AVP2 Marine aim and texture boundary — 2026-11-07

One bounded successor to async recovery. Run: scratch/runs/20261107-avp2-aim-textures/ (ignored evidence). Result: the late-controller Mouse-cache discrepancy is reproduced and locally fixed. Original Marine first-person world appears, then an Out of memory dialog prevents the fresh movement/attack test. No attack consequence, hit, damage, or original-game mouse acceptance is established. No emulator change.

Controller correction and focused verification

aim.cjs, used by the actual late-input.cjs and staged driver, moves using the same page.mouse object that sends DOWN and independent UP. It checks live document/worker/render-owner identity, elementFromPoint, canvas bounds, foreground owner/window bounds, and records canvas/window coordinates. It requires trusted DOWN/UP at the exact aimed client point and target screen. MouseLeft release also checks VK1 physical and asynchronous high bits. Independent UP, held-key cleanup and sticky UNKNOWN Gate remain unchanged. These are run-specific helpers, preserved with new-file binary patches; no tracked runtime source was changed.

local.cjs uses production WineAssembly, guest Worker, Win98Renderer and browserInput with the local boids render endpoint. local-results.json passes:

This local check proves controller routing, not original AVP2 gameplay. Earlier harness readiness failures and intermediate runs remain alongside the final pass. Local Chrome/server/profile cleanup is recorded in local-cleanup.json.

Unchanged physical-key fix 74d173af4, collector 8453f67b4, immutable controls and their passed checks were reused (reused-controls.json). All 619 staged runtime pins match (changed-entry.json); actual Node dependency closure and all staged hash/size checks pass. node tools/verify-avp2-recovery.js passes original18. NODE_OPTIONS=--v8-pool-size=1 flock /tmp/wine-build.lock bash tools/build.sh passes (build.exit=0). Standard module is f9650607181301b5d2a9990420f437bd9f1ca10d40ac73067339f07499d53121 (2,182,500 bytes), also recovered from instantiated original workers.

Single original attempt

Exactly one fresh UUID / box bx_hs48msbv, created 22:27:27Z, initial expiry 23:02:27Z. Detached observable normal-stop and 25-minute watchdog helpers were armed before allocation; immutable watchdog deadline 22:52:27Z. Local and remote launch floors exceeded 5 GB. Original launch 22:31:26Z. Staging retained Git SHA 1db931c53342c3bfc6b088fed1cecdf4fd238aa6 and relevant patches, not a source archive. Media paths/sizes/hashes are in media-refs.json.

Native WebGL2 Intel UHD620 ANGLE/Mesa composition, actual guest CPU and render owners, response-body collector, and original module guards pass with no collector first fault. Late entry seals process1001, hwnd131073, CPU endpoint2 and render endpoint1. Ordinary Enter pulses select Single Player, Marine, Unwelcome Guests / Normal and launch. The loading screens named marine-world, marine-interior and intro-current are not world evidence. world-progress.png shows the conversation interior with four Marines. Escape opens the pause menu; Enter resumes. resume-world.png at 22:39:48.440Z shows first-person corridor, knife lower right, teammate, HUD health100/armor050 and radio. This is native scene progression, not a hit.

Before any fresh movement or attack, corridor-fresh.png at 22:42:40.696Z shows AVP2's Out of memory dialog. No further gameplay input was sent. No MouseLeft pulse was sent in this original attempt. No settings, clocks, private handlers or bindings were invoked. No replay or second box. An observational command11 used type instead of action; it returned a known unsupported-command error without input. Corrected observational command12 produced the OOM capture. Quit helper7 reports missing .done after its intentional quit branch; late-input-terminal.json establishes normal closed state instead. Neither event is an uncertain native input outcome.

Passive texture provenance and limits

textures.cjs takes eight read-only shared-memory state snapshots and a short built-in API trace window, then uses the existing bounded producer diagnostic. It reads region-map bases, direct guest state and own VFS property descriptors; it never invokes lazy .data getters. Trace names are removed after the bounded window; initial creation tracing has a 180-second removal timer. Forwarding remains enabled, so this is not a performance benchmark.

At 22:36:31–33Z during loading, all eight snapshots show 520 live texture-capable surfaces (617 total type2 surfaces), device slot10, stage-zero binding0 and owned texture reference0. Stage-zero defaults are [0,4,2,0,4,2,0,0]. Existing producer receipt endpoint2 accepts30 draws and submits16 times; endpoint1 is idle. This is distinct from the predecessor's retained565 accepted calls/62 submissions and sampled diffuse/no-texture triangles.

texture-api-window.json retains 42,842 initial CreateSurface entry logs, one EnumTextureFormats7 entry, and 468 window entries: 24 CreateSurface, 168 SetTextureStageState7, 276 SetRenderState7, zero SetTexture7/3. Texture creation reaches guest API/WAT object state. In this bounded loading window, the first observed absence is the texture-binding entry and its WAT stage-zero state, upstream of textured producer submission. This does not prove SetTexture never occurs in the Marine world. The later world receipt has 311 cumulative texture uploads; uploads alone do not establish texture binding on the sampled diffuse triangles.

The original d3d.ren call return 0x011f9857 maps to PE 0x10040857: CreateSurface call at 0x10040854, followed by HRESULT test and success branch. Its failure path contains “Unable to create (%dx%d) texture surface.” Disassembly and exact media reference are retained. Entry-only traces have no HRESULT/branch receipt. Prior descriptor stack addresses were sampled after reuse and read zero; they are explicitly not return evidence. Therefore neither failed texture creation nor a faulty WAT SetTexture operation is established. The inspected Device7 wrapper/vtable reaches $d3dim_set_texture, which stores stage0 binding/reference; no speculative renderer patch was made.

Next work needs bounded creation entry/return and first binding provenance while the scene is fresh, plus a fresh reviewed movement/aimed attack test. The OOM allocation failure must first be attributed from passive evidence; do not infer its cause from the trace count or flat textures.

Recovery and shutdown

Unique remote logs, captures, instantiated module and receipts are recovered with exact size/SHA-256 checks before the explicitly delegated normal box stop. remote-artifact-inventory.json, retrieval-verified.json, remote-cleanup-audit.json, lifecycle terminal records and normal-stop-confirmed.json are the terminal evidence. They establish exact owned runtime process, port, profile and held-input cleanup and box state. No public deployment. No archive of source, scratch, docs or repeated media.

Terminal outcome: 118 remote artifacts / 20,779,346 bytes verified. HTTP46473 and CDP43075 closed, owned processes empty, profile absent, held input empty. Box bx_hs48msbv confirmed archived at 22:44:35Z by normal non-forced stop, before the original 23:02:27Z expiry. Both lifecycle children exit0 and supervisor remaining set is empty. Local messageboard watcher is retired. Lifecycle ownership returns to the coordinator; no live assigned box remains.