Namen nachtraeglich in fertige Zusammenfassung uebernehmen

speaker_labels speichert das Label jedes Clusters im LLM-Text; relabelAnalysis()
ersetzt beim Umbenennen wortgenau in summary/todos/transcript_clean und baut die
Teilnehmerliste neu. #74: keine 'Sprecher N' mehr.

app/db.js

@@ -15,6 +15,7 @@ export async function ensureSchema() {
     ADD COLUMN IF NOT EXISTS transcript_clean TEXT NULL,
     ADD COLUMN IF NOT EXISTS wav_file VARCHAR(255) NULL,
     ADD COLUMN IF NOT EXISTS speaker_names TEXT NULL,
+    ADD COLUMN IF NOT EXISTS speaker_labels TEXT NULL,
     ADD COLUMN IF NOT EXISTS analysis_status VARCHAR(16) NOT NULL DEFAULT 'pending',
     ADD COLUMN IF NOT EXISTS analysis_started_at BIGINT NULL,
     ADD COLUMN IF NOT EXISTS analysis_ms INT NULL`);
@@ -187,6 +188,17 @@ export async function setSpeakerName(id, speaker, name) {
   if (name) m[speaker] = name; else delete m[speaker];   // empty name clears back to "Sprecher N"
   await pool.execute('UPDATE recordings SET speaker_names=? WHERE id=?', [JSON.stringify(m), id]);
 }
+// The label EVERY speaker carried in the text the LLM produced ({speaker_07:
+// "Sprecher 5"}). Stored with the analysis so a later rename can swap the name
+// into the finished summary/todos instead of re-running the 20-minute analysis.
+export async function setSpeakerLabels(id, labels) {
+  await pool.execute('UPDATE recordings SET speaker_labels=? WHERE id=?', [JSON.stringify(labels || {}), id]);
+}
+export async function getSpeakerLabels(id) {
+  const [rows] = await pool.query('SELECT speaker_labels FROM recordings WHERE id=?', [id]);
+  try { return JSON.parse((rows[0] && rows[0].speaker_labels) || '{}') || {}; } catch { return {}; }
+}
+
 // Seed names from acoustic identification WITHOUT clobbering existing (manual) ones.
 export async function mergeSpeakerNames(id, map) {
   if (!map || !Object.keys(map).length) return;

app/server.js

@@ -35,15 +35,69 @@ function applyGlossaryText(text, m) {
 // Build the LLM input from stored words. If acoustic diarization tagged speakers,
 // produce a labelled dialogue ("Sprecher 1: …") so the LLM keeps real speakers
 // instead of guessing; otherwise the plain transcript.
+// Also returns `labels` = the exact label every speaker carries in the text the
+// LLM sees ({speaker_07: "Sprecher 5"}); it is stored with the analysis so a
+// rename afterwards can be patched into the finished summary (relabelAnalysis).
 function buildBasis(rec, nameMap = {}) {
   const words = rec.words || [];
   const hasSpk = words.some(w => w.speaker);
-  if (!hasSpk) return { text: rec.transcript || '', labeled: false };
+  if (!hasSpk) return { text: rec.transcript || '', labeled: false, labels: {} };
   const turns = diar.toTurns(words);
-  const map = {}; let n = 0;
-  const label = sp => nameMap[sp] || (map[sp] || (map[sp] = `Sprecher ${++n}`));
+  const labels = {}; let n = 0;
+  const label = sp => labels[sp] || (labels[sp] = nameMap[sp] || `Sprecher ${++n}`);
   const text = turns.map(t => `${label(t.speaker)}: ${t.text}`).join('\n');
-  return { text, labeled: true };
+  return { text, labeled: true, labels };
+}
+
+// Swap speaker labels for real names in the ALREADY FINISHED analysis text.
+// Renaming a speaker after the analysis used to leave "Sprecher 5" all over the
+// summary/todos; re-running the LLM costs ~20 min on this board, while the label
+// each speaker had is known exactly (recordings.speaker_labels). Whole-word
+// replacement in summary, todos, transcript_clean and participants; the stored
+// label is then updated to the new name so a second rename works too.
+async function relabelAnalysis(id) {
+  const rec = await db.getRecording(id);
+  if (!rec) return { changed: 0 };
+  const names = rec.speaker_names || {};
+  const labels = await db.getSpeakerLabels(id);
+  if (!Object.keys(names).length || !Object.keys(labels).length) return { changed: 0 };
+  // A label that TWO speakers carried (the voice-ID once matched "Fabio" onto two
+  // clusters) is ambiguous in the text — leave those alone instead of renaming
+  // both occurrences to one person.
+  const labelCount = {};
+  for (const l of Object.values(labels)) labelCount[l] = (labelCount[l] || 0) + 1;
+  const pairs = Object.entries(names)
+    .filter(([sp, name]) => name && labels[sp] && labels[sp] !== name && labelCount[labels[sp]] === 1)
+    .map(([sp, name]) => ({ sp, from: labels[sp], to: name }))
+    // longest label first: "Sprecher 1" must not eat "Sprecher 12"
+    .sort((a, b) => b.from.length - a.from.length);
+  const esc = s => s.replace(/[.*+?^${}()|[\]\\]/g, '\\$&');
+  const swap = (text) => {
+    if (!text) return text;
+    let out = String(text);
+    for (const p of pairs) out = out.replace(new RegExp(`(?<![\\p{L}\\p{N}])${esc(p.from)}(?![\\p{L}\\p{N}])`, 'gu'), p.to);
+    return out;
+  };
+  const todos = Array.isArray(rec.todos) ? rec.todos.map(t => swap(String(t))) : rec.todos;
+  // Participants are REBUILT (not string-swapped) from the diarized speakers in
+  // order of first appearance — a swap would leave the entries of speakers that
+  // were never mentioned in the text untouched and can duplicate names.
+  const spkOrder = [];
+  for (const w of (rec.words || [])) if (w.speaker && !spkOrder.includes(w.speaker)) spkOrder.push(w.speaker);
+  const participants = spkOrder.length
+    ? spkOrder.map((sp, i) => names[sp] || labels[sp] || `Sprecher ${i + 1}`)
+    : (Array.isArray(rec.participants) ? rec.participants.map(p => swap(String(p))) : rec.participants);
+  const partChanged = JSON.stringify(participants || null) !== JSON.stringify(rec.participants || null);
+  if (!pairs.length && !partChanged) return { changed: 0 };
+  await db.pool.execute(
+    'UPDATE recordings SET summary=?, todos=?, transcript_clean=?, participants=? WHERE id=?',
+    [swap(rec.summary), todos ? JSON.stringify(todos) : null, swap(rec.transcript_clean),
+     participants ? JSON.stringify(participants) : null, id]);
+  const next = { ...labels };
+  for (const p of pairs) next[p.sp] = p.to;
+  await db.setSpeakerLabels(id, next);
+  console.log(`relabel #${id}: ${pairs.length ? pairs.map(p => `${p.from}→${p.to}`).join(', ') : 'nur Teilnehmerliste'}`);
+  return { changed: pairs.length || (partChanged ? 1 : 0) };
 }
 
 const REC_DIR2 = path.join(path.dirname(fileURLToPath(import.meta.url)), 'recordings');
@@ -158,6 +212,7 @@ async function applyMarkers(id) {
   let learned = 0;
   for (const [sp, name] of Object.entries(nameOf)) if (await enrollSpeakerVoice(rec2, sp, name)) learned++;
   const named = Object.keys(nameOf).length;
+  await relabelAnalysis(id).catch(() => {});
   console.log(`markers on #${id}: ${resolved.length} marker(s) → ${named} named speaker(s), ${merged} merged, ${learned} voice(s) learned`);
   return { resolved, named, merged, learned };
 }
@@ -194,7 +249,10 @@ async function analyzeRecording(id) {
       for (const w of rec.words) { const nt = applyGlossaryText(w.word, gmap); if (nt !== w.word) { w.word = nt; changed = true; } }
       if (changed) { await db.replaceWords(id, rec.words); console.log(`glossary: applied soundalike swaps to #${id} words`); }
     }
-    const { text, labeled } = buildBasis(rec, names);
+    const { text, labeled, labels } = buildBasis(rec, names);
+    // Remember which label each speaker had in the LLM input, so renaming a
+    // speaker later can patch the finished text (relabelAnalysis).
+    await db.setSpeakerLabels(id, labels).catch(() => {});
     const a = await llm.analyze(text, { labeled, glossary });
     // Participants = the ACOUSTICALLY diarized speakers (named where known),
     // not the LLM's list — the model returned every name MENTIONED in the
@@ -202,7 +260,9 @@ async function analyzeRecording(id) {
     // LLM participants stay only as fallback when there is no diarization.
     const spkOrder = [];
     for (const w of (rec.words || [])) if (w.speaker && !spkOrder.includes(w.speaker)) spkOrder.push(w.speaker);
-    if (spkOrder.length) a.participants = spkOrder.map((sp, i) => names[sp] || `Sprecher ${i + 1}`);
+    // Same labels as in the dialogue the LLM read — otherwise the participant
+    // list numbers the speakers differently than the summary talks about them.
+    if (spkOrder.length) a.participants = spkOrder.map((sp, i) => names[sp] || labels[sp] || `Sprecher ${i + 1}`);
     // Belt-and-suspenders: also swap in the LLM outputs (cheap, catches any the
     // model reintroduced).
     for (const k of ['title', 'summary', 'transcriptClean']) if (a[k]) a[k] = applyGlossaryText(a[k], gmap);
@@ -486,6 +546,8 @@ app.post('/api/recordings/:id/speaker', wrap(async (req, res) => {
   const r = await db.getRecording(id);
   if (!r) return res.status(404).json({ error: 'not found' });
   await db.setSpeakerName(id, speaker, name);
+  // Patch the name into the already generated summary/todos/participants.
+  const relabel = await relabelAnalysis(id).catch(e => { console.error('relabel:', e.message); return { changed: 0 }; });
 
   let enrolled = false, learnMsg = null;
   if (enroll && name && r.wav_file) {
@@ -508,7 +570,8 @@ app.post('/api/recordings/:id/speaker', wrap(async (req, res) => {
     }
   }
   const r2 = await db.getRecording(id);
-  res.json({ ok: true, speaker_names: r2.speaker_names || {}, enrolled, learnMsg });
+  res.json({ ok: true, speaker_names: r2.speaker_names || {}, enrolled, learnMsg,
+             relabeled: relabel.changed });
 }));
 
 app.get(/^(?!\/api|\/media).*/, (req, res) => res.sendFile(path.join(__dirname, 'web/dist/index.html')));

app/web/src/main.js

@@ -810,7 +810,8 @@ async function openRecording(id) {
       btn.disabled = true; btn.textContent = '…';
       try {
         const res = await api.post('/api/recordings/' + id + '/speaker', { speaker: popSp, name, enroll });
-        toast(res.enrolled ? `„${name}" gespeichert + Stimme gelernt` : (res.learnMsg ? 'Gespeichert (' + res.learnMsg + ')' : 'Gespeichert'));
+        const base = res.enrolled ? `„${name}" gespeichert + Stimme gelernt` : (res.learnMsg ? 'Gespeichert (' + res.learnMsg + ')' : 'Gespeichert');
+        toast(res.relabeled ? base + ' · Name in Zusammenfassung übernommen' : base);
         openRecording(id);
       } catch (e) { toast(e.message, true); btn.disabled = false; btn.textContent = 'Speichern'; }
     };