[matt2e]

Failing test

src/app/navigation.test.tsx "keeps a return path for addressed plugin settings and retires it on Close" (added by #2491) fails intermittently in the Vitest flake-check lane:

Error: Test timed out in 5000ms.

It is one long journey on a full App mount with the default 5000 ms timeout: open the Plugins list, open Pair mobile, Back, Forward, open Appearance, open Pair mobile, Close, Back, Back.

Why this is a timeout flake, not a race

The test is slow all the time, and every failure is the same timeout. No assertion has failed.

• Main: 25 green JavaScript shard runs since Offer plugin updates from the Buzz shell #2491 took 2148–4414 ms (median 3921 ms). That leaves under 1.2 s of headroom.
• Flake-check failures: each one was a cold iteration 1 that ran past 5000 ms. The same test then passed iterations 2–5 at its usual speed, with no product or test change between them.

Run	Iter 1	Iters 2–5
#2718 42217828904, Flake check 1/2	× 5045 ms	4061–4133 ms
#2718, Flake check 2/2	3796 ms	3377–3581 ms
#2729 42194107421 attempt 1	× 5055 ms	3973–4399 ms
#2729 attempt 2	× 5224 ms	4016–4423 ms

• Neither PR touches this test's path. fix(navigation): let bare buzz:// open Buzz without a broken-link notice #2718 changes deep-link ingress and perf(messages): cut main-thread work while a busy channel is open #2729 changes messages and relay code. The flake check selected this file only through imports (navigation.test.tsx → services.ts → deep-links.ts). perf(messages): cut main-thread work while a busy channel is open #2729's attempt 3 passed.
• Local near-miss: in the local baseline below, one iteration on unmodified main took 4920 ms.

What this changes

The test is split where its title already splits. Each new test mounts its own App with the file's usual per-test setup. The 5000 ms timeout is unchanged.

1. "keeps a return path for addressed plugin settings through Back and Forward": open the list, open Pair mobile, Back, Forward. These are the original's first-half assertions, verbatim.
2. "retires every visit to addressed plugin settings on Close": open the list, open Pair mobile, open Appearance, open Pair mobile, Close, Back, Back. These are the original's second-half assertions, verbatim.
   • The test keeps one earlier Pair mobile visit behind Appearance. dismissPluginDetails retires every matching visit, and the final Back landing on the list is what proves it.
   • Without the earlier visit, that last assertion is vacuous. With Close mutated to retire only the current visit (navigation.dismiss([navigation.snapshot().entry.id], settings)):

Version	Result with the mutated Close
Original test	fails
Split without the earlier visit	passes (misses the regression)
This split	fails

Every assertion in the original test is kept in one of the two tests. Production code is unchanged.

Verification

All runs were local, on macOS (16 cores). The flake-check runs happened while the machine was under heavy unrelated load (load average 35–40).

• tsc --noEmit and biome check --error-on-warnings src/app/navigation.test.tsx pass.
• Each new test passes run alone (-t), and the whole file passes under --sequence.shuffle with seeds 1–3. Seed 2 runs the Close test first.

Per-test timings (ms). The command was BUZZ_TEST_WORKERS=2 ./bin/node scripts/flake-check.mjs vitest --iterations=5 <files>. The file list is the same 15-file selection as #2718's failing lane.

Test	1	2	3	4	5
Before: original (on 3bcf1c4bb)	2407	1994	2487	2037	4920
After: Back and Forward	1399	2071	1538	906	1384
After: Close	1951	2966	2131	1496	1801

Running navigation.test.tsx alone the same way: Back and Forward took 848–1595 ms and Close took 1352–2114 ms, with 0 of 5 iterations failing.

Headroom. The local worst case drops from 4920 to 2966 ms. The Close half is the longer one, at about 0.75–0.8× the original's local time. Scaled to CI's 3921 ms median, it should land near 3 s, about 1.8–2 s under the timeout instead of about 1.1 s. That is an estimate. This PR's own flake-check lane will repeat the edited file and give the real CI numbers.

Unrelated local failures. Each 15-file run had one failure in src/app/entity-navigation.test.tsx, which this PR does not touch, and both happened under the load above:

• Before: "opens Inbox for a malformed launch link when the catalog settles after the client" failed a findByRole wait.
• After: "normalizes old Bestie links … retaining { …(2) } scope" timed out at 5608 ms.

No browser journeys were added or removed.

🤖 Generated with Claude Code

[kazemek21]

🤖 Kyle reviewed this PR and the Nexus review-tracker notes and approved with no comments.

Posted by Codex (gpt-6.1-sol) on behalf of kylez.

[wesbillman]

🤖 Star Lord’s automated source review via Wes’s account.

No actionable findings in the test split after independent Rocket, Mantis and Groot reviews and Star Lord integration. The Back/Forward assertions are preserved, and the separate Close test keeps two Pair mobile visits separated by Appearance, so its final Back assertion still detects retirement of only the current visit. Each case uses fresh services and the existing awaited teardown; production code and the 5000 ms timeout are unchanged.

Observed hosted evidence at this head: CI run 42229295012 succeeded. Both Vitest flake lanes ran the full 18-case navigation file five times each, with no failures. Back/Forward measured 1781–2060 ms and Close 2761–3153 ms; the regular shard passed 18/18 in 23.884 s. This supports per-case timeout headroom, not reduced aggregate execution cost or proof that all timing failures are eliminated.

Reviewed head 8a77a99f4394e9f2de605ecdd61d0a3651b20805, base 067db608fbc624e56161ccf3e4ff1c0f11c30baf (PR merge-base 3bcf1c4bb1111d924a84d9b710c1da2db80ccf0b). Source-only review; no local tests or app execution. Author-reported baseline, shuffle and mutation experiments were not independently rerun.