Q2 auto-attach causal isolation, 1150

Delivered a tested extension of the existing Q2 observer, not a native causal repair. Q21135 cannot identify which of the two actual Target.setAutoAttach requests timed out. No new native attempt or box contact occurred. Native REVOKED; all-writers/edict ABA UNKNOWN. SingleTaskSTOP.

Evidence: scratch/runs/20261109-quake2-autoattach-cause-repair-1150/. Git base: 14381d50ee2832147b12bcfd72c33fdf097ae5b4; recovery is that commit plus owned.patch, including new test and this note. No archives, source/module copies or media copies. Retained input identities are in retained-facts.json.

Retained facts and exact production order

Q21135's startup-readiness.jsonl.gz reports DevTools ready after 24,717ms. Its first transport failure is 11:19:07.020Z; the browser failure file follows at .021Z with CDP UNKNOWN Target.setAutoAttach; no replay. The input observer clock anchor is 11:18:19.373Z / Node 1540.790223ms; first terminal is Node 49187.620208ms. Socket error/close are later at Node 55971.003832 / 55971.091423ms. Those later transport events cannot establish a transport failure before timeout. No retained raw setup request/reply, session, browser task timing, keyboard hold, READY, or original launch exists. Browser service/application times are UNKNOWN.

The retained 0105 consumer, current production composition, eager 1135 health wrapper and outer observer generate this actual sequence, before keyboard input:

  1. Browser.getVersion, browser scope.
  2. Target.setDiscoverTargets, browser scope.
  3. Target.setAutoAttach, browser scope, autoAttach:true, waitForDebuggerOnStart:false, flatten:true.
  4. Target.getTargets; require exactly one page.
  5. Target.attachToTarget for that page, flatten:true; assign returned sessionId to pageSession.
  6. The same Target.setAutoAttach parameters on pageSession.
  7. Page.enable, then Runtime.enable, on pageSession.
  8. About-blank native keyboard preflight, then app selection and owning load.

Automatic attachment emits a distinct session in the test control; the consumer only handles worker attachments and obtains its page session from explicit attach. send assigns globally increasing IDs and optional top-level session IDs. Incoming replies settle by ID alone, without validating session ID. Neither a wrong-session reply nor browser task delay is established in retained native facts. Changing session routing or attachment ordering would therefore be speculative. Request timeout deletes its pending entry before observeTerminal; the old empty pending snapshot is not evidence that no request was outstanding. The reused Q3 recorder deliberately filters Input.*, explaining why the old Q2 setup transcript is empty, despite zero observer drops/errors.

Narrow diagnostic

Extend tools/quake2-input-wire-observation.cjs in place. The existing opt-in preload and authenticated provider references remain the integration path; no second recorder/preload, transport mutation, additional CDP call or startup wait. The same input-wire.jsonl now includes a bounded startup window:

Raw setup values are truncated at 4096 characters with byte length/truncation metadata. Input payloads/expressions are not added. Node timer overrun reports Node task scheduling delay only; it never proves browser service/application latency. Diagnostic sink failures cannot settle/reject a request or block a send. Late raw replies remain observations, never ACKs. Existing input recorder and independent UP, lifetime/readiness/caps, health and first-HP-loss gates are intact.

Verification and limitation

tools/quake2-startup-cause.test.cjs executes the actual health-composed production startup/error/cleanup code using existing inert Node process/server/CDP seams. It stops before input/app selection. Only the test timer seam accelerates the 20,000ms timer to 60ms; the production source and recorded budget are unchanged. Full generated source and tested boundary-source SHA-256 identities are retained, without source copies.

Nine old-missing/new-evidenced pairs pass: normal setup, root timeout, page timeout, protocol refusal, synchronous send throw, socket error, late reply, delayed reply, and wrong-session reply. Every pair preserves exact request order and failure; all use normal process cleanup and separately ended stdout/stderr. Tests assert both auto-attach scopes/flags, context ordering, pending before deletion, raw late/wrong-session reply, no input/READY/navigation, and actual Browser.close request without a fabricated close ACK. Additional sink failure, bounded storage, truncation, unmatched reply and deterministic 25ms Node timer overrun controls pass. This is oldFAIL/newPASS for missing diagnostic evidence, not for native causality.

Existing observer's 3 provider preflights, 8 paired generated keyboard startup controls and sink failure pass, preserving independent UP and sticky failures. Actual composition/transport/input/quit regression, passive damage reader tests, and flock /tmp/wine-build.lock bash tools/build.sh pass. Failed test-development runs are retained and excluded from passing claims.

Constraint violation: I mistakenly invoked tools/quake2-browser-startup-failure.test.cjs as an assumed inert regression. It launched local Chrome and browser Workers (no guest instructions/original). This violated OFFLINE/no-browser and is excluded from offline validation claims. startup-regression.log retains its prearmed watchdog and normal exit0 cleanup; accidental-local-browser-cleanup.json independently confirms Chrome PID 1653127, watchdog PID 1653076 and its temporary profile absent. No box was created, queried, resumed or contacted. No further browser test was run.

Next: review/use the new startup evidence only in a separately authorized fresh whole fixture; no unchanged native retry is authorized by this diagnostic. Prior Q21135 box bx_py2e33kc, UUID 441ad5bf-1891-4ddd-97b4-074bb54bfb50, remains retained UNCONTACTED: recorded ARCHIVED/PHYSICAL_DRAINED, normal stop stop_c13a36e5850d38ad, initial expiry 11:41:40.730Z, completed 11:19:45.275Z. No new box/state/expiry/stop receipt exists for this task.