Blood II selective native lifetime watch, 2026-11-18

Evidence: scratch/runs/20261118-blood2-selective-native-lifetime-watch/.

The preceding 1f6f6c7e1 observation closed UNKNOWN before menus after 322 seconds and 6,071 repetitive nonzero Releases. This preparation replaces that JS block observer with a default-off native filter and bounded consumer. Preparation was tested and published as 12ecbef77 before the one newly authorized original. That original closed UNKNOWN at a measured missing module-discovery boundary. No music cause or game progress was established.

The native filter captures ECX at the actual IMRT constructor, publishes that pointer in diagnostic shared memory, and retains the first live state, primary word deviation and refcount zero. It filters the zero-Release branch, destructor and free argument against that actual object. Late CPU instances arm from the module configuration already published by the first instance, before their first guest instruction. There is no JSON/DAT pointer polling, JS callback on unrelated blocks, or retained nonzero reference-count history. The final consumer discovers modules after the real load_dll and init_thread exports return. A proxy exposes these two wrappers while the native exports and every execution call remain unchanged. Original return values and exceptions propagate; this adds no guest callback or input.

All previous region bases remain unchanged. The appended diagnostic region is 128 bytes. The native event cap is 64 per shared scope; the consumer permits 128 rows and 32 instances per Node thread, with a finite deadline. Reuse, post-execution reset, unmapped observation, cap or consumer error closes the observation UNKNOWN. Existing import exceptions propagate. Disposal disarms at a JS boundary; a scope cannot be rearmed. The actual native observing thread, PC, prior block, registers, stack, object and DLL rows accompany each witness. These identify an observation interval, not an exact writer or a concurrent writing thread. A page notification can precede the store; the following boundary can therefore be the first witness.

The existing page/store notifications and block boundaries are used, with post-copy/fill checks for fast REP and native bulk helpers. A native control caught a missing post-REP boundary during preparation. Authentic Worker controls then caught late-instance arming after zero Release; shared native configuration fixes that gap. Render controls exposed initial-load versus post-execution-reset handling. Failed controls and their actual wait receipts remain in the run directory. None was an original game scope or a game defect.

Focused native instructions cover unrelated same-page/other-page writes, suppressed nonzero references, transient primary/refcount zero and restoration within one block, byte/word/dword, MMX, SSE, REP MOVS/STOS, shared CPU instances, first-event claims, reset, cap, disposal, errors and late-instance arming. On/off registers, lazy flags and guest memory agree for these controls. The existing page-watch, sparse string and REP fault-restart regressions are also run. This does not assert coverage of arbitrary unnotified host/SAB writes.

Authentic IMRT25/IMUSIC25 constructor, QI and ownership-transfer checks pass with the old module, new watch off and new watch on. Real native Workers perform retained AddRef/Release on exported thread 2 and final Release/free on thread 3; outcomes agree. The on control witnesses zero, zero-Release branch, destructor, free and primary invalidation. No fake reference, restored vtable, music disable, guest behavior repair, forced polling or suppressed crash is introduced. A normal guest Worker/software-render fixture has identical whole pixel hashes in all three arms. Its complete image was reviewed: teal desktop and gray authored window, not a game menu. Initial authored render elapsed totals were 2,838/2,640/2,158 ms; guest batches were 985/619/1,275. These noisy single samples establish no performance improvement, timing equivalence or scheduling claim. The watch disables chaining/uops and can perturb execution.

The frozen base remains Git 4808701dec187469a900bef81d668e10067f2a76, 1,404 inventory rows and 99 unchanged media pins. The existing 7d347a6a... module is reproduced exactly from that Git source plus its three recorded, unchanged Win16 WIP blobs; its recovery patch is referenced, not duplicated. Applying only this task's five native file changes yields module 21cbbb4ae9a08860007ea3e74799c8a42fc0c377d5f6c182f09b01dff26bbe80. Only the generated region-map row changes in the frozen JS graph. Main and Worker registry/media/whole-render preflights pass through the historical published sealed loader recovered from Git in memory. FoxBear's shared preparation, source-view, launcher and supervisor files are not edited.

The original used the registered normal-Worker/software route, exact qualified descriptors and finite native/disk watchdog. Actual whole frames must establish menus/world before any game input. Existing Enter limits, identity, 500 ms confirmations, independent UP, freshness, operation bounds and Quit reserve are unchanged. UNKNOWN closes the scope without later endpoint requests, replay or a second original; only an already armed release remains permitted. Movement, fire, enemy consequence, ordinary Quit and a music cause are not established by these preparation controls.

Original outcome and repaired preparation boundary

Native PID 2425757 ran the qualified frozen module and sealed graph after strict publication. Five native instances emitted creation records, but none armed: the real registered route did not make the host log call on which the first consumer had depended for module discovery. Thus neither construction nor an invalidating write was observed. IMUSIC guest threads 2 and 3 exited at EIP zero with previous PCs 0x00cd2ccc and 0x00cd275b, respectively; the retained object register was 0x00c189ec. This reproduces the earlier music failure context but does not identify its invalidating writer or lifetime.

The missing boundary closed the scope UNKNOWN. Zero endpoint requests, captures, binds, Enter or other game inputs occurred. No visible menu/world was established, and no endpoint was contacted after closure. Owned host cancellation produced native exit 0 after 37.144 seconds / 16,205 batches; that is not ordinary game Quit. The watch was unarmed, with optimized tiers still active, so these figures cannot measure selective-watch overhead or be compared as a speed improvement against the older full-block observation.

The discovery gap was repaired offline after closure. The consumer now arms at actual native DLL publication and thread initialization, without wrapping host logging, run, or every import/block. New negative controls verify no-log module publication, original return/exception preservation and unchanged export objects. Authentic native Worker/register/flags/unrelated-write, final Release/free/invalidation and software pixel controls are repeated with the repaired consumer. This is a diagnostic preparation repair, not a proved emulator/music defect, so the conditional additional original is not spent. The closed original is not replayed; actual first invalidation, causal repair, movement, fire, enemy consequence and ordinary Quit remain unproved.

Native and watchdog waits are both actual, exit 0 with both EOFs and empty drain errors. Native-supervise reports leaked:false, no containment signal was needed, and the finite disk monitor's observed minimum was 1,987,796,992 bytes. Final evidence verifies owned processes are nonlive, the control port has no listener, and all media/module/source identities still match. No RAM gate, agents, Codex workers, boxes, Python, Wine, deployment or held investigations were used. Box ID null, state NOT_CREATED, expiry and stop receipt null.