AVP2 current browser attempt, 7 November 2026

The original fixture and registered browser manifest are restored. Current browser gameplay remains unverified: the single admitted original route stopped during required-file loading on a recorded Chromium cancellation. No gameplay inputs, scene captures, guest configuration changes, or emulator repair were made. The accepted October gameplay and fixed HUD/DXT diagnoses are unchanged.

Evidence: scratch/runs/20261107-avp2-current-browser/ (ignored).

Fixture recovery and registration

All 18 cabinet files were recovered from the retained scratch/runs/20261018-avp2-local-launch/handoff/. Case-insensitive original names were mapped to the spelling in test/candidate-corpus/avp2-demo-files.json; gzip streams were decompressed, and existing uncompressed files were hardlinked. Every recovered size and SHA-256 matches that inventory. tools/verify-avp2-recovery.js passed locally and remotely. The fixture is now at test/binaries/candidates/avp2-demo/game/.

The existing writeBrowserManifest function in tools/fetch-candidate-corpus.js wrote the ignored candidate-root manifest using the corpus's actual AVP2 browser definition. Its 17 companions exclude the original launcher executable and contain no command-file override. test/test-candidate-browser-manifest.js passed. The retained 136-byte handoff/avp2cmds.txt was neither edited nor copied into the fixture; its before/after SHA-256 is 48d472700862291e96bdb7e1ce2311b91778b38d5cfebf6f8de6fd3ab411b76a. Original autoexec.cfg and Profiles/Default.prf passed the cabinet hashes; no settings were synthesized. No new generated child command was recovered.

Recovery is reproducible without an archive: for each inventory path, read its lowercase retained handoff path, or its .gz stream if the plain file is absent; verify the inventory size/hash before writing the original-case path under game/; run the verifier, then the existing manifest writer. The run's fixture-audit.json records every exact source path, compression state and verified original hash. registered-manifest.json preserves the small generated manifest; it does not duplicate game media.

Tested source and bounded environment

Source commit: 57139b29b3c5be175e012c50a4a7c1964ab7347c. source-diff.patch records relevant uncommitted runtime hunks, including the already-present untracked lib/wininit-boot.js. No source tree was archived. Selected runtime, media, fonts, native DLLs and module inputs were pinned and verified at their actual remote paths: 606 files, zero terminal mismatches. The ignored binaries link resolves to test/binaries locally and remotely. Chunk transfer used temporary remote pieces and verified complete assembled files before admission. Game media remains referenced by fixture path and hash.

The served standard module was e41511fe861363fbe1639d80f20cbcdd7240f46937c93a801eaa256ca094c467 (2,182,493 bytes); its exact bytes are retained as tested-wine-assembly.wasm. The original-route console reports guest Worker slot0 instantiated with 2288 exports and 294 brokered imports, but the actual instantiated module bytecode receipt was not persisted before the collector failed. Consequently the served module identity is verified; independent instantiated-module identity is unknown.

One UUID-idempotent fresh box: bx_z4d6hnmp, AMD Ryzen 7 9700X, four reported cores, Chrome151.0.7922.108, isolated threaded page, 1024x768 viewport, requested WebGL. No rendering endpoint was created, so actual renderer/backend/GPU-provider identity is unavailable and no GPU rendering claim is made. The 35-minute box deadline was 08:45:12.045Z, with input admission closing six minutes earlier and runtime cleanup four minutes earlier.

Two setup mistakes were retained explicitly. Supplementary files initially uploaded beneath a duplicated absolute prefix; assembly failed at its first read with no assembly/runtime effects. Correct intended-path uploads then passed all preflight hashes. Next, the dropdown-based harness closed before guest admission because AVP2 has no dropdown option. index.html explicitly supports registry IDs outside the dropdown through ?app=ID; the first and only original route used that documented direct link in the same box. The earlier selector attempt had no launch-input receipt, no guest host, and normal browser exit0; its artifacts are retained under prelaunch-selector. Neither mistake was treated as a guest failure or used to justify an original runtime retry.

Exact new blocker

At 08:12:42.123Z, static-server request163 completed HTTP200 for test/binaries/candidates/avp2-demo/game/AVP2.exe: 57,344 bytes, SHA-256 36a772c2340ea7e415d8213d92b7a18d9541ef07c53b028d8a8ec475caf79b7d. At .124Z the owning page's response-stage collector captured and verified the same body, then recorded continued. Chromium subsequently reported Network.loadingFailed, canceled:true, net::ERR_ABORTED for the same network ID 7496.264. This became the first sticky collector fault.

The boot capture calls bodies.drain() before writing a screenshot. Its drain threw at tools/browser-response-body-collector.cjs:341, so there is no boot PNG or reviewed screenshot. At 08:13:01.889Z the terminal state was running:false, guest Worker present, main:null, no children, no render endpoints, and status Loading data files 9/17. The runtime closed at 08:13:02.634Z with browser exit0, no signal and empty held-input set. A later read-only screenshot attempt found the already-closed CDP port; it did not relaunch or send input. Detached driver OS wait status is unknown; the driver error and terminal output are retained separately.

The actionable next step is a generated-only reproduction of this exact captured/verified/continued required-HTTP200 cancellation, with an independent consumer completion witness, before another original session. Retained server and collector logs distinguish complete correct server bytes from the client abort. They do not establish why Chromium canceled the request, that AVP2's launcher executed, or an emulator instruction/API defect. Do not bypass the fault, replay the guest, fabricate a callback, or revive the fixed HUD/DXT work on this evidence. There is no semantics-based source repair to commit here.

Tests, evidence and cleanup

Original18 verification passed twice; manifest regression passed; the run-local bounded-input controls passed delayed keydown/UNKNOWN finally-release and serialized-success checks. No emulator source changed, so no new locked-build or emulator-regression claim is made. Ordinary Marine menus, static geometry, causal movement and fire were not reached. No FPS, audio, device, parity or full-corpus claim follows.

Thirty-five unique remote artifacts (1,778,272 bytes), plus the terminal audit, were retrieved and individually size/hash checked before stopping. The audit verifies 28 tracked PID/start identities absent, four ports refused, profile and singleton locks absent, zero input events and zero queued input. Both prelaunch and original harness terminal records refer to the same reused profile pathname, which is absent. Physical/async key maps are null because no guest renderer input owner was established; this is not a fabricated key release receipt.

Delegated nonforced stop completed at 08:14:52.934Z; box state archived, expiry null, no resume/delete. Lifecycle was under five minutes. Local free disk remained above 2GB (2.85GB at final check). The restored ignored fixture and retained historical gzip sources remain intact. Result metadata is written last, after note publication and artifact verification.