Blood II durable causal reproduction: preparation refusal, 2026-11-18
The fourth-Enter failure remains UNKNOWN. The one newly authorized original process stopped before menus because this worker's run-local preparation filter omitted the ignored game manifest. No Enter, movement, firing, enemy consequence or ordinary Quit was submitted. No original retry or endpoint request followed the failure. This is a preparation mistake in this run, not evidence of the historical fourth-Enter cause or termination sender.
Measured failure
The original used published main 47095dcb6c84178f63dde8bb61e9fd142aa1f4b9
and the retained module SHA-256
7d347a6a631ab77335a14935a533feb896be71aa2508fc5aa19b1e5df8f096ce.
Published Git blobs replaced concurrent tracked worktree modifications in
the immutable inventory. The same filter retained package-lock dependency
files, but accidentally dropped the explicitly collected, ignored
Game/.wine-assembly-browser.json. All 98 media files plus that manifest
had passed their unchanged pins. Checking media on disk did not check its
presence in the sealed inventory. Renderer/main-and-Worker whole-PNG
preflight passed and likewise did not exercise the registered asset manifest.
PID 2202628 ran at 01:41:32 UTC. The native log records PE entry 0x00465870
and the unchanged registered arguments, followed by the exact
unsealed runtime input refusal in getAssetFiles during asset mounting.
No menu or original screenshot exists. The retained native wait is exit 1
at 01:41:33.339 UTC; the watchdog actually waited with exit 0 at
01:41:33.369 UTC. Both have separate stdout/stderr EOFs and no drain errors.
The detached owner retained both waits and native stdout/stderr.
Containment reports leaked:false, no termination request and exit 0 for
the receipt-owning wrapper. That wrapper exit is not the native exit:
the original's authoritative native exit remains 1.
An offline check reconstructs those exact immutable source rows and proves the missing-manifest read fails through the production loader. Adding only the explicitly pinned manifest row makes that read pass. This identifies the preparation correction without executing another game or contacting an endpoint. It is not a native old-fail/new-pass fourth-Enter behavior repair, and does not earn another affected-original run. The executed launch helper is retained unchanged so its omission remains directly reviewable.
Reused hooks and focused controls
No runtime, controller, launcher or observation code changed. Existing bounded public-input and HTTP phase hooks were selected; there was no input to observe in the failed original. Historical missing logs remain missing. The original's lack of an input trace is not evidence about the historical failure stage. No deadline, timestamp, identity, freshness or release rule was weakened.
Fresh authored loading controls survive submitter TERM and SIGKILL. Native
Enter confirmation took 27 and 20 ms; the real 900 ms loading loop and 4096
GDI calls completed afterward. Each retains native/watchdog actual waits,
both EOFs, independent UP and clean containment. The TERM loaded whole PNG
shows the completion caption and color patch. The SIGKILL loaded whole PNG
is a blank teal desktop despite the earlier title snapshot; it is retained
as a visual limitation, not claimed as rendered completion. The published 2ffa
renderer/preparation and d34 lifecycle proofs and locked build remain
applicable to unchanged production code.
The native observer-on/off pair both reports UP after DOWN, actual Worker
activity, native DOWN and no late DOWN after UP. Seven paired unit guard
cases preserve input outcomes; bounded-row and observer-fault controls pass.
The Blood II forward/turn/Ctrl, lost-ACK/independent-UP and terminal-no-endpoint
controls pass. The real HTTP phase test passes after providing its required
output directory; the first invocation's configuration error is retained.
An initial offline-check syntax error is also retained separately from the
passing corrected check. None launched another original.
Evidence and closure
Evidence is in
scratch/runs/20261118-blood2-durable-level-entry-causal-gameplay/.
It retains executed small helpers, exact Git/source/module identities,
relevant worktree binary diff (excluded from original execution), new-helper
patches, source-read verification, all actual waits, unique logs, offline
manifest proof, selected named whole PNGs and artifact hashes. Git remains
the source archive; the unique existing module and media were referenced,
not copied. No matching candidate-corpus entry exists, so this run is not
attached to another game.
All known owned runtime, watchdog and release PIDs are checked offline after closure, and the original loopback listener is absent. The independent original watchdog recorded minimum free space 3,171,655,680 bytes, above the 1.5 GB floor. There was no disk or RAM gate change, foreign cleanup, Wine, Python, box, worker, deployment or public publication of a site. Box ID null, state NOT_CREATED, expiry null, stop receipt null. Documentation is committed and strict-published; result.json is finalized last after cleanup and checks. Ownership is released and the worker stops with the causal/gameplay goal unachieved and this concrete preparation blocker recorded.