Q3 raw startup evidence, 1648

The missing native evidence is recovered: request 6 returned a valid reply on the correct page session 7 ms after its unchanged 20-second deadline, after the Browser.close reply. The authenticated production composition records that late reply before routing discards it. Chromium's internal delay cause remains UNKNOWN; no startup-policy repair is justified. Arena movement/firing remain unverified.

Evidence: scratch/runs/20261109-q3-raw-startup-evidence-1648/. Git is the source archive: exact base plus owned.patch; run-specific Python inputs have retained references, SHA/size and small patches. No source tree or media archive is created.

tools/q3-current-module.cjs supplies an assertion-anchored transform applied by tools/q3-current-foreign-closure-v1.cjs to the authenticated generated consumer. It records exact outbound/inbound JSON, request ID, method, reply and request session IDs, protocol error/result, pending-match status, whether an ID was previously sent, and target ID/type/URL/waitingForDebugger. Inbound frames are recorded before parsing/routing, including unmatched, duplicate and late responses. Separate timeout, send-error and transport events retain their request/session context. Logging ends after a successful navigation reply; a terminal startup continues recording through cleanup. Existing routing, first-failure propagation, startup policy, debugger policy, CPU/provider authentication and deadlines remain unchanged. A peer's waiting flag does not model a paused Chromium renderer and is not used to infer one.

The production peer regression now uses fresh native-lifecycle preparation, the actual generated relative-import entry, and the native cause/input-wire hooks. Ten cases exercise single and duplicate page sessions, event ordering, waiting flags, protocol error, socket closure, an unchanged 20-second timeout, late reply, and duplicate/unmatched replies. Every inbound/outbound envelope is checked against the peer's exact transcript; late and duplicate replies remain unmatched without reviving setup. The real stand-in browser and caller actually wait with separate stdout/stderr EOF. No local Chrome, navigation, guest input or capture runs in these controls.

The full lifecycle check caught an initial diagnostic insertion breaking the input-wire hook's literal transport seam. The insertion now preserves that seam; both generated lifecycle paths and all ten peer cases pass. Initial invalid test assertions (plain versus compressed log name, a duplicate-reply case accidentally classified as a timeout) and an early stale-manifest failure from editing the provider while a test ran are retained as failed controls. They are not native evidence. An integration invocation without its required foreign selector and a concurrent negative-control test's 20-second timeout are also retained; production deadlines were not extended.

Fresh own 942-provider preparation independently hashes manifest af96d84fbb3b1483fe06e0ffd752c9b957d93db9140d6931e42d3f6989f3d368. The actual own provision body verifies all upload bytes and generated remote prelaunch syntax offline. Both actual own generated launch paths execute the prelaunch/custody/staged consumer with Git unavailable and the existing DRIVER_PREFLIGHT exit before Chrome. Kernel reaps and distinct outer/runtime parent waits with separate EOF pass. The production controller regression executes the actual fresh generated caller's retained original branch and preserves first failure with successful and failed read-only joins; later capture remains unrequested. The current foreign-closure suite passes all 27 refusal controls. The locked build passes.

The final peer regression also reproduces the native ordering: Browser.close ACK, then request 6's valid late ACK. Setup stays terminal and the late frame retains its original request method/session with matched:false.

Both Q2-reviewed1638 and Redline-validation1648 published RELEASE STOP before allocation. Their unique retained artifacts were independently checked for compressed and decoded SHA/size and gzip CRC, physical drain, exact API archived box/normal unforced stop, and actual watchdog join with separate EOF. Neither had started a native runtime: their kernel and runtime parent waits are NOT_CREATED. Redline's separate runtime/writer lock predicates were not present in its retained census and remain UNKNOWN, alongside its historical browser EOF; no inference replaces those missing predicates.

The one own box was bx_882xv5qc, UUID 2a330650-f9b4-46e0-8e8b-2782b940e48e. Its independently observed immutable 25-minute watchdog was armed before allocation; initial expiry was 2026-11-09T17:28:03.235Z. All 942 production provider uploads and both actual remote prelaunch import checks passed. An offline-only upload journal was moved into the offline evidence directory before allocation to preserve the native writer's fresh write-once path. No production path or provider pin was changed to accommodate that isolation.

The single diagnostic cancellation fixture reached native Chrome 151.0.7922.108 and failed before readiness. The raw stream establishes:

UTC time Native observation
17:04:17.595 Root request 3 auto-attach replied successfully.
17:04:17.597 Manual request 5 attached the about:blank page and returned session 5DBBEBD190A971BE915FA3D4AC74FA1E.
17:04:17.598 Page-session request 6 sent Target.setAutoAttach on that same session.
17:04:37.598 Its unchanged 20-second deadline removed the pending request.
17:04:37.599 First browser failure recorded CDP UNKNOWN; no replay.
17:04:37.600 Cleanup sent Browser.close (7), which replied successfully.
17:04:37.605 Request 6 returned {result:{}} on the correct session, recorded as a known but unmatched late reply.

There are four native attachment events: two sessions on the same about:blank page, and two browser_ui omnibox-popup targets. All four report waitingForDebugger:false. This is actual browser evidence, not the earlier peer's invented target reconstruction. Actual renderer pause state remains UNKNOWN. Browser UI target-info events arrived during the outstanding request, so inbound event delivery continued. No protocol error or wrong-session reply explains this run's request 6 failure. Its valid result arrived after the pending deadline, rather than being permanently absent. The ordering after Browser.close does not prove that shutdown released a Chromium dependency. The internal service-delay cause remains UNKNOWN; changing debugger policy, adopting another session or extending the deadline would be speculative.

The owning observer cancelled on the first browser failure. The caller exits 1 after its read-only joins and preserves that first error. No Page.enable, navigation, guest input, screenshot, original arena, retry, second box or deadline extension followed. The new fixture closes the raw-reply evidence gap but does not establish a safe causal startup repair or gameplay qualification.

Kernel custody records 120/120 actual reaps, complete with no failure or unreaped identities. Three distinct generated-parent, prelaunch-child and runtime-parent waits have separate stdout/stderr EOF. Independent physical drain verifies ports 36063/39625 closed, profile absent, no runtime FD holders or input writer lock, and runtime lock drained. Normal browser-close EOF and nonleader/untraced-clone auditing remain UNKNOWN. The empty native command log is retained and is not used as EOF proof.

All 43 unique retrieved artifacts were independently reread for compressed and decoded SHA/size and gzip CRC before delegated normal shutdown. The exact final API box state is ARCHIVED, archiveAfter null, expiresAt NOT_RETURNED, normal unforced stop_6fd590a9e08ece24 completed 2026-11-09T17:05:35.055Z. The independent watchdog actually joined 0 with separate stdout/stderr EOF; its configuration hash stayed unchanged and its lifecycle parent returned 0 and released the local lock. No inherited box was mutated, deleted or resumed. No local Chrome, workers/agents or deployment ran.

Next work is offline Chromium service-delay diagnosis against these exact raw frames. This task makes no further runtime attempt and ends with final published RELEASE STOP.