Quake II staged-root repair and upper-route attempt, 0820

The concrete 0754 ENOENT is fixed in the exact run inputs and exercised through the actual changed remote-parent invocation. Original guard/onward-door progress was not attempted: the whole native fixture failed ordinary joined shutdown. The first unexpected failure was terminal; no later input, capture or retry followed. Evidence: scratch/runs/20261109-quake2-staged-root-upper-route-0820/.

Concrete root fix and invocation evidence

This defect belonged to the run-specific native-fixture.cjs preload and its control.py launch branch, rather than the tracked production caller. The preload no longer derives its root from process.cwd(). The launch branch passes explicit QUAKE2_STAGED_ROOT from the root whose full provider manifest has just passed native-startup-barrier.require. Before opening the canonical EXE, the preload requires an absolute root owning its own fixture path, a run inside that root, and the barrier's exact root and manifest SHA. Missing or wrong roots fail closed. The runtime root was /home/user/q2-staged-0820.

offline-invocation.py compiles the actual launch branch using the production Q2 Caller, retains the generated replacement invocation of tools/quake2-native-launch-client.py --remote-parent, and executes it from unrelated cwd /home/user. The positive case reaches the composed fixture's DRIVER_PREFLIGHT exit with chromiumStarted:false. Missing-root and wrong-root variants fail in preload before Chrome. All three parents and remote-supervisor children actually waited with separate stdout/stderr EOF. These offline tests authenticate actual local provider bytes; they do not qualify frozen native providers or claim original gameplay. The new runtime separately staged and verified its exact frozen production closure. The prior ENOENT was not retried inside its terminated predecessor runtime.

Recovery keeps Git base plus a small run-input patch, the zero production-source patch, exact pinned-provider Git revisions, and existing media/module path/SHA references. No tree, media or module archive was made. No tracked launcher/helper needed modification. The run inputs are ignored evidence; this tracked note records their recovery and outcome.

Actual fixture, including the terminal shutdown blocker

The 723-provider closure retains caller 1473c5952a52d5d2b4546b93a0d462b64fe5c6cd, custody 4341b87723a7d56bc98223215cca4cdc878fcae2, reader 5bc5f86f9 integrated in terminal repair 4ba9f3a1b08c11219c50023f4abffb3bf6c6f9cf and the exact historical host/Worker providers. CPU remains 2185596 bytes, SHA-256 a320d4c85156644fd12b6134d66b11784f3100b9b135756206c087cdacd43295. Source-matched accepted reader/composition/PIPE/build proofs were reused rather than rerun as separate qualification tasks. Production code did not change.

The actual retained PIPE caller launched the changed staged-root path. Chrome, native keyboard/type/combo preflights, strict observation flags and the owning Worker/module/instance/endpoint1/slot0 gates ran. PreCPU/preDLLmain records have zero execution entries before release and NOT_ARMED native stores. The startup capture identifies Worker target D21BB504882C4F7CB7254CFA1A5B7688, instance 9b2ebd66-225f-4517-8b55-82eb991755bb:1, and the exact CPU hash above.

I personally reviewed startup.png, fixture-comma.png and fixture-period.png. They show the expected desktop/readme for the authored idle fixture; they are not Quake II gameplay. Each 500ms key command explicitly used requireMotion:true and captureHeld:true. Comma had four wholly-held physical-high samples; period had three. All held passive/motion reads were SNAPSHOT, both reverse key releases ACKed, fresh post-action captures showed physical keys zero, and both numbered command ACKs succeeded. The fixture initializes authored received-frame bytes in native code; this proves neither original command consumption nor movement/contact/damage cause.

Ordinary quit then failed. The browser-close receipt records closeRequest:ACK, a sent SIGTERM, Chrome terminal code0/signal null, drained true, elapsedMs:12113, capMs:19827, and normal:false. The first browser terminal record is 2026-11-09T08:21:53.772Z, abnormal browser shutdown; UNKNOWN; no replay; the caller then latched owned transport failed: request-fixture_quit. There was no successful quit ACK. No cap was extended or SIGTERM result promoted to ordinary success. The reason the close helper required escalation remains unestablished; latency or a guest defect is not inferred.

Whole-fixture qualification therefore FAILED. Independent no-input caller cancellation joined the existing owner; no original launch, route, game command, gameplay PNG or later capture occurred. The next concrete blocker is ordinary joined quit through this actual caller, before any original upper-route attempt.

Retrieval and final lifecycle

The exact Redline0756 UUID 69ce5bd6-8e7d-4300-a29a-4f0bd7715acf ARCHIVED PHYSICAL_DRAINED release and stop stop_f4cdb22d404ad552 completed 08:08:42.905Z were checked before create. One box only: bx_wckgsd9r, UUID c9b59dc4-de93-4b60-851c-509654899ca9. The observable independent immutable watchdog was prearmed before create, with 25-minute deadline 08:43:19.195Z. Create returned TTL1493 and initial expiry/archiveAfter 08:43:18.904Z.

All 79 remote regular-file rows, 57 newly stored unique contents, were retrieved and independently SHA/size reread before archive. All 11 remote gzip rows passed GNU gzip CRC and independent decompression/hash. Kernel custody records 171 owned / 171 actual reaps, complete true and command-log EOF true, with no remaining/unreaped/unresolved entries. Both remote supervisor and caller parent actually waited1 with separate stdout/stderr EOF. Local caller children actually waited -9,0,0,1,0 with separate EOF; the retained controller waited1. These are failed-run cleanup receipts, not normal qualification.

Ports 37903/33145 were closed, profile writers empty and runtime/input locks free. The profile directory remained. Physical absence supports this census and never replaces actual wait evidence. All-writers/nonleader coverage and lifetime edict ABA remain UNKNOWN; native lifetime remains REVOKED.

Final state ARCHIVED, stopped 08:22:50.540Z. Receipt stop_f07c5c431fca7a35, completed 2026-11-09T08:22:50.631Z, forced false, error null. Final archiveAfter is null; expiresAt was not returned. The independent watchdog and lifecycle supervisor actually waited0 with separate EOF. The local runtime lock was released and PHYSICAL_DRAINED posted at 08:23:16Z. A local reporting attempt used a nonexistent completedAt field; it was corrected to the actual endedAt field without any runtime retry.

Outer logs were gzip drains from first byte. Final local verification includes PNG chunk CRC/decompression, retained source/provider pins and reused proof hashes. Free /home stayed above 1.5GB. No agents/workers, peer/shared host/custody/ WAT edits or public deployment occurred; Heroes II and Claude were untouched.

The 0635 damage cause remains UNKNOWN. No blind descent, crate-side turn, ladder or Space assumption was attempted. Received-frame origin/velocity/raw flags SNAPSHOT is not contact, consumption or cause; no flag names are invented. Single-task STOP.