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/spk.js

@@ -7,8 +7,20 @@ import readline from 'node:readline';
 
 const BIN = '/opt/callrec/spk_embed';
 const DIR = '/opt/sherpa/sherpa-onnx-v1.13.3-linux-aarch64-shared-cpu';
-const EMB = process.env.DIAR_EMB || '/opt/sherpa/nemo_titanet_small.onnx';
+// Identity/clustering model. Benchmarked 2026-07-27 on our own conference audio
+// (#74, 1181 mean-centred windows, same-turn vs cross-turn cosine):
+//   TitaNet-large  d' 3.75 / 197 s   CAM++ (3D-Speaker) d' 3.72 / 82 s
+//   TitaNet-small  d' 3.54 /  57 s   wespeaker VoxCeleb resnet34 d' 2.43 / 185 s
+// Fed through the full v2 pipeline the three good models ended up within one
+// speaker of each other (21/22/21 on a 23-person weekly), so TitaNet-small stays
+// the default: it is the fastest AND keeps every already-learned voiceprint
+// valid. SPK_MODEL switches models — prints are then stored per model (see
+// db.upsertVoiceprint), because cosine across models is meaningless.
+const EMB = process.env.SPK_MODEL || '/opt/sherpa/nemo_titanet_small.onnx';
 export const THRESHOLD = +(process.env.SPK_THRESHOLD || 0.5);
+// Voiceprints are only comparable within the SAME embedding model — stored with
+// every print so a model swap can't silently match against stale vectors.
+export function modelId() { return EMB.split('/').pop().replace(/\.onnx$/, ''); }
 
 let proc = null, ready = null, rl = null, pending = null, chain = Promise.resolve();
 
@@ -56,4 +68,4 @@ export function bestMatch(emb, voiceprints, threshold = THRESHOLD) {
   return (best && bs >= threshold) ? { id: best.id, name: best.name, score: bs } : null;
 }
 
-export function status() { return { model: EMB, present: fs.existsSync(EMB), running: !!proc, threshold: THRESHOLD }; }
+export function status() { return { model: EMB, modelId: modelId(), present: fs.existsSync(EMB), running: !!proc, threshold: THRESHOLD }; }