Baldur’s Gate II native cage-key attempt, 2026-11-08

The native key pickup and Jaheira release remain unverified. This fresh attempt stopped at character naming because the worker sent lowercase a through an uppercase-only typing helper. This is a concrete harness blocker, not a game-pathfinding or emulator failure. No opening-scene, pause, weapons-room, inventory, key or cage-release evidence was obtained in this run. Historical movement, Jaheira dialogue and Minsc recruitment are not counted as new progress.

Evidence is retained in scratch/runs/20261108-bg2-native-cage-key/ (ignored). The large-file gameplay and cage-rescue handoffs, CLAUDE.md and shipped README.TXT control guide were read. The guide documents I for inventory, G for return to game, portrait drag for party order, and chests/bookshelves/boxes as containers. All 645 unchanged originals matched exact size/SHA-256, totaling 850,839,386 bytes. BGMain.exe remains 944db091d82c6144db8c9a21c085e4504bdc549fca8003c31e309646567bb98b. The registered manifest lists the other 644 files.

The released 1 MiB chunk staging and qualified native controls were reused; no upload/preflight project was repeated. Complete remote verification and atomic publication passed for 1,284 entries, including the 1,283 runtime/input pins and their manifest. Git source identity is 7d8dd962d2043e0bffc542cc7fbaae93904a1dbc, with relevant-source.patch, new-files.patch, dependency references/patch and individual media references retained. The sole non-run runtime difference from the preceding attempt was the published app-file-size metadata correction. No source/media archive was created.

One UUID-idempotent box, bx_zanu7767, was created at 00:54:11 UTC after the detached watchdog was observed armed and /home exceeded 5 GB free. Its original expiry was 01:34:11.564 UTC. The immutable browser deadline was 01:16:02.987 UTC; the independent watchdog deadline was three minutes later. The sole original launcher started at 00:54:56.199 UTC. Actual runtime was Chrome 151.0.7922.108, V8 15.1.206.16, six guest Workers, process 1000/endpoint 1/slot 0, and Win98Renderer/Canvas2D without a render-worker endpoint. The recovered live module is f9650607181301b5d2a9990420f437bd9f1ca10d40ac73067339f07499d53121. Toolbar WebGL selection does not qualify GL coverage.

Fresh captures were personally reviewed before dependent input. A short New Game click showed no transition; a separately reviewed longer native click entered ordinary creation. Creation selected male Human Fighter, Lawful Good, unchanged first roll 13/15/13/12/13/12, two proficiency points each in Axe, Long Sword and Two Handed Sword, and default portrait/colors/sound. The last reviewed screen, review/name.png, shows the empty name dialog and caret.

Published command 32, name-type, requested ordinary typing of lowercase a. keyboard-bridge.cjs requires [A-Z]{1,12} and throws bounded uppercase name before its first CDP event. However, that validation occurred inside the driver's admitted dispatch, after command publication. The driver recorded first-failure.json and first-browser-failure.json; the client recorded sticky transport-closed.json. There are 32 exact successful ACKs and one unacknowledged command, with 30 independently released mouse actions. No input was replayed, no transport marker was cleared, and no further game action was attempted. The zero-send helper rejection explains the cause but does not reopen the UNKNOWN stream.

After shutdown, the owned local request.py gained the same uppercase-name check before review/request records or remote publication. type-admission-regression.json proves the exact lowercase request now fails locally without a request record or remote call. keyboard-contract-test.json exercises the actual helper: lowercase makes zero CDP calls, while uppercase A produces paired down/up events. The launched helper bytes remain recoverable in new-files.patch; the post-run repair is separately preserved in owned-inputs-after-repair.patch. This run changed no tracked emulator/tool code and required no code build. No generic emulator defect was demonstrated.

Chrome closed normally at 01:00:11.798 UTC, exit 0/no signal. All 394 terminal artifacts (22,817,294 bytes), including unique captures/logs and the live module, were individually retrieved and SHA-verified before normal shutdown was requested at 01:00:42.269 UTC. Stop stop_67041c0b1ca63e6f completed without force at 01:00:42.745 UTC. Final box state is archived, expiry cleared. All terminal runtime pins matched; runtime PIDs were absent, both ports closed and the profile removed. The watchdog observed normal stop and exited; the local board watcher was terminated. No extra box, resume/delete, guest/save/configuration/clock edit, private handler, fabricated restore, archive or public deployment occurred.

The worker launcher ended with signal 143 after shutdown and local evidence checks, before publication or outer worker result.json/final.txt. Its termination cause is unknown. The coordinator confirmed the matching processes were absent, reclaimed this unfinished note and verified the retained artifact hashes and shutdown records. The run-level result remains a harness error; this recovery does not turn it into a clean worker pass. No original input was repeated or old box resumed.

Next: a separately assigned fresh original run should reuse the local pre-publication type guard, use uppercase naming or the qualified ordinary key route, complete the opening scene, verify actual pause and portrait selection, follow Imoen’s west/slightly-north hint, inspect native container/loot UI, verify the key in inventory, then release Jaheira. Do not resume this archived box. Combat, FPS, audio, GL and full playability remain separate.