perf(desktop): instant channel switching — non-blocking first paint, persisted snapshots

[wesbillman]

What

Client-side channel-switching performance bundle — the goal is Slack/Discord-tier switching: never show a skeleton for a channel you've seen before, and make first-ever opens paint in one round trip.

Three commits, reviewable independently:

1. Don't block first paint on the row-floor top-up

The cold-load queryFn awaited up to 3 extra serial 200-event relay pages (pageOlderMessagesUntilRowFloor) whenever the first 60-event window rendered fewer than 30 top-level rows — which reply-heavy channels almost always trip. That made a cold open up to 4 serialized relay round trips before the skeleton dropped. Now the first window commits immediately and the top-up runs in the background (same self-merging pattern as the aux backfill), guarded by a per-channel in-flight map so a scroll-up fetch shares the running pass instead of duplicating REQs.

2. Persisted per-channel message snapshots + longer gcTime

Channels revisited after gcTime expiry or an app restart went fully cold. Now the newest 80-event slice per channel is persisted to localStorage (keyed per relay + channel, LRU-capped at 20 channels/relay, GC'd on workspace removal, pending optimistic events excluded) and fed to placeholderData — a cold revisit paints instantly and revalidates behind the paint. selectTimelineLoadingState gains one narrow pre-settle rule: placeholder rows drop the skeleton; live-sub-seeded partial data still holds it, so the intro-flash and older-fetch-spinner traps stay covered. gcTime raised 5→60 min.

Ghost-resurrection guard: a snapshot row deleted or edited while the app was closed never reappears in a history fetch (the relay soft-deletes), so its tombstone/edit is only reachable by #e over that row's id. Cold snapshot loads therefore widen the aux backfill to the snapshot-merged timeline (mergeHistoryOverSnapshot), not just the fresh fetch window — otherwise the deleted message would repaint from the snapshot and the post-settle snapshot rewrite would persist it forever. Warm refetches keep the narrow fresh-window backfill.

3. e2e contract update

The scroll-history cold-load test asserted the old contract (skeleton held while the paced top-up ran); it now asserts the new one (rows paint well inside the paced top-up window, skeleton gone once they do).

Dropped from this PR

An earlier revision included sidebar hover prefetch. Cut after review: with persisted snapshots, every previously-visited channel already paints instantly, so hover prefetch only helps the first-ever visit — and it paid for that with an uncancellable fetch+top-up pipeline per hovered row and unbounded cross-channel fan-out. If first-visit latency matters later, it should return as its own PR with a hover-intent delay.

Result

• First-ever channel open: 1 relay round trip to paint (was up to 4).
• Any previously-seen channel (including across app restarts): instant paint from snapshot, revalidate behind it.

Testing

• pnpm typecheck, pnpm check, pnpm test — 1491 unit tests (16 new: messageSnapshot round-trip/eviction/bounds, loading-state placeholder rules, merged-timeline aux-backfill window).
• Playwright: touched spec (scroll-history.spec.ts) run locally — 12 passed; the 2 failures (preserves user scroll…, does not teleport upward…) reproduce identically at the unmodified merge-base (c88799a), so they're pre-existing, not from this branch.

Follow-ups (separate PRs, discussed in #faster-channel-switching-and-loading)

• S1: GIN index on event tags for #e backfill queries (relay migration).
• After fix(read-path): reach complete threads, dense-second timelines, and all people in the GUI #1418 lands: S2 concurrent bridge filter execution, C4 get_channels round-trip diet.