TR2 retained-source native gameplay — 2026-11-08
The published e4238acfcc51ec4bb664f017cde3670966408e5b complete-source
handoff now has original-game validation and a personally reviewed new forward
movement consequence. A qualified publication consumed the actual owed
descriptor 3483180 / serial 525, worker slot 0, surface slot 6, front
131805376, retained generation 491. The source remained committed before and
after readback. All 307,200 decoded pixels matched its RGB565 bytes, and
all 397,485 visible output pixels matched the decoded layer through the
actual nearest-filter viewport: zero mismatches in either comparison.
Evidence: scratch/runs/20261108-tr2-retained-native/. This is native gameplay
validation, with no new production repair or FPS claim. Full27, parity, audio
and FPS remain open.
Source and presentation authority
Before the single UUID-idempotent allocation, the exact preflight verified
612 source/helper pins, 226 accepted control references and 23 existing media
references, with JavaScript/CommonJS/Python syntax checks. The production
closure changes only the four published host.js, lib/host-imports.js,
lib/render-worker.js and lib/d3d-render-worker.js blobs. Fresh physical
startup checks verified 33 owned helper files before allocation. Launch disk
exceeded 8 GB. Peer WIP was excluded.
The preceding repair's 49 artifact hashes were independently verified and its
accepted native actual-WAT decoder/compositor, 29 guards, 16 new negatives,
writer hold, source lifetime, transient writer/scene, producer identity,
Manager admission and locked build results were reused without reruns.
repair-controls-reused.json records those checks.
CDP actual parsed-source retrieval confirms the exact host, imports, both
render-worker modules and renderer hashes (parsed-source-verification.json).
Both actual CPU and owning renderer modules match
d8facd0658b7f158c6e60bef2de0936471f438f8bf41e9240fd75635f1d319dc.
The owning context is WebGL2 on Intel UHD 620 through ANGLE. One original
index.html?debug&threads&app=tomb_raider_2_demo&gl-renderer=webgl&d3d-renderer=webgl
was launched; no alternate route or original was used.
The production observation used its bounded 4096-record journal. The first
settled capture already contained four qualified retained-source rows, then
continued display reached 40 before movement. A bounded passive post-publication
logpoint sampled the actual completed composition, reading the current DX
token, its qualified row, the actual owed upload from the surface presentation,
the committed retained reader copy, the decoded layer and display.output.
It compared every pixel without calling a guest handler or changing guest
memory. The logpoint was removed. review/webgl/qualified-visible-pixel-proof.json
binds publication 86, cache token and generation 491 to the exact serial 525;
qualified-visible-display.png is the matching actual visible output.
No newest-record identity was substituted. Ring expiry, active writer, poison,
stale review, owner, scene and source lifetime guards were preserved.
Reviewed movement and continued display
After that qualified pixel proof, I personally reviewed before-forward.png:
Lara stood in the narrow brick alley facing the sloping path and lattice.
One ordinary ArrowUp hold of 2000 ms was admitted against that current
review and independently released. The native trusted keydown and keyup are
exactly 2000 ms apart; the independent release timer also measures 2000 ms.
The command ACKed. No further movement, weapon or scene input was sent.
I personally reviewed after-forward.png and settled-after-forward.png.
Lara advanced along the alley; the sloping ground and lattice/temple ahead
drew nearer and the wall perspective changed. The settled image retained the
new position after release, with continued ordinary idle pose/display. The
input proof records every checked key released. Qualified retained-source rows
advanced from 53 before movement to 61 afterward and 67 in the settled capture.
These counts establish observation continuity, not simulation rate or FPS.
Final retired accounting is 616 raw = 67 qualified + 549 discarded; all
67 qualified rows used retained sources.
Preserved diagnostic failures
Two stale Graphics-dialog Enter reviews were refused before dispatch, with zero effects and explicit ACKs. Fresh capture/review then admitted the only Enter that started the game; its requested 500 ms hold measured 501 ms. All 12 queued commands, including both zero-effect refusals and quit, ACKed. No input/capture UNKNOWN was replayed.
The first passive pixel helper had a circular diagnostic JSON object; its
known failure is retained. A corrected bounded sample correctly declined a
different current owed identity. An initial post-publication diagnostic tried
toDataURL on an OffscreenCanvas and recorded that error. The next compared
the 640x480 logical canvas against the 605x657 letterboxed output geometry,
giving 293,324 mismatches despite exact source-to-layer pixels. One console
subscription also recovered that older buffered diagnostic. The final helper
uses a unique event marker and the actual display.output; its current serial
525 proof has zero mismatches. These are retained diagnostic corrections,
not new native production defects. An early local image-open preceded capture
retrieval completion; the completed image was subsequently reviewed.
The initial local audit lacked Pillow; the final audit uses standard-library
PNG chunk CRC validation and verifies all 13 reviewed/captured PNGs.
The first drain found the controller still exiting after Chromium's normal exit. Its assertion is retained. A distinct subsequent process census found the controller absent without a signal, then the complete settled drain passed.
Lifecycle and recovery
An observable detached independent 25-minute watchdog was armed before create. Overall, browser and cleanup deadlines were immutable. The remote custodian held the remote runtime lock before provisioning through descendant drain, with subreaper adoption and explicit child reaping. No local heavy runtime or local runtime lock was used during provisioning, supervision or retrieval. The settled drain verified all 25 recorded PID/start identities, bootstrap, custodian, adopted descendants, profiles, ports, command writer and held input absent, with the remote lock released. Read-only diagnostic subprocesses were synchronously parent-awaited, with retained command exit records.
Before normal stop, 170 artifact references / 273,530,723 bytes, covering
150 distinct hashes, were retrieved and independently hash-reread. Seven
adopted child reaps are recorded. Normal/watchdog children exited zero; the
supervisor ended empty and all three local lifecycle PID/start identities were
absent. final-audit.json passes source, owed identity, pixel, input, command,
PNG CRC, artifact, drain and lifecycle assertions, with the done condition met.
Box bx_p462sgmc is archived, expiry null. Normal stop
stop_0d3764bc5c018692 completed at 14:04:37.851 UTC, forced=false,
error=null. No task box remains running.
Recovery uses the exact published Git/blob pins, source-head.txt,
relevant.diff, small owned helper patches, and existing media path/SHA-256
references. Git is the source archive. No source tree/media archive, peer WIP,
old-box resume/delete, public deployment or subworker was used.
Next work can build on the validated native presentation and released forward movement. Full27, parity, audio and FPS require separate evidence.