fix(desktop): refresh @ search when channel membership changes

[loganj]

🤖

Summary

Draft 2 of 8, based on #7190.

Refresh agent discovery on governing membership events, cancel stale requests, coalesce updates and clear queued work across communities. Mechanical restack only: behavior and immediate-parent source diff are unchanged.

Stack and exact revision

#7190 → #7191 → #7192 → #7196 → #7323 → #7197 → #7239 → #7240

Head b411da58017d7d06304073e46dc1b6dc4cd7e3fd; immediate base bb8cca21e13099bc6206b18c62272b72928574ca. Fixed root 802ba6170e57ac71d1ecfd6cf6fb12f02ecf910c — no main chase. Existing PR identities preserved; new editor predecessor is #7323. Related #7124/#7128.

Measured review size

Production +86/−0; tests +140/−0; docs +0/−0. Total +226/−0, net +226, churn 226 against the immediate parent.

Validation

Amended PR1 prefix: 14 focused browser passes. Corrected editor prefix 6b78dec1: 11/11 focused native chat/forum/caret + DM-loading browser passes, with fresh TypeScript/E2E build and strict-port cwd/HTTP asset proof. The existing later-slice one-time DM fixture was moved to the editor prefix; it checks no premature signing, empty post-send draft and exact final recipient, not persistent @alice . The unchanged editor production/plugin tree retains 50/50 plugin passes.

Chooser prefix: 62/62 hook tests. Integrated final tree: 6077/6077 package tests + 156/156 browser tests, zero skips/flakes, TypeScript/E2E build, production build, desktop check and size gate passed. PR5/6/7 whole-tree hashes exactly equal their previously validated 6372d0e1/5e10b5db/36fdadbb trees after replay; this is evidence reuse by exact content equality, not a claim that old CI ran on new commit IDs.

Exact-head CI run: success, attempt 1 (final observation 2026-09-04T12:15:39.457156+00:00). DCO Check: success. Conflict check: MERGEABLE. Submitted reviews: 0; review threads: 0. Draft status retained; this is not review approval.

All-eight immediate-parent review churn remains 5529 (old seven: 5544). Against the old integrated tip, tests only +19/−34, net −15; production/docs byte-identical. These are distinct baselines. No independent review approval, merge, native WebKit or live-relay acceptance claimed. Full unrelated local just ci not run.

Known inherited failures and external gate

All eight PRs now have successful current CI runs on unchanged heads (historical cancelled runs are not current failures). #7240's attempt3 passed: smoke shard3 job101031131396 and dependent Desktop job101036797640 both succeeded. Only the failed shard and dependent aggregate were rerun; prior successful jobs were carried forward. Two of the owner's three authorized reruns used (attempt1 was the original run); stopped on success, with no attempt4. The pass condition for deferring the hover-fixture repair is met: repair DEFERRED, not fixed or flake-free. The fixture samples channel color before its 100 ms transition settles (attempt1 observed alpha .030/.016/.027), then compares the settled profile (.040) to that frozen intermediate value. The fixture, sidebar components and common 4% hover token are byte-identical to fixed root 802ba617. Exact artifact9935770782 contains three failures and retry1 trace. Deferred repair owner: @block/buzz-oss-team. Follow-up action: await channel-hover settlement before sampling and retain the actual final-color equality assertion, then validate the correction. The owner permitted deferral after a passing rerun; this known race is not repaired. This is source/trace attribution, not a claimed third baseline runtime reproduction; no product styling change or assertion waiver was made here.

Earlier independent failures also remain material evidence: #7190 attempt1 empty-edit cancellation accepted a stale channel choice inside the inherited 120 ms deletion debounce; superseded #7240 DM retry was intercepted by a hover-paused error toast. At fixed root 802ba617, the two ordinary scenarios passed once; separate controlled probes with the observed preconditions reproduced both failures. Baseline TypeScript failed on its existing unused import; only the unchanged-root Vite E2E diagnostic build was used. Owner @block/buzz-oss-team should separately invalidate/revalidate channel queries synchronously on deletion and make the DM retry fixture move away and await toast dismissal, retaining failure/draft assertions. None of these separate repairs was folded into the mention split.

Known recovered CI flakes include original #7190 mention-recipients.spec.ts:279 and #7240 attempt1 messaging.spec.ts:2437 (root-thread-link witness passed on retry2). #7240 attempt2 shard3 reports 323 passed, 1 failed, 5 flaky, 1 skipped: messaging.spec.ts:2437, :4306, :4487, navigation.spec.ts:153, and onboarding-agent-defaults.spec.ts:908 each recovered on retry1. The latter four are newly observed recovered failures here, not diagnosed or waived; detailed assertions are preserved in OUTBOX/MENTION_RERUN_D02EE206/FAILURE_EXCERPT.log. navigation.spec.ts:131 was skipped. Attempt2 hover samples were .035/.016/.016 versus the settled .040 on all three executions (artifact9937283288); this was the same terminal blocker in attempt2, not a new terminal failure. Attempt3 shard3 reports 327 passed, 0 failed, 2 flaky, 1 skipped: the hover test (message-feedback-snapshots.spec.ts:97, assertion108) initially sampled .016 versus settled .040, and messaging.spec.ts:4306 initially duplicated the reply prefix; both recovered on built-in retry1. navigation.spec.ts:131 remained skipped. Exact logs and artifact9938303997 preserve these recovered failures. Green CI does not erase either attempt2's five flakes or the older channel-query/toast diagnosis. The zero skips/flakes claim above applies only to the 156 focused local browser tests, not the full CI smoke suite. Full source/trace/probe diagnosis: OUTBOX/MENTION_CONTINUE_B78AD8BA/FAILURE_DIAGNOSIS.md and OUTBOX/MENTION_DELIVERY_63C2264C/NEW_FAILURE_DIAGNOSIS.md in the originating workspace. The previously failed external CI gate is now green. Ordinary maintainer review/acceptance remains outstanding; no merge is authorized by this report.

Draft for ordinary review. Origin: Buzz channel 8dcd7afb-0193-4a3e-8ae9-72838de89a1d.