Blood II immutable renderer dependency repair, 2026-11-18

The preparation defect is repaired and published as 2ffa7f555. The single authorized fresh original produced reviewed whole main-menu, character, difficulty and Demo Level 1 images through the immutable loader. Enemy consequence and ordinary Quit remain unproved: Enter on the level card timed out at the unchanged confirmation boundary and permanently closed the runtime UNKNOWN. There was no further endpoint request or replay.

Measured preparation repair

The preceding original remains closed and was never queried or replayed. Its preparation omitted the external pngjs package while gathering relative imports. The production immutable loader correctly refused live resolution; the native CLI consequently had no renderer. The earlier controller tests had used live sources and missed this preparation difference.

tools/blood2-immutable-preparation.cjs now includes the installed pngjs JavaScript files and package manifest in the same indexed immutable inventory as the runtime sources. It uses the unchanged tools/lmarbles-native-source-view.cjs resolver and descriptor transport. No shared loader, emulator, controller, allowlist or deadline was changed. Git plus the retained binary patches recover project sources; the dependency is pngjs 7.0.0, recoverable through the repository package-lock integrity pin.

The self-contained focused test removes that package from an otherwise identical inventory and observes the old failure through the actual loader. A separate inventory missing its entry file also refuses preparation before any original process. The repaired inventory constructs the real Win98Renderer, encodes and decodes a complete 1024×712 raster PNG, and repeats that construction and encoding in a real Node Worker. This tests considerably more than successfully requiring the codec.

The same frozen preparation then runs an authored native PE with a genuine guest Worker. Its complete desktop PNG was personally reviewed. The fixture starts and cancels with actual native wait, both output EOFs and a clean native-supervise receipt. A terminal-controller test rejects a capture before calling the endpoint. Both focused invocations passed; the final invocation also retains the old-failure messages. The canonical locked build passed before the explicit own-path commit and immediate strict pull/push. Published 70064d010 native input/independent-UP proofs were reused without modifying their controller or tests.

One fresh original

Runtime HEAD was 2ffa7f555cff6640dc7f0df8b7c9d79594867458 on main (the exact full SHA is authoritative in runtime-inputs.json). The existing actual module was reused from the preceding retained run, SHA256 7d347a6a631ab77335a14935a533feb896be71aa2508fc5aa19b1e5df8f096ce. It was not rebuilt or downloaded for this original. The actual module read, 23 codec files and all observed source reads matched the immutable inventory. Renderer construction, whole PNG encoding and Worker preflight ran against those exact descriptors before the original's runtime clock.

All 98 owner-designated local BUILD109 game files plus the manifest matched the retained November inventory before preparation, at dispatch and after closure. Client.exe and manifest hashes remain those recorded in blood2-demo.md. The registered blood2_demo arguments and media were used unchanged: -rez B2Demo.rez -windowtitle "Blood2 Demo" -config defaults.cfg. There were no asset copies, configuration edits, cheats, guest-state edits, Wine runs or compilation-archive downloads.

The sole original PID 2151022 began at 01:07:42 UTC, with threads, real ticks, default software D3DIM, 1024×712 desktop and a prearmed 600-second runtime bound with 180 seconds reserved for ordinary Quit. Initial whole captures were black while batches advanced; the menu appeared during the bounded startup wait. Personally reviewed scenes show Single Player selected with BUILD109, Caleb selected, Homicide selected, then Demo Level 1 / Center for Disease Management. Three Enter gestures returned native CONSUMED receipts with independent physical UP and subsequent progress.

The fourth finite Enter gesture, on the reviewed level card, did not obtain confirmation within the unchanged 500 ms boundary. Its independent prearmed release returned ACK and was actually waited, exit 0 with both EOFs. UNKNOWN closed permanently. No later endpoint was requested. No in-level enemy was reviewed and no movement or firing was submitted. There is no claim about enemy reaction/damage/defeat, ambient or friendly damage, objective completion, or ordinary game Quit. Historical ammo 50→48 remains separate evidence and does not establish an enemy consequence.

Cleanup limits and retained evidence

The launch tool session returned signal status 143 at approximately 121 seconds of runtime. Its native and supervisory parents disappeared before publishing native-wait, watchdog-wait or containment receipts. The sender and cause of termination are UNKNOWN; timing alone does not identify them. Buffered native stdout/stderr were also lost. The later owned cancellation request cannot establish how the original exited. No native exit code, native output EOF, ordinary Quit, or original leaked:false claim is made.

The remaining owned finite watchdog was stopped with an identity-recorded TERM. Its original parent was gone, so an actual wait cannot be recovered. All seven known original/group/watchdog/release PIDs were subsequently absent, and the loopback listener was absent through local process-table inspection. This verifies no known survivors; it does not replace the missing actual waits or prove an unrecorded descendant inventory. The focused native controls retain their complete waits and clean containment receipts. The original cleanup-receipt gap is an explicit remaining blocker.

An earlier preparation-only refusal exceeded the default Git listing output buffer while writing recovery metadata. It happened before any original PID or runtime clock. Restricting that run-local listing to runtime paths fixed the refusal; no live-source fallback or deadline extension was used. A local manual-helper duplicate review write was also corrected before submitting input; the existing immutable review was preserved.

Evidence is in scratch/runs/20261118-blood2-frozen-renderer-dependency-repair-combat/. It contains exact Git/source inventories, relevant binary/new-file patches, dependency/module load observations, the retained-module recovery path, selected named whole PNGs, scene reviews, input/release receipts, focused tests, locked-build log and explicit missing-wait records. Redundant captures were removed with hashes and recovery details recorded. result.json is written last. Free /home remained above 1.5 GB in all recorded checks; the final check was approximately 2.42 GB. Continuous original minimum disk telemetry was lost with its watchdog parent and is not invented. No box was created: ID null, state NOT_CREATED, expiry null, stop receipt null. No agents or Codex workers were spawned; no foreign process shutdown or deployment was used. Ownership is released and this task stops without another original.