fix(transcript): tlang en dernier recours + garde anti-blob

Le fallback &tlang= peut renvoyer la piste source en un seul bloc
de plusieurs Ko (throttle YouTube) : il passait devant les pistes
directes et son blob anglais etait mis en cache comme 'fr'.
tlang desormais etape 3 (apres directes signees), isBlobTranscript()
refuse les cues uniques > 2000 chars (fetch, yt-dlp et post-dedupe).
Tests 27/27.

server/transcript.mjs

@@ -352,6 +352,19 @@ export function isEmptyTimedTextBody(text) {
   return s.trim().length < 32;
 }
 
+/**
+ * True when parsed lines look like a degraded server-side translation blob
+ * rather than real cues: a single giant cue (several KB) instead of
+ * timestamped lines. YouTube serves these under throttle on `&tlang=`
+ * requests (source-language text, untranslated, in ONE event). Caching or
+ * labeling them as the target language would poison the UI + cache.
+ * A legitimate single-cue transcript (very short video) is always tiny.
+ * Pure and testable offline.
+ */
+export function isBlobTranscript(lines) {
+  if (!Array.isArray(lines) || lines.length !== 1) return false;
+  return String(lines[0]?.text || '').length > 2000;
+}
 /**
  * Merge fetch candidates, signed yt-dlp URLs first (they carry a fresh
  * `signature`/`expire` and are the only ones plain fetch can still read),