fix(desktop): derive unread state from NIP-RS + relay catch-up only

[tlongwell-block]

What

Make GUI unread state depend solely on (a) the NIP-RS read marker and (b) a per-channel relay catch-up query that asks "are there messages newer than this marker?" — no localStorage cache, no Channel.last_message_at dependency, no schema extension.

Why

useUnreadChannels had two halves: a NIP-RS read marker (which persists fine, locally and cross-device) and an in-memory latestByChannelRef Map that got built up only from the live subscription (since: now). On every restart the Map reset to empty, the comparison latest > readMarker always returned false, and badges silently disappeared until new live activity hit the channel.

The original fix in this branch persisted the Map to localStorage. Tyler's review surfaced a better model: the relay already knows what's in each channel, so just ask it. The NIP-RS spec is explicit that contexts carries read positions — channel "latest" is the wrong shape for that blob. And we don't need any client-side latest cache if we can query "newer than marker?" on demand.

How

• Delete the latestMessageStorage.ts module added earlier in this branch.
• Delete the (no-op) Channel.lastMessageAt opportunistic backfill effect and the read-state seeding effect from useUnreadChannels. Backend never populated last_message_at, both effects were dead branches.
• Add a catch-up effect: once NIP-RS is ready and channels are known, for each channel we haven't already caught up this session, REQ the channel's trigger-kind events with since: readMarker + 1 (strict-newer; client-side > readAt belt-and-suspenders), limit: 1000, in parallel. Take the max external-author event timestamp per channel; feed the existing predicate.
• Effect is keyed on a sorted channelIdsKey string so React Query refetches with identical contents don't churn. Optimistic claim of caughtUpChannelsRef prevents duplicate REQs from re-renders; cleanup releases claims so the next effect run retries; individual failed fetches release their own claim so transient relay errors don't suppress unread for the rest of the session.
• isCancelled guard around the post-settle mutation handles identity reset mid-flight.
• markChannelUnread falls back to latestByChannelRef.current.get(channelId) when Channel.lastMessageAt is null (always today), so right-click "mark unread" syncs across devices when we've observed any external message.
• Export CHANNEL_MESSAGE_EVENT_KINDS from shared/constants/kinds so useLiveChannelUpdates and the catch-up share one source of truth for "unread trigger kinds."

The live subscription remains since: now — unchanged.

Verification

• pnpm --dir desktop typecheck ✅
• pnpm --dir desktop check (biome + file-size) ✅
• All pre-commit and pre-push hooks pass.

Reviewers

Reviewed in #sprout-bugs by @max. Two rounds of review caught: (1) limit: 1 is unsafe when self-author filter discards the only returned event, (2) the original retry path left failed channels permanently "caught up" — both fixed. Approved at v3.

Closes

The "unread indicators disappear on restart" regression Tyler reported in #sprout-bugs.