AVP2 Marine allocation boundary and fresh world observations, 2026-11-07

The historical Out of memory failure is localized to a sparse-heap backing commit refusal, but its cause remains unresolved. The one fresh box reached the original Marine world without reproducing OOM. No WAT/JS allocation disagreement was demonstrated, so this task changes no emulator code and claims no allocation fix. Ordinary movement is visible; aimed mouse attack input was delivered and released, but neither an attack animation nor a hit is established.

Evidence is retained in scratch/runs/20261107-avp2-allocation-cause/ (below, RUN). Earlier evidence and controls are described in the aim/texture note and the async recovery note.

Historical failure: what is known

RUN/retained-allocation-boundary.json records the source path, SHA-256 and numbered excerpts from the retained aim/texture remote-webgl/console.jsonl:

UTC Observation
22:41:32.954 1,048,576 bytes, sparse arena reserved range could not be committed (reason 3)
22:41:32.961 87,048 bytes (0x15408), bump/low reserve/sparse paths refused (reason 1)
22:41:32.964 / .972 Same two failure sizes and reasons recur
22:41:34.472 Original game's Out of memory message box

In src/10-helpers.wat, $heap_sparse_alloc emits reason 3 when $virtual_map_commit(new_top, chunk) fails and empty-arena reclamation either cannot help or its retry still fails; it releases the reservation and returns zero. Here the minimum chunk is 1 MiB. $heap_alloc then emits reason 1 for the 87,048-byte request. This identifies a backing-commit refusal after reservation, not a proven guest-address-space exhaustion, DIB exhaustion, GPU texture heap exhaustion, or host OOM. The adjacent worker-exit records do not recover the allocation caller's stack. The 42,842 CreateSurface entry logs and flat textures alone do not connect a surface request to that refusal.

Missing evidence is the failed commit's allocated-region map, arena/live-object occupancy and caller at that instant, plus the relevant API return chain. The historical run cannot supply these retrospectively. The fresh run supplies those observations only for successful allocations and unsupported surface requests. Consequently this is a concrete blocker to selecting a defensible emulator fix; it is not evidence sufficient to prescribe the next code change.

Fresh bounded observation

The checkout identity recorded during preparation is 05babd650578b76572ca8f39ca975ce6f78e9d17, with RUN/runtime-relevant.patch preserving relevant uncommitted hunks. The runtime reused the predecessor's immutable 619-file pins, whose source identity is 1db931c53342c3bfc6b088fed1cecdf4fd238aa6 plus that run's runtime-relevant.patch; those references and hashes are in source-provenance.json. The runtime pins are in runtime-pins.json; the standard module is 2,182,500 bytes, SHA-256 f9650607181301b5d2a9990420f437bd9f1ca10d40ac73067339f07499d53121. media-refs.json references existing game files by path, size and hash. No source or media archive was made. Original 18-file recovery and the original launcher route were retained; no guest patches, settings changes or clock manipulation were used.

The run-specific observer used existing import trace hooks for HeapAlloc, VirtualAlloc, CreateSurface, SetTexture and heap OOM. It preserved original hook calls and returns, retained bounded descriptor/stack-argument/caller/output records, and read RegionMap's allocated bases rather than assumed offsets. allocation-boundary-summary.json gives compact return records; raw snapshots are under remote-evidence/remote-webgl/allocation-*.json. analyze.py produces allocation-analysis.json by unioning backing intervals and reporting gaps, overlaps, DIB usage and arena live bytes. These are read-only samples, not an atomic stop of every worker; count stability is retained with each sample.

The first captured successful CreateSurface at 23:01:39.904 returned HRESULT 0 and output object 136454192 to caller return address 18825922. The recorded last-error value 183 is not a failure indication for a successful HRESULT API. The first captured SetTexture7 at 23:01:54.072 returned HRESULT 0, last error 0, with device 136454224, stage 0, surface 136457920 and caller return address 18833772. This is the first observed startup binding, not a claim about the first Marine-world texture binding or correct texture contents.

The bounded failed CreateSurface records return 0x80004001 (E_NOTIMPL), output zero. For example, at 23:01:39.917 caller return address 18850365 requested a 16×16, 32-bit surface with descriptor size 124, caps 0x1000, caps2 0x40090, caps3/caps4 zero; last error was 2. The WAT v4/v7 descriptor guard rejects the unsupported caps2 bits against allowed mask 0x2009c before allocating. This is an intentional unsupported descriptor result, not an allocation failure. The late failure guard counted 39,374 such results and captured no HeapAlloc or VirtualAlloc failures. No WAT/JS semantic defect was established from this path.

At final observation (23:12:55.990), the main worker had equal entry/return counts: HeapAlloc 252,598, VirtualAlloc 496, CreateSurface 259,902, SetTexture7 380,231, with no observed heap OOM. Counts establish observation coverage only.

Resource Final observed occupancy
Wasm linear memory 536,870,912 bytes, fixed
Virtual backing pool [0x08000000, 0x1a000000), 301,989,888 bytes
Allocated backing 207 entries, union 157,487,104 bytes, no overlaps
Free backing 144,502,784 bytes, one contiguous largest gap
DIB backing 16,965,632 / 66,060,288 bytes
Surface objects / texture-capable objects 1,316 / 1,127
Surface vidmem accounting ledger 5,642,016 bytes
Heap arena live-byte accounting 103,313,752 bytes across 166 arenas

The surface ledger is not GPU heap occupancy. GPU texture heap occupancy and host allocation refusal were not measured or established. Fresh healthy occupancy cannot be substituted for historical failure occupancy.

Reviewed world and native input

Actual rendering was WebGL 2.0 through ANGLE on Intel UHD Graphics 620, Mesa 23.2.1 (forward-world.json); retained renderer counters show zero fallbacks. The original Marine selection and Unwelcome Guests/Normal launch were reviewed, followed by loading, the Marine introduction, then the first-person world. intro-progress.png shows corridor, teammate and 100/050 HUD. The earlier menu-entry.png is a logo and loading-bar captures are not world evidence.

The passed immutable input/collector controls and corrected same-Mouse aim helper were reused. D for 300 ms produces visibly changed corridor geometry in after-strafe.png; A for 600 ms gives corridor-return.png; W for 600 ms gives forward-world.png with teammate and illuminated floor. Each has trusted screen DOWN/UP events and final physical/async/high-bit state false in its input ACK.

Two MouseLeft pulses at client (430,450), 1,200 ms and 300 ms, have trusted screen DOWN/UP events and cleared final physical/async/high-bit state. The read-only capture helper took knife-edge.png shortly after the second DOWN; its state saw the high bit true. Its separate physical probe read the wrong mouse slot and was false; that field is not evidence of a missing press. knife-during.png, knife-edge.png and after-attack.png show the knife HUD and a dark wall after aiming changed the view. They do not establish a swing, target reaction, damage or a hit. No attack success is claimed. All commands completed with known acknowledgements, no UNKNOWN or replay; every press received an independent UP. The final input terminal has held [], gate failure null.

Diagnostic limits, validation and shutdown

The first startup navigation trapped before any native input because the new observer passed a SharedArrayBuffer-backed view to Chromium's TextDecoder. Its logs are preserved under remote-evidence/startup-observer-failure/. The original driver/browser were closed, with ports/profile/process cleanup verified. The observer was corrected to decode copied bytes and fail open; the original was restarted on the same box. Thus there were two startup navigations, one completed world run, and no second box or unknown-input replay. This diagnostic repair is not an emulator allocation fix. All ten worker observers were removed before normal application shutdown (allocation-final.json).

Validation retained in RUN:

Only box bx_rnbsqq9m was created, UUID-idempotently at 22:56:18.395 UTC. Its initial expiry was 23:31:18.412 UTC. Before launch, an independent observable detached watchdog was armed for 23:21:18.562 UTC, alongside the normal-stop delegate. Launch free-space floors exceeded 5 GB locally and remotely. retrieval-verified.json confirms 208 artifacts / 141,153,874 bytes retrieved and SHA-256/size verified before requesting stop. Cleanup verified both ports closed, no live owned processes, browser profile absent and held inputs empty. Normal delegated shutdown completed at 23:13:43.516 UTC, receipt stop_c52b02c8efe15cb9, forced false, no error. Final box state is ARCHIVED, archiveAfter: null. Both lifecycle children exited 0 and supervisor remaining children is {}. No resume, deletion or public deployment occurred.

Next work requires a separately authorized reproduction of the historical failure with the failing sparse commit's caller and live backing/arena snapshot. Until that exists, there is no justified allocation patch or regression to land.