Diarization v2: Fenster-Clustering mit gelernter Schwelle + Live-Namensmarker

#74 (22 Teilnehmer) ergab automatisch 44 Sprecher, dieselbe Stimme mehrfach.
Ursache gemessen: ein Zentroid pro sherpa-Cluster aus max. 3 Schnitten, fester
Kosinus auf rohen Embeddings (Kreuz-Turn-Kosinus ~0,4 = keine Trennschaerfe).

v2: sherpa liefert nur Turn-Grenzen, Identitaet aus 3-s-Fenster-Embeddings,
Mittelwert-Zentrierung, Average-Linkage-AHC, Weak-Fold, Schwelle pro Aufnahme
per 2-Gauss-EM gelernt. Validiert: #74 44->20, #70 ->4, #68 ->2.

Live-Namensmarker: Namen waehrend der Aufnahme tippen; Zeitstempel benennt
spaeter den Cluster und dient als must-/cannot-link-Zwangsbedingung.

app/recorder.js

@@ -158,13 +158,35 @@ function wavDurationSec(wavPath) {
   try { return Math.max(0, (fs.statSync(wavPath).size - 44) / 32000); } catch { return 0; }
 }
 
-let current = null; // { proc, wavPath, startTs, source, title }
+let current = null; // { proc, wavPath, startTs, source, title, markers }
 
 export function status() {
-  return current ? { recording: true, startedAt: current.startTs, source: current.source } : { recording: false };
+  return current ? { recording: true, startedAt: current.startTs, source: current.source,
+                     markers: current.markers } : { recording: false };
 }
 export function isRecording() { return !!current; }
 
+// --- live speaker markers ---------------------------------------------------
+// While recording, the user types the name of whoever is talking. We stamp the
+// offset from recording start; after diarization the name is attached to the
+// acoustic cluster that was speaking around that offset (see diar.resolveMarkers).
+// This is the only reliable way to name 20+ participants of a conference call.
+export function addMarker(name) {
+  if (!current) throw new Error('not recording');
+  const n = String(name || '').trim();
+  if (!n) throw new Error('Name fehlt');
+  const m = { t_ms: Date.now() - current.startTs, name: n.slice(0, 128) };
+  current.markers.push(m);
+  return { ...m, index: current.markers.length - 1, markers: current.markers };
+}
+export function listMarkers() { return current ? current.markers : []; }
+export function removeMarker(index) {
+  if (!current) throw new Error('not recording');
+  const i = +index;
+  if (i >= 0 && i < current.markers.length) current.markers.splice(i, 1);
+  return current.markers;
+}
+
 // List real capture devices (one entry per recording device). A "source" in
 // PulseAudio/PipeWire includes monitor sources (the loopback of every OUTPUT)
 // and both A2DP/HFP variants of a BT card — confusing for the user. We only
@@ -263,7 +285,7 @@ export async function startRecording({ source, title, codec, numSpeakers } = {})
   // in the list (mSBC 16 kHz vs CVSD 8 kHz).
   const codecTag = hfp && hfp.profile ? (/cvsd/i.test(hfp.profile) ? ' [CVSD 8k]' : ' [mSBC 16k]') : '';
   current = {
-    proc, wavPath, startTs: ts, source: source || 'default', numSpeakers,
+    proc, wavPath, startTs: ts, source: source || 'default', numSpeakers, markers: [],
     title: (title || `Aufnahme ${new Date(ts).toLocaleString('de-DE')}`) + codecTag, err: () => err
   };
   return { startedAt: ts };
@@ -328,13 +350,14 @@ export async function stopAndSave() {
   catch { fs.copyFileSync(c.wavPath, path.join(REC_DIR, wavFile)); try { fs.unlinkSync(c.wavPath); } catch {} }
 
   return { mp3_file: `${base}.mp3`, wav_file: wavFile, duration_ms: dur,
-           language: 'de', title: c.title, device: c.source, numSpeakers: c.numSpeakers };
+           language: 'de', title: c.title, device: c.source, numSpeakers: c.numSpeakers,
+           markers: c.markers };
 }
 
 // Transcribe a saved wav (in REC_DIR) with whisper (thermal-guarded) + diarize.
 // Returns { transcript, words }. Idempotent — safe to re-run for recovery of a
 // recording whose transcription was interrupted (crash) or failed.
-export async function transcribeWav(wavFile, numSpeakers) {
+export async function transcribeWav(wavFile, numSpeakers, opts = {}) {
   const wavPath = path.isAbsolute(wavFile) ? wavFile : path.join(REC_DIR, wavFile);
   if (!fs.existsSync(wavPath)) throw new Error('wav-Datei fehlt: ' + wavFile);
   const outBase = `/tmp/${path.basename(wavPath, '.wav')}`;
@@ -389,7 +412,7 @@ export async function transcribeWav(wavFile, numSpeakers) {
     // Acoustic diarization under the A76 clock cap (both paths). Non-fatal.
     await setA76Max(A76_CAP);
     try {
-      const segs = await diar.diarize(wavPath, numSpeakers);
+      const segs = await diar.diarize(wavPath, numSpeakers, { markers: opts.markers || [] });
       diar.assignSpeakers(words, segs);
     } catch (e) { console.error('diarize failed:', e.message); }
     try { fs.unlinkSync(`${outBase}.json`); } catch {}