Blood II level-entry causal limit and durable launcher, 2026-11-18

The actual Blood II level-entry failure remains unresolved. No new original was launched: authorization requires an applicable native-proved behavior repair, and the authored loading controls did not reproduce the failure. Movement, fire, enemy consequence and ordinary Quit remain unproved.

The previous scope remains permanently UNKNOWN. Its fourth Enter dispatch, operation, independent release ACK/actual release wait, launch-tool 143 and cleanup evidence were audited without an endpoint request or replay. Missing native consumption/response traces, native stdout/stderr and native/watchdog actual waits remain missing. Neither the failed 500 ms stage nor the sender of termination can be established from that evidence. Historical receipt absence is not repaired retroactively.

Authored native loading control

test/test-blood2-loading-lifecycle.cjs builds its own PE, using no game media or retained-run dependencies. Enter reaches the real native USER consumer, then its guest wndproc performs a 900 ms GetTickCount computation and 4096 actual SetPixel calls before changing its public window title. A genuine guest Worker remains active. Whole before/loaded desktop PNGs were inspected; the loaded scene has the completion caption and rendered color patch. This exercises GDI work, not Blood II's D3DIM level loader.

Preparation uses the unchanged 2ffa7f555 indexed immutable loader/codec helper, including actual renderer/whole-PNG/main-and-Worker preflight before the native runtime clock. The tested module is the existing 7d347a6a631ab77335a14935a533feb896be71aa2508fc5aa19b1e5df8f096ce. 70064d010 controller behavior and all deadlines/freshness/identity predicates remain unchanged; its native forward/turn/Ctrl, independent-UP, lost-ACK and terminal-no-endpoint focused tests also passed.

The final controls retain actual immutable source-read observations, bounded public input traces and host HTTP phase traces. With submitter TERM, Enter was consumed 80 ms after dispatch and the response ended at 81 ms. With submitter SIGKILL, consumption took 166 ms and response completion 167 ms. Host application preceded native consumption; complete loaded captures followed consumption by 977 and 1227 ms respectively. Independent release and the complete finite operation passed. Expensive guest loading/rendering alone therefore did not reproduce a confirmation failure. These controls do not identify which stage failed in Blood II and do not earn a guard change or another original run.

Durable ownership repair

tools/blood2-durable-launcher.cjs introduces an app-specific detached receipt owner. The submitter completes immutable preparation, then transfers the same read-only descriptors through exec. The detached owner has ignored stdin and regular-file output; it retains the existing Linux native-supervise.cjs containment (actual available level: pdeathsig). The supervised run prearms a separate finite watchdog before native launch. Native and watchdog pipes are streamed to files as observed and fsynced at close, rather than held solely in the submitter's memory. Their direct parent records actual close waits and both EOFs. The outer owner reports missing waits as missing, rather than assuming successful receipt retention. Worker cancellation remains host cleanup, never ordinary guest Quit.

Both final controls deliberately kill the submitting process before native interaction: TERM in one and SIGKILL in the other. Each detached run remains controllable, completes Enter/loading, and retains native exit 0, watchdog exit 0, actual waits, separate stdout/stderr EOFs, empty drain errors and leaked:false containment. The native processes receive worker-owned orderly cancellation after the test scope closes. This demonstrates current submitter-death survival; it does not identify the historical signal sender or establish the historical native exit.

Validation and evidence

Evidence: scratch/runs/20261118-blood2-level-entry-input-loading-cause-repair/. The final dispatch-verified-controls pair contains timing analyses and actual source-read inventories. Earlier authored attempts remain explicitly distinguished. One added assertion mistakenly counted prearming against the 500 ms confirmation window; it was corrected to measure from DOWN dispatch. The production controller's own unchanged confirmation guard passed.

Focused lifecycle/loading tests, published native controller controls and flock /tmp/wine-build.lock bash tools/build.sh passed. Canonical build output retained the same module hash. All 98 BUILD109 files plus the manifest still match the retained pins; arguments/configuration were not changed. Exact Git identities, relevant binary diffs, owned new-file patches and selected named PNGs are retained; no source snapshot, module duplicate or media copy was made. An unrelated unexecuted inventory member changed during initial controls; that limitation is recorded separately from verified actual source reads.

All owned native/watchdog scopes were closed and checked without querying closed endpoints. No box was created: ID null, state NOT_CREATED, expiry null, stop receipt null. Native controls remained above the 1.5 GB disk floor. After all owned runtimes had closed, shared free space fell to 1,488,637,952 and then 1,250,779,136 bytes during finalization. The board was alerted; this worker's evidence occupies 4.6 MB and foreign files were not deleted. No new runtime was launched after the floor breach. No agents, workers, Wine, Python or deployment were used. The causal blocker is concrete: the actual original's native/HTTP stage evidence is absent, and the applicable authored native loading control passes without a behavior change. There is no demonstrated old-fail/new-pass level-entry repair. Ownership is released after publication and this task stops.