fix(e2e): stabilize shard-3 diary flakes

[steilerDev]

Summary

• Root-caused and fixed the recurring E2E shard-3 (16-shard desktop matrix) diary flakes that were blocking beta -> main promotions on PRs release: promote beta to main (cost basis filter, diary fix, auto-itemize refactor) #1791/fix(deps): scope js-yaml override by version range to unbreak docs deploy #1828.
• diary-drafts.spec.ts Scenario 14: test.slow() only triples the overall per-test timeout, not per-step actionTimeout/expect.timeout — the test paired a 45s ceiling with several 30-45s step timeouts, starving the finally-block cleanup under CI load. Replaced with an explicit test.setTimeout(60_000), proportionate 15s step ceilings, and a guaranteed-cleanup-time pattern (testInfo.setTimeout(testInfo.timeout + 15_000) + try/catch) in the finally block.
• diary-r2-uat.spec.ts Scenario 10: fixed a real race where a stale in-flight response from a prior UI transition could satisfy the next waitForResponse() predicate, leaving a shared requests[] capture array empty. Now awaits each transition's own response explicitly.
• document-linking.spec.ts Scenario 7a: was collateral maxFailures:1 cancellation from Scenario 14, not an independent bug — added the same defensive cleanup-time pattern anyway.
• playwright.config.ts: comment-only documentation (verified against Playwright 1.61.1 dispatcher source) that maxFailures already tolerates one retry-passing test before fail-fast cancels a shard.
• cornerstoneContainer.ts: incidental fix for a testcontainers@12.0.4 API rename (withCopyFileToContainer -> withCopyFilesToContainer) discovered during local verification.

Fixes #1829

Test plan

• eslint clean on all touched files
• prettier clean on all touched files
• tsc --noEmit in e2e/ shows no new errors vs baseline (one pre-existing error in a touched file is incidentally fixed)
• CI Quality Gates
• Full E2E matrix, shard 3 in particular, green

Co-Authored-By: Claude dev-team-lead (Sonnet 4.6) noreply@anthropic.com

[steilerDev]

[product-architect] APPROVED

Reviewed as the always-on architect reviewer for this e2e-only change (per the PR Review Gate, security-engineer/ux-designer correctly don't trigger). Scope is clean: only e2e/ and agent-memory touched — zero production (server/, client/, shared/) changes, so no API-contract/schema/architecture surface is affected. I verified the technical claims rather than taking them on faith.

Test-strength preservation — confirmed, no weakened assertions:

• diary-r2-uat Scenario 10: the full assertion block is intact — type param present, contains ≥1 manual type, and contains zero automatic types. Reading type off manualResponse.url() (the exact response the Manual click produced) is stronger than the old shared-mutable requests[] array, which is precisely the race being removed. Good fix.
• diary-drafts Scenario 14: the 45s→15s step-ceiling reductions are not a weakening — assertions fire only after entryLoadPromise confirms the API response arrived, so the remaining work is a client-side React re-render that completes in well under 15s. The 45s values were budget-reservations masquerading as ceilings; tightening them is correct.

Timeout arithmetic — sound. 60s test budget + proportionate 15s "something is actually stuck" ceilings + additive testInfo.setTimeout(testInfo.timeout + 15s) cleanup extension. Because testInfo.timeout is the full original budget and the body can consume at most that before timing out, the finally block always gets ≥15s of guaranteed teardown headroom. The test.slow() correction is accurate (it triples only the overall per-test timeout, not actionTimeout/expect.timeout), as is the maxFailures note (a retry-passing "flaky" test never increments the failure count — only all-attempts-failing tests trip fail-fast).

Incidental testcontainers fix — verified genuine. withCopyFileToContainer (singular) is gone from testcontainers 12.0.4; withCopyFilesToContainer(filesToCopy: FileToCopy[]) is the correct replacement. The proxy-CA path would not compile/run without this, so it's a necessary carry, not scope creep.

Reusable patterns — worth codifying, and correctly placed. Both the guaranteed-cleanup-time pattern and per-transition response sync are genuinely reusable e2e conventions. They belong in the e2e-test-engineer's domain, and the PR already lands them in general-e2e-patterns.md. No architect-owned wiki page covers E2E test conventions, so no wiki update is required from me — agent memory is the right home. (If these keep recurring, a dedicated "E2E Conventions" wiki page could be considered later, but that's not a blocker.)

5-consecutive-runs criterion — pragmatically satisfied. A full 16-shard × 3-viewport matrix isn't run pre-merge on beta PRs, so 5 green pre-merge matrix runs aren't achievable here. More importantly, this PR does real root-cause analysis (the test.slow() misconception + the stale-in-flight-response race) rather than a retry/timeout bump — that is a materially stronger signal of a durable fix than 5 green runs of otherwise-unchanged flaky code. On-PR evidence (all 3 target tests green on shard 3 at 1–2s) + post-merge monitoring is a reasonable close-out, and the eventual beta → main promotion PR's full E2E Gates run serves as the confirming pass.

Non-blocking / informational:

1. The testInfo.setTimeout comment phrases the extension as "additive to the current remaining deadline" — Playwright actually sets the deadline relative to test start (start + timeout), not to remaining time. The end guarantee (≥15s cleanup headroom) is correct because testInfo.timeout is the full budget, so the wording is a touch loose but not misleading.
2. In Scenario 14, the four 15s step ceilings sum to the 60s test budget, leaving no theoretical slack before cleanup — but realistic step durations are 1–2s and cleanup gets its own +15s extension, so this is only a pathological-worst-case concern.

No architecture, contract, or schema implications. Approving.

[github-actions]

🎉 This PR is included in version 2.13.0-beta.24 🎉

The release is available on GitHub release

Your semantic-release bot 📦🚀