v2.0.10 — snappier boot + non-blocking relay push on save

DB
- Six indices on the cards table (position+created_at, bucket_id,
  parent_id, updated_at, caldav_uid, relay_uid). The /api/cards
  ORDER BY position was a full-table scan + sort on every board
  reload; now it's an index walk.
- PRAGMA synchronous=NORMAL (safe under WAL) + temp_store=MEMORY.

Save path
- pushCardToRelayRead and deleteCardFromRelayRead now run AFTER the
  response is flushed via fastcgi_finish_request (falls back to
  ignore_user_abort + flush on non-FPM setups). The user's save no
  longer blocks on the relay round-trip.
- cURL timeouts dropped from 10s/30s to 5s/10s, plus a 2s
  connect-timeout. Worst-case on a dead relay goes from ~160s down
  to ~15s — and now that worst case is in the background anyway.

Client
- loadCards phase-2 batches bumped from 25 to 500 cards per fetch.
  For 300 cards that's 1 HTTP trip instead of 12 — spinner wall time
  drops from ~3s to ~0.25s.
- renderBoard() is now requestAnimationFrame-coalesced. The boot
  sequence calls renderBoard from cache paint, meta paint, body
  paint, and the cards.reloaded plugin hook; they used to each
  trigger an innerHTML swap + layout pass. Now they collapse into
  one actual render per frame. Callers that really need a sync
  render (rare) can use the new renderBoardSync().

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

public/js/app.js

@@ -108,11 +108,14 @@ async function loadCards() {
     updateSortUI();
     renderBoard();
 
-    // Phase 2: Load full cards in batches → replace spinners with content
-    const batchSize = 25;
-    for (let i = 0; i < meta.length; i += batchSize) {
-        const batch = meta.slice(i, i + batchSize);
-        const ids = batch.map(c => c.id).join(',');
+    // Phase 2: Load full card bodies. One fetch up to the ID cap — with
+    // the /api/cards(?ids=) endpoint now indexed on cards.position +
+    // created_at a single big batch is faster than many small ones, and
+    // shorter perceived spinner time matters more than memory. Fall back
+    // to smaller batches only if the URL would blow past ~2000 chars.
+    const IDS_MAX_PER_FETCH = 500;
+    for (let i = 0; i < meta.length; i += IDS_MAX_PER_FETCH) {
+        const ids = meta.slice(i, i + IDS_MAX_PER_FETCH).map(c => c.id).join(',');
         const full = await api('api/cards?ids=' + ids);
         if (!full) continue;
         for (const card of full) {

public/js/modules/board.js

@@ -493,8 +493,30 @@ export function renderCardHtml(c, inGroup = false, recent = []) {
 }
 
 // ── Render ──
+//
+// Public renderBoard() is coalesced via requestAnimationFrame: multiple
+// callers in the same tick (cache paint → meta paint → body paint →
+// plugin-hook re-render → ...) collapse into a single actual render on
+// the next frame. Dropped the "three innerHTML swaps per reload" cost
+// that showed up as the "Gekringel" during boot. Callers that need the
+// DOM in a specific state before doing something else can call
+// renderBoardSync() explicitly — but in practice every existing caller
+// just wants the board to reflect the latest state.
 
+let rafScheduled = 0;
 export function renderBoard() {
+    if (rafScheduled) return;
+    rafScheduled = requestAnimationFrame(() => {
+        rafScheduled = 0;
+        renderBoardSync();
+    });
+}
+export function renderBoardSync() {
+    if (rafScheduled) { cancelAnimationFrame(rafScheduled); rafScheduled = 0; }
+    renderBoardImpl();
+}
+
+function renderBoardImpl() {
     // Diff state.cards against the last seen ids so we can flag arrivals
     // as "unseen". The very first renderBoard runs with state.cards = []
     // (initial empty state, before the API has returned), so we wait