Current original TR3 native WebGL Jungle actions — 2026-11-08
Ordinary native welcome → setup → Game → New Game reached Jungle on an actual owning WebGL context. Space2000 visibly drew the pistols. Persistent translation and heading remain unproved: the run stopped when shared free disk fell below the owner's 5 GB requirement. Software D3DIM was not started.
Evidence: scratch/runs/20261108-tr3-current-native-gl-gameplay/ (ignored).
This is a new current-module run. Earlier movement and armed poses are separate
prior evidence. No demonstrated emulator defect or implementation repair follows.
Route and reviewed actions
The unchanged registered ten-file TR3 fixture ran in one fresh browser on one UUID-idempotent fresh box. Both ordinary host renderer queries requested WebGL, with production guest Workers. No guest/config/save/clock patch, forced map, private handler, console command, source/media archive or public deployment was used. Native welcome OK and setup OK were dismissed before a renderer was required. Setup retained Primary Display Driver, Direct3D HAL, Hardware Acceleration, 640×480 High Colour (16 Bit) and ZBuffer.
Fresh raw PNGs were personally inspected at original 945×889 resolution. All 17
retained PNGs pass chunk CRC and full pixel decoding; 16 fresh captures were
opened, while the initial startup.png was superseded by the reviewed welcome.
The welcome and setup PNG hashes differ:
welcome-reviewed.png:845c3957843ff2f42574ada2e6641069a404b56e80abf420f95e313201f6fac2.setup.png:d383c6bd1376fd151d2c352810c9d5283d5d0ddeb3713e133f270bf42a0898a3.
The sidecars bind each review, timestamp and PNG SHA-256 to the actual owner. No unchanged screenshot was interpreted as a disappeared dialog.
The open passport visibly said New Game before the second balanced Enter300.
jungle-before.png shows the Jungle starting slope and Controls overlay,
including Up Run, Left Left and Space Arm. Short Space100, a later distinct
Space300, Up100 and Left1000 completed but did not establish persistent
terrain-relative movement, heading or weapon draw in the reviewed images.
These short holds are not accepted as gameplay effects merely because they
received ACKs. During Left1000 the host reported Left physically down; passive
samples showed sparse guest/render progress. That is not a logical-update or
keyboard-poll measurement.
A separately reviewed, safe stationary scene preceded balanced Space2000.
long-arm-after.png, captured after four seconds without further input, visibly
shows raised pistols and the Controls overlay gone. This establishes a visible
weapon draw. The planned additional idle-persistence capture was refused locally
before command publication because free disk was below 5,000,000,000 bytes.
No extra persistence, firing, target consequence, translated location or changed
heading is claimed. Firing was not attempted.
The first sampled free space was 5,182,377,984 bytes; the later sample was 4,481,933,312 bytes. The assigned run itself occupied about 7.7 MB at the block. The shared-space drop's cause was not attributed. No peer evidence was removed. Only necessary closeout and evidence retrieval followed the failed local gate. This resource requirement is the concrete terminal blocker; it does not prove an emulator input defect. Restoring disk above 5 GB and performing a fresh bounded run is needed for current movement/turning and paired software work.
Actual execution and ownership
Published staging source: 4c922e3ade224f6d11e2b7df25cc2516118c1f20.
Tracked runtime files came from individual published Git blobs, excluding peer
WIP. runtime-pins.json, the excluded worktree diff and retained new-file/helper
patches preserve reconstruction without archiving a source tree. Existing
standard-module identity and source identity are recorded separately; no fresh
build or source-to-module equivalence is claimed.
Actual guest and render Worker Wasm is 2,182,404 bytes, SHA-256
f5cfb1c6dea588eb538831d42069c699f6b7a6306cb530b8eb0abb437af754a7.
Recovered live bytecode at all 45 recorded module/checkpoint references matches
that pin. All 189 complete successful HTTP responses and four recovered loaded
host scripts match the staged bytes. All 602 staged pins remained unchanged at
final drain. Chrome was 151.0.7922.108, preserving its native viewport, with a
640×480 game buffer.
The native-input stage records actual process, document, canvas, renderer, guest Worker instance/target and module continuity while no render endpoint exists. It then binds one way to the actual D3DIM endpoint, render Worker, module, epoch and observed WebGL context. Native-stage input does not wait for graphics creation. Binding forbids return to native-stage admission.
Guest target EFD39A6A92A6E0BF2DEEB76AB88826D2, endpoint 1, slot 0,
reports GPU enabled. Actual render owner
3D47E6681C62BB9608F58874910FE136, endpoint 1, is legacy/webgl, with one
observed owning OffscreenCanvas WebGL2 context and healthy shared control.
The context reports:
ANGLE (Intel, Mesa Intel(R) UHD Graphics 620 (KBL GT2), OpenGL 4.6 (Core Profile) Mesa 23.2.1).
This is an observed renderer string, not an independently measured physical-GPU
claim. Direct owner/module/context evidence and advancing production draw calls
establish actual WebGL execution; toolbar selectors alone do not. Recorded
Jungle checkpoints retain errors=0, fallbacks=0, native software completions=0
and GPU-fallback software completions=0. analysis.json retains complete raw
counter/identity references. No FPS, audio quality, backend parity, physical
scanout, full27 or complete-playability measurement was made.
Guards, retrieval and shutdown
Local chain controls passed native-before-graphics input, exact owner continuity, one-way binding, stale zero-effect refusal, replacement-owner rejection and independent key release despite slow sampling. Actual trusted blank-page native Up/Left/Space/Enter/Escape key pairs passed before the original launch; invalid holds produced zero input events. Physical holds remain 50–2000 ms, and the scene guard remains 12 seconds. The released numbered ACK transport, native mouse helper and actual Worker/module recovery were reused. Local transport controls also passed atomic publication, exact ACK identity, duplicate rejection, stale-refusal capture-only recovery, sticky UNKNOWN closure and writer-lock release. The synthetic UNKNOWN was confined to its local test fixture; it was not an original-browser command.
All 20 original commands have exact complete ACKs. Three were stale refusals before dispatch, effects=0, followed only by fresh captures and new IDs. No UNKNOWN, partial action, consumed-ID replay or second backend occurred. The local post-refusal fetch helper initially attempted to retrieve nonexistent post-action captures twice; those failed read-only fetches are retained as helper errors. They are not unknown input outcomes. Later helper handling recognizes zero-effect refusals. The first short Space result's filename contains “armed” but its pixels were explicitly not accepted as weapon draw.
Chrome exited normally, code 0, at 05:39:30.586 UTC. Independent drain at 05:39:46.529 UTC found controller/Chrome and profile processes absent, HTTP/CDP ports closed, profile absent and writer lock clear. No owned browser remained. All 215 remote artifact references, 193 unique payload hashes, were retrieved and independently size/SHA-256 verified at 05:40:35.573 UTC, before stop. Unique logs, captures and actual modules are retained.
The independent observable 25-minute watchdog was armed before allocation.
Fresh box bx_v2pkwf6w is archived, final expiry null; initial expiry was
05:59:10.158 UTC, watchdog deadline 05:59:09.979 UTC.
Normal stop stop_6b72604a559803ec completed 05:40:36.203 UTC,
forced=false, error=null. Both delegates exited normally and the supervisor
records no remaining children. No RAM gate, old403 resume, box deletion or
nested worker occurred. Only this findings note is committed; no emulator code
or guest media changed.