fix(chat): dedup parallel loadMessages/loadMoreMessages per channel

Live network trace via playwright showed every channel mount firing TWO
parallel GET /channels/:id/messages requests (AppLayout + MessageList
both call loadMessages on mount; neither can be removed because
MessageList is also rendered by VoiceChatPanel which doesn't call it),
and every fast-scroll burst firing N parallel loadMoreMessages calls
(handleScroll closure shares !isLoadingMore=true across the burst).

The hasMore-based guard only deduplicates calls AFTER the first
completes — it does not collapse parallel callers. On a NAT hairpin'd
LAN where some federated TCP connections hang, having N hung parallel
fetches keeps the pagination skeleton stuck while any one is still
waiting on the 30 s api-client timeout.

Fix: in-flight Promise dedup at the chatStore level. Module-level Maps
keyed on channelId; concurrent callers receive the same Promise.
Cleared in finally with self-check so a force-reload mid-flight isn't
dropped. force=true bypasses dedup (WS reconnect intent).

docs/systems/message-list.md

@@ -99,4 +99,5 @@ These items were considered and rejected for the 2026-04-25 work; they live here
 - 2026-04-27 — smooth-scroll intent + `scrollend` final pin: typed `smoothScrollIntentRef` (`'bottom' | 'message' | null`) with an 800ms deadline. `handleScroll` suppresses the at-bottom flip while a `'bottom'` intent is active and the user hasn't wheeled past the 5000px threshold; the gate stays open so Effects B/C re-pin to bottom as media loads mid-animation. Effect D fires a final defensive instant pin via the native `scrollend` event (Chrome 114+, Safari 18+) or a `setTimeout(800)` fallback. `'message'` intent (jump-to-message from search) does NOT suppress — the gate flips honestly so the user is left at the targeted message. Closes the "Jump to Present doesn't fully reach the bottom" residue reproducible on `nova.ddns.net` Orbit → general. UX call: smooth-scroll animation preserved, not replaced with an instant jump.
 - 2026-05-05 — skeleton-as-overlay fix: the initial-load skeleton was rendered as an early return that replaced the JSX containing `containerRef` / `contentRef`. On slow loads (mobile hotspot, throttled connection), `useDelayedLoading`'s 200ms threshold flipped `showInitialSkeleton=true` before messages arrived, unmounting the scroll container. When messages then arrived (`messages.length` 0→N), Effect A's rAF ran against `containerRef.current === null` and bailed at the null guard at the top of its callback. Effects B/C/D and the scrollend listener — all keyed on `[hasMessages, channelId]` — re-fired exactly once when `hasMessages` flipped false→true (which happened *while the skeleton was still up* due to `useDelayedLoading.minDisplay`'s 300ms enforcement), hit the same null guard, and never re-attached because no dep changed when the skeleton finally cleared. Net effect: chat opened scrolled to the top instead of the bottom; ResizeObserver never observed the content; scrollend never registered; saved-anchor restore was equally broken. Fix: render the skeleton as an `absolute inset-0 z-10 bg-surface-chat pointer-events-none` overlay alongside the (always-mounted) scroll container, so all refs stay live across the loading transition. Added the "ContainerRef invariant" section above to encode the constraint for future loading-UI additions. **Manual repro recipe** (regression check): DevTools → Network tab → throttle to "Slow 3G" → click into a channel that hasn't been opened this session → pre-fix: list lands scrolled to the top; post-fix: list lands at the bottom. Saved-anchor variant: scroll up in a channel, switch away, return under throttle — pre-fix: lands at top; post-fix: lands at the saved anchor.
 - 2026-04-27 — pagination cross-channel race fix: `handleScroll`'s load-more block (`scrollTop < 50 && hasMore`) leaked `isLoadingMore = true` and the `prevScrollHeight` value across channel switches that raced the async `loadMoreMessages` await. The `useState` slot is the same component instance across channel changes, so the new channel inherited the flag (phantom pagination skeleton) and the post-await `requestAnimationFrame` applied the outgoing channel's `prevScrollHeight` to the incoming channel's container DOM (wrong-position scroll). Fix: capture `channelId` into a `requestChannelId` local at the start of the load block, mirror the live `channelId` prop into `currentChannelIdRef` synchronously each render, and compare twice — once before scheduling the rAF and once *inside* the rAF callback (the ~16ms gap between scheduling and firing is enough time for a click to switch channels). Wrap the await in `try/finally` so `setIsLoadingMore(false)` always runs even on throw. Belt-and-suspenders: the channel-switch effect (Effect 3) also calls `setIsLoadingMore(false)` so a never-resolving await (network hang) cannot strand the flag on the new channel. The store's `currentChannelId` was rejected as the live source — it lags one render behind a URL-driven channel switch (set in an `AppLayout` effect that fires after `MessageList` renders with the new prop), which would let the guard mis-fire during that one-frame window. Request cancellation was deliberately deferred — the channelId guard already silently drops stale results, and `AbortController` plumbing through `chatStore.loadMoreMessages` is a larger refactor; network waste on an abandoned page is not a correctness issue.
+- 2026-05-05 — sticky pagination skeleton on federated channels (parallel loadMessages/loadMoreMessages dedup): the 2026-05-05 hook + clamp-scroll fix above did not cover the federated-LAN reproduction. Playwright trace against the live deployment showed every channel mount firing TWO parallel `GET /channels/:id/messages?limit=50` requests, and every fast-scroll burst firing N parallel `loadMoreMessages` requests. Two cooperating root causes: (a) `AppLayout.tsx:289` AND `MessageList.tsx:267` both call `loadMessages(channelId)` in their channel-mount effects — `MessageList`'s call is required for surfaces that don't have the AppLayout chrome (`VoiceChatPanel`), so neither can be removed unilaterally; (b) `handleScroll`'s load-more block sees `!isLoadingMore` as true in every scroll event of a fast burst because all events share one closure created before the first `setIsLoadingMore(true)` commits. The chatStore guard `if (!force && get().hasMore.has(channelId)) return;` only deduplicates calls *after* the first completes — it does not collapse parallel callers. On a NAT hairpin'd LAN where some federated TCP connections succeed and others hang, having N hung parallel fetches per channel mount keeps `chatStore.isLoading` and component-level `isLoadingMore` true while any one of them is still waiting on the 30 s api-client timeout, producing the visually-stuck pagination skeleton across channel switches. Fix: in-flight Promise dedup at the chatStore level. `loadMessages` and `loadMoreMessages` each consult a module-level `Map<channelId, Promise>`; concurrent callers receive the same Promise. The map entry is cleared in a `finally` block, with a `inFlightLoads.get(channelId) === promise` self-check so a force-reload (WS reconnect) scheduled mid-flight isn't dropped. `force=true` callers bypass dedup intentionally — WS reconnect needs a fresh fetch even if a stale one is still pending. Net effect: one federated fetch per channel mount, one per fast-scroll burst, predictable timeout window, skeleton clears within ≤ 30 s + minDisplay even on full hairpin failure.
 - 2026-05-05 — sticky pagination skeleton (clamp-scroll trigger + minDisplay deadline drift): two cooperating bugs produced a "ghost" pagination skeleton that stuck at the top of the chat across channel switches, sometimes until reload. (1) When the user switched from a tall channel to a shorter one, the browser clamped `scrollTop` and dispatched a synthetic scroll event after Effect 3 ran. `handleScroll` saw `scrollTop < 50 && hasMore && !isLoadingMore` — for any channel where `hasMore` defaulted to `true` (unvisited or LRU-evicted; see `MessageList.tsx` `hasMore` selector with the `?? true` fallback) the load-more block fired even though the user never scrolled. (2) `useDelayedLoading`'s threshold-timer callback unconditionally wrote `displayStartRef.current = Date.now()`, including when `show` was already true. With `isLoading` cycling true→false→true at periods near `threshold` (200 ms), each re-fire refreshed the deadline and the next `false` rescheduled `minDisplay` from a fresh start — extending the visible skeleton by up to one full `minDisplay` window per cycle. Together: the clamp event re-armed `isLoadingMore` against an empty-bail `loadMoreMessages` call, the hook over-extended the visible window, and rapid channel switches kept feeding the cycle. Fix: (a) `useDelayedLoading.ts` — only write `displayStartRef.current` on the false→true `show` transition, using `setShow`'s functional form to avoid stale closures. (b) `MessageList.tsx` — `suppressNextLoadMoreRef` is armed in Effect 3 on every channel change and consumed by the load-more block on the first scroll event after the switch. A 250 ms `setTimeout` fallback disarms the flag if no clamp event fires (new channel's content fit without clamping), so a real user scroll-to-top isn't permanently suppressed. Suppression is scoped to the load-more block only — `isAtBottomRef`/`isNearBottomRef` updates and visible-message tracking still run because the clamped scroll genuinely changes scroll position. Regression test: `useDelayedLoading.test.ts` (`REGRESSION: minDisplay deadline measured from FIRST show, not refreshed by re-fired threshold`).