sync channel sort preferences across clients

[tellaho]

Category: improvement
User Impact: Users' per-group channel sort choices now follow them across desktop clients on the same relay.
Problem: The channel sort controls from #1505 persisted only in the desktop local cache, unlike nearby sidebar preferences such as sections, stars, and mutes that sync through the relay. That meant another desktop client could show the same channel groups but lose the user's chosen Recent / A–Z ordering.
Solution: This adds encrypted NIP-78 relay sync for the existing per-group sort map using the same local-cache, startup fetch, live subscribe, reconnect, and debounced publish pattern as channel sections. The UI and default behavior stay unchanged; localStorage remains the instant/offline cache while the relay blob becomes the cross-client source of truth.

File changes

desktop/src/features/sidebar/lib/channelSortSync.ts
Adds a sync manager for channel sort preferences using kind 30078, d-tag channel-sort, encrypted-to-self content, debounced whole-blob LWW writes, and workspace-switch destroy guards.

desktop/src/features/sidebar/lib/useChannelSortPreference.ts
Wires the existing sort preference hook into startup remote fetch, live subscription updates, reconnect refresh, pending publish flush, and seed-publishing of non-default local preferences when no remote blob exists.

desktop/src/features/sidebar/lib/channelSortSync.test.mjs
Adds regression coverage for destroy/cancel behavior so pending or in-flight publishes cannot leak preferences to the wrong relay during workspace switches.

desktop/src/shared/constants/kinds.ts
Documents channel-sort alongside the other NIP-78 d-tags and adds a named KIND_CHANNEL_SORT constant for the shared 30078 kind.

Reproduction Steps

1. Open the desktop app with PR feat(channel list): add persistent channel sort toggle #1505's sidebar sort controls available.
2. Change a sidebar group's sort mode from A–Z to Recent.
3. Confirm the group reorders immediately and the local cache still preserves the choice after refresh.
4. Open another desktop client or window with the same identity and relay.
5. Confirm the same group's sort mode is fetched from the relay and applied without changing other groups.
6. Switch relays/workspaces while a sort publish is pending and confirm the pending write does not publish to the wrong relay.

Validation

• tsc — clean
• biome — clean
• full desktop node suite — 1660/1660 pass, including 4 new sync regression tests
• lefthook commit hooks — passed

[wesbillman]

LGTM — approve. CI fully green (all checks pass/skip, incl. both E2E Integration shards + relay e2e). This is a disciplined, near line-for-line mirror of the existing channelSectionsSync pattern, which is exactly the right move here.

What's good

• Faithful pattern replication: manager class, NIP-44 encrypt-to-self, kind 30078 + d-tag, debounced whole-blob LWW, destroyed flag with both abort guards (post-fetch and pre-publish), monotonic created_at = max(now, lastRemote+1) — all consistent with channelSectionsSync.ts on main. No invented conventions.
• The workspace-switch regression tests are real: the in-flight doPublish abort test holds fetchEvents open with a manual-release promise, fires the timer, destroys, then releases — that exercises the actual cross-relay leak vector, not a sequential approximation.
• Seed-publish adaptation is correct: Object.keys(local.groups).length > 0 is the right non-default check for a map-shaped store (sections uses sections.length > 0).
• KIND_CHANNEL_SORT doc-comment update keeps the d-tag inventory in kinds.ts accurate.

Minor notes (none blocking)

1. Stacked PR — merge order matters. Base is tho/channel-sort (#1505, still open); don't merge this one first.
2. Whole-blob LWW is a shade lossy for this shape. The sort store is a per-group keyed map (Record<groupKey, mode>) — structurally identical to mutes/stars, which use per-key mergeStores. Two clients changing different groups within the same debounce window will drop one change under whole-blob LWW. The PR body's low-frequency argument is convincing for sort toggles — but if a "my sort reset" ghost report ever shows up, per-key merge is the known fix.
3. Fourth copy of the same manager. sections, mutes, stars, and now sort each carry a near-identical ~225-line sync manager + test scaffolding. Matching the convention rather than refactoring mid-flight is the right call for this PR, but a generic NIP-78 sync-manager factory is now worth a queued follow-up — the next preference blob shouldn't be copy #5.

Reviewed by Pinky (Buzz agent) on behalf of Wes.