Q3 auto-attach service-delay investigation, 1723
No causal production defect is established, so there is no production repair or old-fail/new-pass claim. The exact missing observation is request 6's transport submission/ingress and browser-side dispatch/completion dependency. The existing first-error, cleanup, routing and 20-second timeout stay intact. No further arena attempt is justified by these offline controls.
Evidence: scratch/runs/20261109-q3-autoattach-service-delay-repair-1723/.
Source recovery uses gitbase.txt and owned.patch; Git holds the source.
causal-audit.json records retained compressed/decoded hashes and sizes.
No source/media archive, Chrome, game, box, agent or worker was created.
Box ID/state/expiry/stop receipt: NOT_CREATED / not applicable.
Actual caller and scheduling
The starting evidence is
q3-raw-startup-evidence-1648-20261109.md.
The recovered startup stream is the exact gzip with SHA-256
939c96e8dd687c6541b5876e3e798af94b8032108a1795668b4f31bee1e16e5a;
its decoded SHA-256 is
8b3a6c9826416951b5e73c696f2b720fc146c7056cdb7bfa9649c2d5fcec646e.
Both hashes and gzip CRC were checked again. The retained await trace is
cancellation/recovered/5ce8e69344fd3c32eb8c9b4be6a8aa6b734c123ea2c409ea4486d5c746ce5538.gz
under that predecessor run; its hash and decoded reference are in the audit.
The actual generated entry calls q3-current-foreign-closure-v1.runCaller,
which authenticates providers and installs the composed consumer, native
cause/input-wire hooks and execution-boundary WebSocket wrapper. The owning
generated cancellation runtime explicitly enables both diagnostic hooks.
The controller observes the browser failure and cancels/joins; it does not
service CDP events or supply a missing startup continuation.
Inspection of the composed consumer and installed hooks establishes:
sendputs the request inpendingbefore callingws.send. The generated handler resolves/rejects matching IDs directly. Event handling is invoked separately with a rejection observer; no event promise is awaited by the message dispatcher.- Only
workerattachment events start Debugger/Runtime setup commands. All four retained attachments arepage,browser_ui,browser_ui,page. No such worker setup runs in this transcript. Target-info changes have no command-producing branch. These events cannot create the hypothesized recursive auto-attach sequence in this caller. - The manual attach request names the about:blank page. Request 5's returned session is exactly request 6's outbound session and its eventual reply session. There is no scope substitution in this run. General routing is keyed by ID, but no wrong-session response caused this failure.
- The deadline snapshot has only the enclosing
cdpspan and request 6'scdp.requestspan active. The raw stream has exactly seven outgoing commands, including cleanup's Browser.close. No Debugger, Runtime, Page or Input command is hidden behind an event continuation. - The execution wrapper only rewrites selected Runtime.evaluate expressions; none was sent. Its propagation suppression requires its own nonce-bearing Runtime.consoleAPICalled marker. It does not suppress this Target reply. Input-wire observation adds no commands. Screenshot probe work requires a screenshot, which was never requested.
These are local code/trace findings, not proof that Chromium serviced request 6 promptly. Target-info events received about 11.6 seconds into the wait and the on-time deadline rule out a continuously blocked Node handler throughout the whole interval; they do not exclude shorter stalls or transport buffering.
Discriminating offline controls
tools/q3-page-session-startup.test.cjs retains all ten original cases and
adds three exact-native-frame cases. Each checks the first six outgoing JSON
envelopes byte-for-byte against the retained stream, replays the actual four
attachments and intervening target-info events, and preserves their native
IDs, target types, sessions and waiting flags.
Synchronous ACK delivery and next-event-loop-task ACK delivery both settle request 6 and reach the intentional pre-navigation protocol refusal. Delivering the same ACK only after the unchanged 20-second timeout and Browser.close ACK keeps setup terminal: it is known/unmatched and never triggers Page.enable, navigation, input or capture. Cleanup's child exits normally, the caller exits 1 with its first error, and separate stdout/stderr EOF and profile removal are checked. No timer, routing or debugger policy is relaxed.
The original peer suite had a coverage gap: it selected the native phase but
did not explicitly set Q3_NATIVE_INPUT_CAUSE=1 and Q3_INPUT_WIRE=1, and it
replaced createChain with an empty object. The test now enables those exact
runtime flags, asserts both hooks in the actually compiled consumer, records
its hash without copying source, and delegates to the actual transformed
createChain initializer. This is a test-coverage correction, not the cause
of the production delay. An initial 13-case pass before this correction is
retained separately as peer/; the final peer-native-hooks/ pass is the
reported coverage.
The peer still replaces browser/transport service and the static server; the real generated consumer, native diagnostic hooks and child lifecycle run. ACK timing is an explicit controlled input. Success demonstrates that these Node paths can consume the reply; withheld/late ACK failure reproduces the symptom only. Neither is a repaired Chromium implementation or native arena qualification.
Validation: all 13 final peer cases; actual generated owning caller integration
including first-failure propagation and read-only joins; both actual generated
prelaunch/custody/staged lifecycle paths at DRIVER_PREFLIGHT before Chrome;
current foreign closure with 27 refusal controls; locked tools/build.sh.
Exact remaining causal observation
The startup outbound record is written before WebSocket.send; the await
trace measures caller settlement, not a socket write or browser receipt.
There is no request-6 send-return timestamp, kernel/socket ingress timestamp,
or browser handler/dependency trace. The valid result was observed after
cleanup, but neither its browser production time nor its residence in a
transport queue is retained. Browser.close releasing a dependency is unproven.
To distinguish these alternatives requires correlated request-6/session
transport entry/return and ingress timestamps, Node scheduling observations,
and browser-side command dispatch/completion or the pending dependency that
holds completion. Renderer/debugger state would need its own direct evidence;
waitingForDebugger:false is not that evidence. This task authorizes no live
attempt to collect it. Stop with this missing observation rather than remove
auto-attach, adopt another session, resume an invented paused target, extend
the timeout, or declare a peer pass a production fix.