perf(chat): instant channel re-entry — per-channel message cache + persistent WebSocket

[haowei2000]

Problem

Opening a channel was slow, and every switch paid the full cost again:

1. No client-side caching — ChannelView wiped messages and cold-fetched 50 rows on every entry, even re-entering a channel viewed seconds earlier.
2. A fresh WebSocket per channel switch — the realtime hook was keyed on channelId, so each switch tore down the socket and paid TCP/WS + auth + subscribe (2 RTT) again.
3. Duplicate first page — if the subscribe ack won the race against the initial history load, the catch-up ran with since_seq=0 and re-fetched a full page the load was already fetching.
4. Auto-scroll only fired on message-count growth, which a cache seed doesn't trigger.

The backend needed no changes — list_channel_messages is already keyset-paginated with batched hydration (3 queries/page) and the right indexes.

Changes (Discord/Telegram-style, frontend only)

• chatCache.ts (new) — per-channel in-memory LRU cache (messages + members, 30 channels). Re-entry renders the cached window instantly, then a since_seq catch-up merges anything missed. Seeds trim to the newest 100 rows (hasMore flips back on) and clear interrupted streaming flags.
• useChatRealtime.ts — the browser socket is now a module-level singleton shared across switches: switching sends unsubscribe/subscribe on the live connection. Late frames for a left channel are dropped by channel_id; an idle socket closes after 30s; logout closes it immediately. Public API and exports unchanged.
• ChannelView.tsx — warm path seeds from cache and catch-ups incrementally; the subscribe-ack race is fixed (catch-up defers until the real seq cursor exists); every async response is guarded so a reply landing after a channel switch is discarded instead of merged into the wrong channel's state/cache.
• MessageList.tsx — a channel switch jumps straight to the bottom of the new timeline.

Verification

Headless Chrome against the kind cluster (Vite → in-cluster gateway):

• Warm re-entry: rows visible in 134–152ms, network shows only a since_seq delta — 0 full-page refetches.
• 0 new WebSocket connections across A→B→A→B switches (the extra opens in dev logs are StrictMode double-mount noise + the by-design separate user-scope notification socket).
• After 5 rounds of subscription churn, a REST-posted message still arrives live over the reused socket with no refetch.
• vitest 83/83 (incl. new chatCache tests: LRU eviction, seed trimming, streaming-flag clearing); tsc clean except the known PdfViewer baseline.

🤖 Generated with Claude Code

[haowei2000]

Added the iOS counterpart in 520b1b2.

iOS already had two of the three web fixes by construction — a single app-wide socket with per-channel subscribe/unsubscribe, and a catchUp guarded against the since_seq=0 race — but ChatsHomeView rebuilt ChatModel on every switch (.id(channelId)), so each channel entry still cold-reloaded 50 messages + members.

Now AppModel.chatModels (new ChatModelStore, LRU of 12) retains one ChatModel per channel: re-entry renders the in-memory history instantly, scrolls to bottom, refreshes members in the background, and heals via the existing since_seq catch-up. Retained history is trimmed to the newest 100 rows on detach; the store is cleared on sign-out. Verified with xcodebuild (Debug, iOS Simulator) — build succeeds.

[haowei2000]

1b90be3 adds the two iOS platform-convention upgrades:

Offline-first persistence (SwiftData) — new MessageStore keeps each visited channel's newest 200 finalized messages on disk as full-DTO JSON (schema-stable), so a cold app launch renders history instantly and the network refresh lands on top. Handles the non-contiguous case (>1 page arrived while offline → keep only the fresh page so upward pagination stays gap-free), debounces writes, flushes on channel exit, deletes tombstoned messages from disk, and wipes on sign-out.

Memory-pressure eviction — ChatModelStore sheds every cached model except the open channel's on didReceiveMemoryWarning; safe now that evicted history reloads from disk instead of the network.

Notable bug caught by a macOS harness driving the real store: SwiftData does not apply fetchOffset windows reliably to pending (unsaved) objects — a pre-save trim fetch classified every fresh insert as overflow and deleted the batch it had just written. save() now persists first and trims by slicing in code. Harness covers round-trip/ordering, unique upsert, trim + hasMoreBefore, in-flight skipping, delete, channel isolation, removeAll, and cross-instance relaunch persistence; xcodebuild (Debug, iOS Simulator) succeeds.