Blood II level-card loading and native music NULL call, 2026-11-18
One authorized original reached the personally reviewed BUILD109 menus and Demo Level 1 card with three finite Enter gestures. All three received native USER consumption receipts and independent physical UP acknowledgements. No fourth Enter was sent. Passive observation established continued filesystem and DirectX activity behind the card, followed by a native NULL call before an actual world frame. Movement, firing, enemy consequence and ordinary Quit remain unproved. The bounded task ends at this concrete blocker with cleanup verified; there was no runtime repair or second original.
Same frozen graph and preparation
The source/media inventory is exactly the preceding 4808701de inventory:
1,404 rows, the unchanged manifest plus 98 mounted game files, source SHA-256
f804ccd3c1a2d4cf077c39ba8fa56d88ff16b66d0d12afc16a00b54c36fe6c4e,
index SHA-256
44545c5af3faf15fd5ec1d18ef293b606cd10d5c1a5d3dd78facea569591779a.
Published 4808701dec187469a900bef81d668e10067f2a76 includes the 70062e2eb
native process/window metadata correction. The retained module remains
SHA-256 7d347a6a631ab77335a14935a533feb896be71aa2508fc5aa19b1e5df8f096ce,
referenced at its existing recovery path without copying it.
The unchanged production preparation helper collects that published revision
and the same 99 media pins. No new source filter was written. Main/real Worker
mount/hash preflights and actual renderer/whole-PNG/native-Worker preflight
passed, then repeated with the passive observer enabled through the same
descriptors subsequently inherited by the original. Native source/module FD
hashes also match directly through /proc. All 547 observed reads across 12
process/thread readers match the frozen inventory; all 99 disk pins still
match after cleanup. Concurrent foreign worktree changes were recorded as
unapplied binary patches and excluded from execution.
The route is existing native Node CLI, normal guest Workers and software D3DIM, with the preceding original's registered BUILD109 arguments, real ticks, screen, runtime limit and Quit reserve unchanged. The d34 durable receipt owner and existing native-supervise process-group/pdeathsig route were reused. No cooperative/software async-owner graph was substituted.
Passive observation qualification
The run-owned observer.cjs is an explicitly identified preload before the
unchanged sealed-source loader. It wraps selected existing native imports,
returns each original result unchanged, and samples existing read-only native
getters. It records counters and the latest import arguments/results. It does
not inject input, execute guest slices, switch API logging, advance the guest
clock, modify guest memory, request snapshots, or replace runtime sources.
Limits are 250 ms sampling, at most 2,600 rows per Node thread, and a finite
600-second deadline. Separate timers and synchronous import-return samples
allow already-emitted progress to guide capture availability.
An authored PE autonomously waits on the ordinary native clock, paints 4,096 pixels through guest GDI and exits without any input requests. Observed and unobserved runs used the same frozen graph and produced identical complete pixels; actual native/watchdog waits, EOFs and clean containment passed. A small real-Wasm test additionally verifies native instance identity, exact import return and exception identity, one original call per invocation, unchanged memory, no injected input, and time/row bounds. These tests establish the measured outcomes, not identical performance or scheduling. Initial authored branch/audit corrections and their actual waits are retained; empty non-Wasm Worker witness files caused the initial offline JSON audit errors. The corrected audit passed without another original.
No tracked production/test JavaScript changed. The preceding published
4808701de focused tests and locked build remain the runtime qualification;
the new helpers and their tests are run-local evidence only.
Measured original phase and fault
Original PID 2258231 launched at 02:17:46 UTC. Whole images were personally reviewed for Single Player, Caleb, Homicide and the level card. Each ordinary Enter used a fresh reviewed menu and a finite 90 ms intent. Native consumption, 500 ms confirmation, independent UP within 1,750 ms, 3-second operation, freshness, identity and Quit-reserve limits were unchanged. The final immediate action frame still showed difficulty; the later whole card frame established the actual transition. The original config binds Forward to Up and Fire to Ctrl; no gameplay key was sent before a world frame.
The level card capture completed at 02:19:42 UTC and showed Center for Disease Management / Demo Level 1 with physical inputs UP. Already-emitted main-timer samples established a serviceable event loop before this capture. The card has no visible additional-confirmation prompt. Four bounded 20-second passive windows made zero endpoint or input requests. Loading-associated native execution continued after the third Enter: EIP/API counts changed, filesystem reads and seeks continued, and DirectX method activity later increased.
From the sample adjoining the card capture to the first main EIP-zero sample,
the counters increased by 7,645 fs_read_file_result, 3,713 fs_seek_result
and 24,660 dx_trace calls. check_input, input_consumed,
gdi_surface_upload and gpu_stream_kick did not increase. Observed read
results of zero mean filesystem success, not bytes read; byte/offset progress
was not measured. DirectX calls alone do not prove a rendered world. This run
establishes active loading-associated work with USER polling paused during
the observed card interval. It does not establish an eventual timed
transition, a later confirmation requirement, or the historical fourth-Enter
failure cause.
At batch 112287, native stdout reports guest called through NULL, previous
EIP 0x00cd1c61, and return address 0x00cd1c81 at ESP. The first subsequent
passive main-zero sample is 02:22:12.581 UTC. Main EIP and API count then stay
zero/fixed while other guest threads still run. No new surface upload, USER
poll, world frame or later endpoint request was observed.
The fault site maps to IMUSIC25.DLL: its pinned PE entry RVA 0xff49 and the
observed DllMain entry 0x00cdff49 imply relocated base 0x00cd0000. Static
disassembly at RVA 0x1c7f shows call *(%ebx) with return RVA 0x1c81;
the native dump shows the relevant object starts with zero. This identifies
the music-interface indirect-call failure site. The bad pointer/vtable's
initialization provenance and source-level root cause remain UNKNOWN.
Earlier native stdout also records music Worker T2/T3 unexpected EIP-zero
termination before menus; no causal connection is asserted.
The worker closed the scope after this concrete fault. No fourth Enter, movement/fire, postfault snapshot, replay or additional original was used. The last endpoint request was the card capture at 02:19:42.504 UTC, over two minutes before the native fault. The earlier run's fourth-Enter timeout was not reproduced and is not retroactively explained by this fault.
Cleanup and recovery
Worker-owned cancellation requested orderly SIGTERM after the fault. Native
exit 0 at 02:22:41.512 UTC is host cancellation, not game Quit. The native and
watchdog have actual exit-0 waits, separate stdout/stderr EOFs, no drain
errors, and containment leaked:false. All 26 known owned process IDs are
nonlive and port 43945 has no listener. The own board watcher has actual tool
wait 143 following the worker's intentional SIGTERM. The independent disk
watchdog stayed above the 1.5 GB floor; exact minimum and final space are in
cleanup-verification.json. No RAM gate was used.
Evidence is in
scratch/runs/20261118-blood2-level-card-loading-world-gameplay/: native/API
counter observations, original controls/receipts, selected named whole PNGs,
fault disassembly, same-inventory preflights, direct FD equality, media and
source readbacks, run-helper patches, relevant unapplied runtime patches,
actual waits and cleanup. The Git revision plus those patches and the prior
unique module path recover the tested graph. No source archive, frame series,
capture bundle or media copy was made. result.json is written last.
Box ID null, state NOT_CREATED, expiry null, stop receipt null. No agents, workers, boxes, Python, Wine, public deployment or foreign-owner edits were used. Ownership is released at closure.