Stabilize intermittent text-edit IPC timeout in the full Windows regression gate

[RolandUI]

Confirmed failure

Final campaign CI 36025169889, code 2f3894f, Windows full Release suite: 801 passed, 1 failed, 5 explicit native skips. RuntimeTextEditTests.MultilineUnicodeRangeSelectionInsertionAndUndoPreserveSurroundingText failed at RuntimeTextEditTests.cs:108 while issuing the first replace_range request over actual local IPC. Result: bridge_ipc_unavailable / Bridge IPC request timed out; dispatched=unknown, with retained request-id recovery guidance. Test duration was 12 s. Job 107719746236, log retained as artifacts/agent-qa/final-ci-windows-test.log.

The same test passed in the preceding full local Release suite and focused Release/Debug groups. A fresh native direct Unicode edit/replay also passed, with one actual edit independently recorded. Thus this confirms a regression-gate failure, not yet a specific production text-edit defect. Cause is unknown; do not hide it with unconditional retries, skipped assertions or a speculative timeout increase.

Scope

• Reproduce or characterize the failure using repeated focused runs and the full Windows test context.
• Inspect dispatcher/frame preparation, IPC handling and test deadline ownership; distinguish a real product fault from insufficient integration-test time or a transient runner failure.
• Fix the demonstrated cause with the smallest change, preserve mutation idempotency and production timeout behavior unless evidence requires changing it.
• Retain the original failure and validation evidence; update Stabilize AvaScope through native agent exploration, fixes and regression testing #166 only after a fresh combined regression round succeeds.

Acceptance

• Failure cause or externally observable boundary is supported by evidence.
• Any correction has meaningful before/after validation and does not blindly retry uncertain mutations.
• Focused repeated text-edit checks and the complete Release suite pass.
• Hosted Windows gate and downstream platform/native jobs pass on the selected corrected source.

No version bump, release or publication.

Validated checkpoint and blocker (2026-09-24)

Every job of CI 36031948678 passes on a6783cc. Windows and macOS full suites each pass 803 tests plus five explicit native skips; all six native lab combinations and package/provider/installer gates pass. Windows TRX retention is verified: original Unicode case 0.636 s, controlled default-deadline/recovery case 5.054 s. All 12 focused text-edit tests, 15 fresh-process original-case repetitions and the bounded 40-edit payload-size probe pass. New diagnostics retain timing and observed editor state; the deterministic boundary regression proves recovery with exactly one edit and no automatic uncertain-mutation retry.

The original attempt-1 failure remains unexplained and is not relabeled fixed. Its retained log has no per-operation timing or observed editor state; the selected unchanged source no longer reproduces the failure in the completed checks. No product correction is supported by that evidence. Further causal validation needs a reproducible failing environment or a captured recurrence using the now-retained diagnostics. This ticket therefore remains open/blocked at the nearest validated state, with the original evidence preserved. Repeating unchanged passing suites or increasing production timeouts would not establish the missing cause.

[RolandUI]

Starting #171 as the sole active investigation; #166 is in review. The failure is retained from CI attempt 1 before any rerun. The shared text-edit fixture uses the production-default 5-second IPC timeout; the failed test reached its first range replacement at line 108. There is not yet evidence that increasing this timeout or changing dispatcher priorities would fix the cause.

All 15 fresh-process focused repetitions on the unchanged 2f3894f source pass (TRX records under artifacts/agent-qa/text-timeout-repro). I will compare a complete hosted rerun of the same source and retained timing evidence before making any behavioral change. No uncertain edit is retried by the product/test; rerunning the isolated test process creates a fresh application/session.

[RolandUI]

Diagnostic slice committed and pushed as cdaaeeb. The failing assertion now preserves elapsed time, the configured deadline and bounded observed focus/caret/selection/initial-vs-expected-text state. Windows CI writes and always uploads full TRX results. Production IPC timeout, dispatch, mutation idempotency and retry behavior are unchanged.

Validation: a temporary one-tick client deadline intentionally failed the isolated case and retained all diagnostic fields (2.96 ms elapsed; initial text unchanged; expected replacement absent; focus false). The injection was immediately restored and is not committed. The restored normal five-second path passes all 17 text-edit/observation/MCP tests with no build warnings/errors. Negative and positive logs/TRX are retained under artifacts/agent-qa/timeout-diagnostic-* and timeout-diagnostics-positive*.

The unchanged full local context rerun also passes 802 tests with five explicit native skips (11m14s); the formerly failing case takes 0.208 s there. All 15 prior fresh-process repetitions pass. Hosted attempt 2 of 36025169889 has passed Test and is finishing Verify artifacts before downstream jobs. It validates 2f3894f, not this new diagnostic-only commit. Do not cancel it to start another same-branch workflow. After it completes, validate the committed diagnostic/retention changes in hosted CI and inspect the retained TRX.

The original sporadic timeout cause is still unknown. This is improved failure evidence, not a claim that a passing rerun fixed the original event. #171 remains open/in progress; #166 remains in review. No release/publication.

[RolandUI]

Additional isolated timing probe on Avalonia 12.1.3 exercised 40 actual read + range-edit pairs over local IPC for each measured run: 30, 542, 4126 and 8030 UTF-16-unit inputs, ten repetitions per size. Two measured runs verify every edit and final text, with exactly one dispatched edit per request. In the cleanly terminated run, maximum read time is 188.58 ms and maximum edit time 86.19 ms; the five-second production-default deadline is unchanged. Evidence: artifacts/agent-qa/text-ipc-probe/evidence-cleanup-fixed/{measurements,summary}.json.

The first standalone diagnostic completed its measurements but its own HeadlessUnitTestSession disposal needed moving off the completing UI thread. Only the verified owned probe process was stopped, the diagnostic helper was corrected, and the rerun exited normally with zero owned probe processes remaining. This was a diagnostic-harness cleanup issue, not an AvaScope product failure; both evidence sets are retained.

Hosted attempt 2 now has a successful Windows Build/Test/Pack job. Its full test log explicitly records 802 passed, five skips, 9m26s; the Windows native input/dialog artifact records 20 passing checks. Logs and artifacts are downloaded under final-ci-windows-attempt2.log and final-hosted-windows. Linux/macOS downstream validation is still live. No speculative timeout increase or mutation retry was introduced, and the original CI cause remains unproven.

[RolandUI]

Continuing #171 with one deterministic boundary regression: hold an actual TextInput callback until the default five-second IPC client deadline expires, retain the uncertain response, release the callback, inspect actual text, and retrieve the exact original request outcome. Assert one dispatch and one edit, with no automatic mutation retry or production timeout change. This characterizes timeout recovery; it will not be presented as reproduction of the unexplained hosted cause. Validate focused text-edit tests, then the new instrumented full hosted gate after the currently running attempt finishes.

[RolandUI]

Commit a6783cc is pushed. Added a deterministic actual IPC-deadline regression: the application's dispatched TextInput callback blocks until the unchanged five-second client timeout; the response retains dispatched=unknown and the original edit id. After releasing the callback, independent text inspection and exact-payload/id recovery prove one edit, one callback, and a replayed retained outcome. All 12 RuntimeTextEditTests pass in Release (new case 5.058 s), with no failures/skips. Evidence: artifacts/agent-qa/timeout-boundary.log and timeout-boundary-results/timeout-boundary.trx. No production deadline/dispatch/retry change. This is boundary characterization, not a claim to know the original hosted cause.

The earlier unchanged-source CI 36025169889 attempt 2 has passed Windows, Linux, and all three native lab jobs; macOS packaging remains live. Once that run finishes, the full CI gate will run on a6783cc to validate both the recovery regression and cdaaeeb diagnostic/TRX retention. #171 stays open pending that evidence.

[RolandUI]

Every job of unchanged-source CI 36025169889 attempt 2 now passes. Full Windows/macOS each report 802 passing tests and five explicit native skips. Downloaded Linux/macOS artifacts confirm 20 input/dialog checks each and five native tests each; X11 paired CLI/MCP compares all 3840x2160 pixels, while hosted macOS is actual 1x. These successful reruns do not establish the original timeout's cause.

Full CI 36031948678 is now running on a6783cc, covering cdaaeeb's diagnostic/TRX retention and the new deterministic deadline/recovery test. Expected full-suite total: 803 passing plus five native skips. Keep the original failure separate; no speculative production deadline or automatic mutation retry was introduced.