Blood II complete frozen media and measured fourth Enter, 2026-11-18

The immutable preparation omission is fixed and published as 4808701de. One explicitly authorized original reached the reviewed BUILD109 menus and Demo Level 1 card. Fourth Enter reached the host handler and then remained DELIVERY_PENDING until the 500 ms client timeout disconnected transport. No subsequent native USER return was observed for that ticket. Its independent UP was acknowledged. UNKNOWN closed the scope; no later endpoint was used apart from the already armed release. In-level gameplay, enemy consequence and ordinary Quit remain unproved. The historical failure cause and historical termination sender remain UNKNOWN.

Preparation correction and tests

tools/blood2-immutable-preparation.cjs now selects published Git source inside the reusable preparation tool and adds separately pinned media after that selection. It verifies every media row's exact size and SHA-256; there is no arbitrary live-source fallback or launcher-side source filter. The unchanged ignored Game/.wine-assembly-browser.json is an explicit media row, SHA-256 a46bb10441eeb182ff5a28effd8112f507d6a510a25fba48d06870b8fea19519, 10,186 bytes. The same unchanged 98 game files accompany it.

The tracked preparation test is self-contained. An inert registered-app fixture fails with the manifest row missing, fails with binary media missing, and passes with the complete inventory. Main and a real Node Worker both mount/read/hash those inputs through the production immutable loader. Replacing live media and mutating the caller's row array after freezing cannot change their bytes. Altered live media fails the next pinned collection. Existing missing-pngjs fail-fast and authored native input/startup/cancellation tests also pass. Renderer preflight constructs the actual renderer, encodes/decodes a complete 1024×712 PNG in main and Worker, then starts and retires the real GuestThreadHost without loading a PE or running guest slices.

The first focused attempt hit its unchanged 55-second supervisor deadline while repeatedly invoking Git for individual blobs. Batched Git blob reads removed that preparation overhead; the final focused test and locked build passed. The first attempt's containment receipt and failure remain evidence. There was no original in those focused tests and no retained-run dependency in the tracked test.

Actual app preflight and unchanged inventory

Before the original, the read-only preflight resolved APPS.blood2_demo from published source, read the exact pinned manifest, mounted all 98 files in VirtualFS, and checked every mounted hash. It repeated the operation in a real Worker. Neither preflight loaded a PE nor executed game instructions. Both reported identical complete inventories and source/index hashes. Only then did renderer/whole-PNG/native-Worker preflight run through those same descriptors, followed by the original through the unchanged d34 durable launcher. No rows were filtered or mutated between these phases.

The tested graph is published main 4808701dec187469a900bef81d668e10067f2a76, including NFS3's published 70062e2eb CLI window process/lifetime metadata fix. Concurrent foreign worktree edits were recorded as a binary diff and excluded from execution. The retained module is the earlier unchanged unique artifact, SHA-256 7d347a6a631ab77335a14935a533feb896be71aa2508fc5aa19b1e5df8f096ce. It was referenced by existing path, not copied.

The 1,404-row frozen source/media inventory has SHA-256 f804ccd3c1a2d4cf077c39ba8fa56d88ff16b66d0d12afc16a00b54c36fe6c4e; its index SHA-256 is 44545c5af3faf15fd5ec1d18ef293b606cd10d5c1a5d3dd78facea569591779a. Hashes were checked before and after launch and directly through the live original's inherited source/module file descriptors. Node Workers share those process descriptors, and the production loader checks the expected index hash in each Worker. All 530 observed reads across 12 process/thread readers match the frozen inventory. The game-input preflight main and Worker each read all 99 media rows. The original read 13 media rows during its lazy mount/use; all were sealed matches. This is not a claim that the original opened every mounted companion. All 99 disk pins still match after cleanup.

Fourth-Enter measurement and limit

Original PID 2232506 started at 01:57:18 UTC, with the unchanged registered arguments, threads, real ticks, software D3DIM, 1024×712 desktop, 600-second native/watchdog bound and 180-second Quit reserve. Full reviewed frames show Single Player/BUILD109, Caleb, Homicide, then Center for Disease Management / Demo Level 1. Three finite Enter gestures returned native-input-return CONSUMED receipts, independent physical UP ACKs and later native progress. The immediate third-gesture frame still showed difficulty; a later reviewed frame established the level card. Consumption alone did not establish that transition.

Fourth Enter nonce 008ce7f7-7311-4d85-b6fe-8f540f573081 was dispatched at 02:00:40.335 UTC. The bounded native observation retained handler-enter, handler-return and delivery-pending, followed by purge and UNKNOWN on transport disconnect. Its final ticket contains 10 rows, zero overwritten. There is no subsequent renderer poll or native USER return for this ticket. The HTTP trace independently retains request/command-return, transport disconnect, then serialization of UNKNOWN, without response-finish. Thus this measured failure precedes an observed native USER consumption receipt; it is not merely evidence of late serialization after a proven consumption. Handler-side effects may still affect DirectInput/game state; no absence of game effects is inferred.

The independent release returned physical UP ACK 1,471 ms after dispatch, within the unchanged 1,750 ms bound, and its child actually waited exit 0 with both EOFs. The 500 ms confirmation, 3-second operation, 5-second game freshness, menu freshness and current-identity rules were unchanged. No movement, firing, enemy consequence or ordinary Quit was attempted after UNKNOWN. No runtime behavior repair or additional original was attempted.

Cleanup and retained evidence

Worker-owned cancellation at 02:00:55.571 UTC requested SIGTERM. Native stdout records orderly SIGTERM shutdown and held critical-section diagnostics; native exit 0 at 02:00:55.834 is a host-cancellation result, not game Quit. The watchdog actually waited exit 0, and both native/watchdog waits have separate stdout/stderr EOFs and no drain errors. The detached receipt owner retained those waits and native-supervise.cjs containment leaked:false. Its independent disk/deadline watchdog recorded minimum free home space 2,895,572,992 bytes. No RAM gate was used.

Evidence lives in scratch/runs/20261118-blood2-complete-frozen-manifest-gameplay/: exact Git/module/media identities, relevant foreign-WIP diff (not applied), owned code/new-helper patches, actual same-inventory preflights, source-read verification, bounded input/HTTP traces, actual waits, selected named whole PNGs, cleanup and artifact hashes. Redundant black PNGs were removed with hash receipts; no source archive, frame series, media copy or capture bundle was made. The initial offline audit incorrectly expected 99 original reads and exactly four original Worker readers; retained errors distinguish these audit assumptions from the measured 99 preflight reads, 13 original reads and six original Worker readers. The corrected audit passes.

All known runtime/watchdog/release/test PIDs are nonlive and the original listener is absent. The first board watcher ended with actual tool wait 143, sender UNKNOWN; this did not terminate the durable original. Final evidence is published as result.json last. There is no matching candidate-corpus entry, so no unrelated candidate linkage is invented. Box ID null, state NOT_CREATED, expiry null, stop receipt null. No agents, Codex workers, boxes, Wine, Python, deployment or foreign-owner edits were used. Ownership is released at closure.