before beginRead/find/cacheDxLayer; the actual layer had no version token or proof. Logpoints were removed after bounded observation. They invoked no guest handler and changed no guest state, clocks, seed, configuration or scene.

The earlier full-facade fixture had no nonempty WAT client rectangle. It therefore skipped this decoder-time geometry branch, even though it covered the later compositor geometry writes. The new fixture registers an actual native HWND and fills its native client rectangle through the generated region map. Exact published host-imports SHA-256 e1316adcc71654a6182f25e04316fd1fb5c8142361ae3f722eae32d94b71d5e7 reproduces pre-decode host version 0→4 and qualified0. The unchanged production dispatcher, WAT swap/page watches, observed facade, real scratch host, RGB565 decoder, canvas transfers and compositor run in both comparisons.

Generic production repair and tests

Exclusive DX layers now prepare their own canvas directly. They defer getWindowCanvas until a GDI fallback actually needs it. The unused GDI preparation no longer invalidates the source before decoding. Windowed client clipping keeps its existing preparation; failed layer allocation still obtains the GDI canvas. Native geometry exports and callbacks retain conservative writer observation. No getter classification or publication guard changed.

native-client-rect.json records fixed qualified1 with an actual red display pixel, plus a live-writer refusal and all 29 guards under nonempty native client metadata. The compositor still performs four real rectangle-scratch callbacks after decode; exact composition-local token admission from 4003f6502 remains necessary. Owner, request, pages, host freshness at read, poison, live writers, layer/canvas/token, transfer and unsupported-scene refusals remain intact. exact-baseline.cjs reloads the exact preceding Git host-imports blob and retains the expected qualified0 assertion failure.

Focused native-client regression, window-placement and browser Worker upload ordering checks pass. flock /tmp/wine-build.lock bash tools/build.sh passes (build.log); both module hashes remain unchanged. The previously accepted full-facade baseline0/fixed1/red/29 guards, seven focused tests and locked build were reused by hash, without standalone reruns. 190 retained references, 78 explicitly recovered retrieval paths and 23 existing media providers were independently reverified in accepted-controls-reused.json.

Lifecycle, retained failures and recovery

A detached independent 25-minute watchdog was observably armed before the UUID-idempotent single box creation. Overall/browser/cleanup deadlines were immutable. The accepted parent startup queue, subreaper and WNOHANG reaper were reused: remote /tmp/wine-runtime.lock held before provisioning and through launch, runtime and descendant drain. No local runtime lock was held for remote work; focused local execution released its lock before analysis.

The initial extra syntax pass treated ES modules as CommonJS. Its shell continued through allocation after that failure, so the final corrected 615-file closure check occurred after allocation, before provisioning or runtime. This violates the requested before-allocation ordering and is retained explicitly; it is not reported as a fully closed admission gate. The earlier 607-file source/media/helper checks had passed before allocation.

An observation command omitted waitMs and failed in the transport's budget check before ID creation/queue/dispatch. A new explicitly bounded observation was then submitted. The final capture response was truncated and remains sticky UNKNOWN in the local transport history: no capture or input was replayed, and the next command was normal shutdown only. Its unique artifacts and remote ACK were later recovered separately; this does not erase the earlier UNKNOWN. The first drain correctly refused the still-live Node controller; later settled drain passed. An initial audit incorrectly required beforeFence to equal post-fence state; its failure is retained. The corrected audit checks the actual beforeSwap/sealEntry/sealExit contract.

Before normal box shutdown, 95 unique artifact references (168302266 bytes) were retrieved and independently hash-reread, including five CRC-validated PNGs, exact modules, raw journals, command ACKs and custody histories. All 28 registered native PID/start identities, including adopted Chrome descendants, were absent/reaped; bootstrap and custodian were absent; ports, endpoint, profile, writers and held input were drained. Read-only remote helpers were parent-awaited and their structured exit receipts retained. Local normal and watchdog children exited zero; the owning supervisor ended with no children. final-audit.json records the independent checks.

New box bx_2b52t3h9 is archived, expiry null. Normal stop stop_734c8dd2681a7a69 completed at 12:30:50.932 UTC, forced=false, error=null. No box resume/delete, public deployment, second original or subworker occurred. Inherited boxes were untouched. Launch free disk exceeded 5 GB.

Recovery uses source-head.txt, runtime-pins.json, final-relevant.diff, small owned-file patches, exact baseline Git loader and existing media path/hash references. Git is the source archive; no source/media archive was created. The next authorized original must test this new pre-decoder repair and actual qualified DX joins before any reviewed Up/Space/Control. Other remaining causes, Full27, parity, audio and game/simulation FPS stay open.