perf(story-mode): parse transcript tails per line and yield to the event loop

The first read decoded the whole tail as one string and parsed it in one
synchronous pass. Lines are now decoded one at a time straight from the buffer
and the loop yields every 1 MB. Measured on the largest local transcripts
(110, 64, 49 MB): a cold read takes 39-50 ms instead of 100-123 ms, and the
longest event-loop stall is 8-10 ms.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

src/story-reader.ts

@@ -87,7 +87,16 @@ async function readRange(path: string, start: number, end: number): Promise<Buff
 
 const entryChars = (e: StoryEntry): number => JSON.stringify(e).length;
 
-function ingest(entry: CacheEntry, buf: Buffer, startsMidLine: boolean): void {
+/** Bytes of lines parsed between yields to the event loop: a big first read must not stall the server. */
+const YIELD_EVERY_BYTES = 1024 * 1024;
+const yieldToLoop = () => new Promise<void>((resolve) => setImmediate(resolve));
+
+/**
+ * Parses the complete lines in `buf` into the entry. Lines are decoded one at a time (never the whole
+ * buffer as one string: a 64 MB tail would be ~128 MB of UTF-16), and the loop yields every
+ * `YIELD_EVERY_BYTES`, so even the largest read keeps the event loop moving.
+ */
+async function ingest(entry: CacheEntry, buf: Buffer, startsMidLine: boolean): Promise<void> {
   let begin = 0;
   if (startsMidLine) {
     const firstNewline = buf.indexOf(0x0a);
@@ -100,19 +109,29 @@ function ingest(entry: CacheEntry, buf: Buffer, startsMidLine: boolean): void {
     entry.offset += begin;
     return;
   }
-  const text = buf.subarray(begin, lastNewline).toString('utf8');
   entry.offset += lastNewline + 1;
-  for (const line of text.split('\n')) {
-    if (!line.trim()) continue;
-    let parsed: unknown;
-    try {
-      parsed = JSON.parse(line);
-    } catch {
-      continue; // a malformed line is skipped, never fatal
+  let pos = begin;
+  let sinceYield = 0;
+  while (pos < lastNewline) {
+    let nl = buf.indexOf(0x0a, pos);
+    if (nl < 0 || nl > lastNewline) nl = lastNewline;
+    if (nl - pos > 1) {
+      let parsed: unknown = null;
+      try {
+        parsed = JSON.parse(buf.toString('utf8', pos, nl));
+      } catch {
+        parsed = null; // a malformed line is skipped, never fatal
+      }
+      for (const reduced of entry.reducer.reduce(parsed)) {
+        entry.entries.push(reduced);
+        entry.chars += entryChars(reduced);
+      }
     }
-    for (const reduced of entry.reducer.reduce(parsed)) {
-      entry.entries.push(reduced);
-      entry.chars += entryChars(reduced);
+    sinceYield += nl - pos + 1;
+    pos = nl + 1;
+    if (sinceYield >= YIELD_EVERY_BYTES) {
+      sinceYield = 0;
+      await yieldToLoop();
     }
   }
   while (entry.entries.length > MAX_ENTRIES || (entry.chars > MAX_CHARS && entry.entries.length > 1)) {
@@ -139,7 +158,7 @@ async function load(path: string, sidechain: boolean): Promise<StoryEntriesResul
   ) {
     // The file only grew: read what was appended since the last complete line.
     const buf = await readRange(path, cached.offset, info.size);
-    ingest(cached, buf, false);
+    await ingest(cached, buf, false);
     cached.size = info.size;
     cached.mtimeMs = info.mtimeMs;
     cache.set(key, cached);
@@ -161,7 +180,7 @@ async function load(path: string, sidechain: boolean): Promise<StoryEntriesResul
       reducer: createStoryReducer({ sidechain }),
       partial: start > 0,
     };
-    ingest(fresh, await readRange(path, start, info.size), start > 0);
+    await ingest(fresh, await readRange(path, start, info.size), start > 0);
     if (start === 0 || hasTurnStart(fresh.entries)) break;
   }
   const done = fresh as CacheEntry;