@gotcos/glasses-server 6.95.0 → 6.98.1

package/server/lib/mac-speaker-shadow-repair.ts

@@ -0,0 +1,307 @@
+/** Pure PREPARATION only. Hashes bind supplied bytes; they do not authenticate native
+ * execution, reconstructed PCM, model provenance, or a caller's complete export.
+ * A future trusted caller must verify those before any durable transaction and
+ * run this bounded but potentially expensive full-cohort computation off-thread
+ * under a wall-time/memory budget. This is not an HTTP/main-thread entry point. */
+import { createHash } from 'node:crypto'
+import { decide, exactOverlap, nativeTurns, type ShadowEmbeddingRow } from './speaker-shadow-decide.js'
+import { shadowEntries, shadowProductionSegments, type ShadowTimelineRow } from './speaker-shadow-timeline.js'
+import { hasMacSpeakerProtectedRows, isMacSpeakerHumanProtected } from './mac-speaker-protection.js'
+import { hashMacSpeakerDocument, projectMacSpeakerRepair, type MacSpeakerRepairPlan, type MacSpeakerWordBinding } from './mac-speaker-repair-projection.js'
+
+type Doc = Record<string, any>
+export interface MacShadowRepairSource {
+  originalJSON: string
+  currentJSON: string
+  ledgerJSONL: string
+  rawAudio: Array<{ rawChunkIndex: number; wavSHA256: string; pcmSHA256: string; sampleCount: number }>
+  timeline: { rows: ShadowTimelineRow[]; total: number; pcmSHA256: string }
+  cohort: { nativeJSON: string; embeddingsJSON: string; modelSHA256: string; cliSHA256: string }
+}
+export interface MacShadowRepairInput extends MacShadowRepairSource { sourceBindingSHA256: string }
+export interface MacShadowRepairPreview {
+  plan: MacSpeakerRepairPlan | null
+  provenance: { productionQualified: false; nativeAuthenticated: false; sourceBindingSHA256: string; originalSHA256: string; currentSHA256: string; ledgerSHA256: string }
+  protectedRawChunkIndices: number[]
+  withheld: Array<{ rawChunkIndex: number; reason: string }>
+  counts: { raw: number; spoken: number; embeddings: number; acousticAcceptedMerges: number; candidateChunks: number; assignedChunks: number; protectedChunks: number; withheldChunks: number; beforeLabels: number; afterLabels: number; beforeRootLabels: number; afterRootLabels: number }
+}
+const HASH = /^[a-f0-9]{64}$/
+const VOICE = /^Voice (?:[1-9]|[1-9]\d)$/
+const sha = (value: string) => createHash('sha256').update(value).digest('hex')
+const equal = (a: unknown, b: unknown) => JSON.stringify(a) === JSON.stringify(b)
+const fail = (message: string): never => { throw new Error(`Mac shadow repair: ${message}`) }
+/** Integrity helper, NOT a native execution/qualification certificate. */
+export function bindMacSpeakerShadowRepair(source: MacShadowRepairSource): string { return sha(JSON.stringify(source)) }
+function boundedJSON(text: string, max: number): any {
+  if (typeof text !== 'string' || Buffer.byteLength(text) > max) fail('JSON byte bound')
+  return JSON.parse(text)
+}
+function closed(doc: Doc, immutableView = false) {
+  if (!doc || !Array.isArray(doc.chunks) || !Array.isArray(doc.chunkEntries) || doc.finalizationState !== 'complete' || doc.sourceType !== 'mac_system_audio' || !/^mac_/.test(doc.sessionId) || doc.active || doc.isActive || doc.claimPending || doc.status === 'recording' || doc.mergedFrom || doc.sourceSessionIds || doc.isMerged) fail('closed original Mac capture required')
+  // Even a valid prior reader proof binds the complete speech document. A new
+  // label patch would invalidate it. This adapter cannot compose those receipts.
+  if (!immutableView && ['speakerRepair', 'speakerProfilePass'].some(key => Object.hasOwn(doc, key))) fail('prior repair/profile receipt requires proof composition')
+}
+const ANNOTATION_KEYS = new Set(['speaker', 'originalSpeaker', 'speakerCorrectionBatchId', 'speakerCorrectionId', 'manualSpeaker', 'speakerManuallyAssigned', 'confirmedByHuman', 'manuallyAssigned', 'manuallyCorrected', 'speakerOrigin', 'speakerSource', 'speakerPolicy', 'speakerProvenance', 'source'])
+/** Compare the complete source row recursively, allowing only speaker annotation
+ * changes. Annotation differences are handled by protection, never by copying
+ * labels back from ORIGINAL. Array presence, text, clocks and diagnostics remain
+ * exact on every support/competitor row, not just eventual assignments. */
+function sourceGeometry(value: any): any {
+  if (Array.isArray(value)) return value.map(sourceGeometry)
+  if (!value || typeof value !== 'object') return value
+  return Object.fromEntries(Object.keys(value).filter(key => !ANNOTATION_KEYS.has(key)).sort().map(key => [key, sourceGeometry(value[key])]))
+}
+function rows(chunk: Doc): Doc[] {
+  const out = [chunk]
+  for (const key of ['segments', 'finalSegments', 'words']) {
+    if (chunk[key] !== undefined && !Array.isArray(chunk[key])) fail('malformed label surface')
+    for (const row of chunk[key] ?? []) {
+      if (!row || typeof row !== 'object') fail('malformed label row')
+      if (key !== 'words' || 'speaker' in row) out.push(row)
+    }
+  }
+  return out
+}
+function mapping(doc: Doc): Array<{ raw: number; chunk: Doc; entry: Doc }> {
+  const entries = shadowEntries(doc)
+  if (entries.length > 6000 || entries.some((e, i) => e.chunkIndex !== i)) fail('complete ordered raw map required')
+  const spoken = entries.filter(e => e.chunk.text)
+  if (spoken.length !== doc.chunks.length) fail('compacted map count')
+  return spoken.map((e, i) => {
+    const chunk = doc.chunks[i]
+    if (!chunk || ['text', 'elapsed', 'audioSha256', 'audioDurationMs'].some(k => !equal(chunk[k], e.chunk[k]))) fail('ambiguous compacted raw join')
+    if (typeof chunk.text !== 'string' || !HASH.test(chunk.audioSha256)) fail('unbound spoken source')
+    return { raw: e.chunkIndex, chunk, entry: e.chunk }
+  })
+}
+function ledgerProtection(text: string, map: ReturnType<typeof mapping>, original: ReturnType<typeof mapping>): Set<number> {
+  if (Buffer.byteLength(text) > 8 * 1024 * 1024) fail('ledger byte bound')
+  const history = new Map<string, Doc[]>()
+  const replay = original.map(m => m.chunk.speaker)
+  const lines = text.split('\n').filter(x => x.trim())
+  if (lines.length > 20000) fail('ledger row bound')
+  for (const line of lines) {
+    const row = JSON.parse(line)
+    if (!row || typeof row.id !== 'string' || !row.id || !['intent', 'applied', 'failed', 'confirmed', 'confirmed-chunks', 'reverted'].includes(row.phase) || !Array.isArray(row.chunks) || row.chunks.some((i: number) => !Number.isInteger(i) || !map[i]) || new Set(row.chunks).size !== row.chunks.length) fail('ambiguous correction ledger')
+    const prior = history.get(row.id) ?? []
+    if (prior.at(-1)?.phase === 'intent' && !['applied', 'failed'].includes(row.phase)) fail('incomplete correction intent')
+    const intended = prior.at(-1)
+    const resolvedIntent = intended?.chunks.length ? intended.chunks : replay.flatMap((label, i) => label === intended?.from ? [i] : [])
+    // CorrectionRow documents an explicit empty intent as all rows carrying from;
+    // applied rows normally carry the resolved list. Missing chunks is not that
+    // explicit contract and was rejected above.
+    if (row.phase === 'intent' && prior.length || row.phase === 'applied' && (intended?.phase !== 'intent' || row.chunks.length && !equal(resolvedIntent, row.chunks) || intended?.from !== row.from || intended?.to !== row.to) || row.phase === 'failed' && intended?.phase !== 'intent' || row.phase === 'reverted' && !prior.some(r => r.phase === 'applied')) fail('ambiguous correction history')
+    if (row.phase === 'applied' && row.chunks.length) {
+      if (typeof row.from !== 'string' || typeof row.to !== 'string' || !row.to || row.chunks.some((i: number) => replay[i] !== row.from)) fail('correction history source label mismatch')
+      for (const i of row.chunks) replay[i] = row.to
+    }
+    if (row.phase === 'reverted') {
+      const applied = prior.find(e => e.phase === 'applied')!
+      if (!applied.chunks.length || applied.chunks.some((i: number) => replay[i] !== applied.to)) fail('ambiguous correction revert')
+      for (const i of applied.chunks) replay[i] = applied.from
+    }
+    history.set(row.id, [...prior, row])
+  }
+  const protectedRaw = new Set<number>()
+  for (const events of history.values()) {
+    if (events.at(-1)?.phase === 'intent') fail('incomplete correction intent')
+    // Retain protection even for a historical reverted decision. No machine write
+    // should silently reinterpret human history. Empty confirmations keep the
+    // existing offline CLI's conservative protect-all behavior.
+    for (const e of events.filter(e => ['applied', 'confirmed', 'confirmed-chunks'].includes(e.phase))) {
+      if (!e.chunks.length) for (const m of map) protectedRaw.add(m.raw)
+      else for (const i of e.chunks) protectedRaw.add(map[i].raw)
+    }
+  }
+  return protectedRaw
+}
+function batchBindings(original: Doc, current: Doc, entries: ReturnType<typeof shadowEntries>, samples: Record<string, number>, protection: Set<number>) {
+  const wordSpans: Array<{ binding: MacSpeakerWordBinding; raw: number[]; protectedRaw: number[]; protected: boolean }> = []
+  if (current.batchSegments !== undefined && !Array.isArray(current.batchSegments)) fail('batch schema')
+  if (!equal((original.batchSegments ?? []).map((b: Doc) => [b.startChunkIdx, b.endChunkIdx, b.speakerWords?.length]), (current.batchSegments ?? []).map((b: Doc) => [b.startChunkIdx, b.endChunkIdx, b.speakerWords?.length]))) fail('current batch topology drift')
+  for (const [si, batch] of (current.batchSegments ?? []).entries()) {
+    if (!Number.isInteger(batch.startChunkIdx) || !Number.isInteger(batch.endChunkIdx) || batch.startChunkIdx < 0 || batch.endChunkIdx < batch.startChunkIdx || batch.endChunkIdx >= entries.length || !Array.isArray(batch.speakerWords)) fail('batch raw clock range')
+    let offset = 0
+    const intervals = []
+    for (let raw = batch.startChunkIdx; raw <= batch.endChunkIdx; raw++) { const end = offset + samples[String(raw)] / 16000; intervals.push({ raw, start: offset, end }); offset = end }
+    for (const [wi, word] of batch.speakerWords.entries()) {
+      if (!word || typeof word.word !== 'string' || !Number.isFinite(word.start) || !Number.isFinite(word.end) || word.start < 0 || word.end < word.start || word.end > offset) fail('batch word clock')
+      const raw = intervals.filter(r => word.end === word.start ? word.start >= r.start && word.start < r.end : word.start < r.end && word.end > r.start).map(r => r.raw)
+      const oldWord = original.batchSegments[si].speakerWords[wi], changedWord = !equal(oldWord, word)
+      // Preserve even an anonymous, unmarked edit. An altered word clock protects
+      // both its original and current source intersections; never normalize it.
+      const priorRaw = changedWord ? intervals.filter(r => oldWord?.end === oldWord?.start ? oldWord.start >= r.start && oldWord.start < r.end : oldWord?.start < r.end && oldWord?.end > r.start).map(r => r.raw) : []
+      if (changedWord) for (const r of priorRaw) protection.add(r)
+      wordSpans.push({ binding: { segmentIndex: si, wordIndex: wi, word: word.word, start: word.start, end: word.end, from: word.speaker }, raw, protectedRaw: [...new Set([...raw, ...priorRaw])], protected: changedWord || isMacSpeakerHumanProtected(word) || !/^(?:Voice [1-9]\d?|Ext|Unknown|Unidentified(?: \d+)?|Speaker \d+)$/.test(word.speaker) })
+    }
+  }
+  for (const word of wordSpans) if (word.protected) for (const raw of word.protectedRaw) protection.add(raw)
+  return wordSpans
+}
+/** No remap/proposal input: recompute unchanged floor15/.8/1.5 decision from the
+ * entire bound export and production segments reconstructed from ORIGINAL rows. */
+function inspectSource(input: MacShadowRepairInput, immutableView: boolean) {
+  if (Object.keys(input).some(k => !['originalJSON', 'currentJSON', 'ledgerJSONL', 'rawAudio', 'timeline', 'cohort', 'sourceBindingSHA256'].includes(k))) fail('unsupported proposal/remap input')
+  for (const [value, maximum] of [[input.originalJSON, 32], [input.currentJSON, 32], [input.ledgerJSONL, 8], [input.cohort?.nativeJSON, 64], [input.cohort?.embeddingsJSON, 96]] as const) if (typeof value !== 'string' || Buffer.byteLength(value) > maximum * 1024 * 1024) fail('input byte bound')
+  if (!Array.isArray(input.rawAudio) || input.rawAudio.length > 6000 || !Array.isArray(input.timeline?.rows) || input.timeline.rows.length > 6000) fail('raw/timeline schema bound')
+  const { sourceBindingSHA256, ...source } = input
+  if (!HASH.test(sourceBindingSHA256) || bindMacSpeakerShadowRepair(source) !== sourceBindingSHA256) fail('full source binding drift')
+  const original = boundedJSON(input.originalJSON, 32 * 1024 * 1024), current = boundedJSON(input.currentJSON, 32 * 1024 * 1024)
+  closed(original, immutableView); closed(current, immutableView)
+  if (original.sessionId !== current.sessionId) fail('capture mismatch')
+  const oldMap = mapping(original), map = mapping(current)
+  if (oldMap.some(m => !equal(sourceGeometry(m.chunk), sourceGeometry(m.entry)))) fail('original root/raw source geometry disagreement')
+  if (['startTime', 'durationMs'].some(key => !equal(original[key], current[key]))) fail('current capture clock drift')
+  if (oldMap.length !== map.length || oldMap.some((m, i) => m.raw !== map[i].raw || ['text', 'elapsed', 'audioSha256', 'audioDurationMs'].some(k => !equal(m.chunk[k], map[i].chunk[k])))) fail('current speech source drift')
+  if (oldMap.some((m, i) => !equal(sourceGeometry(m.chunk), sourceGeometry(map[i].chunk)))) fail('current compacted source geometry drift')
+  const entries = shadowEntries(original), currentEntries = shadowEntries(current)
+  if (entries.length !== currentEntries.length || entries.some((e, i) => ['text', 'elapsed', 'audioSha256', 'audioDurationMs'].some(k => !equal(e.chunk[k], currentEntries[i].chunk[k])))) fail('current raw source drift')
+  for (const entry of [...entries, ...currentEntries]) rows(entry.chunk)
+  if (entries.some((e, i) => !equal(sourceGeometry(e.chunk), sourceGeometry(currentEntries[i].chunk)))) fail('current raw source geometry drift')
+  if (!Array.isArray(input.rawAudio) || input.rawAudio.length !== entries.length || !input.rawAudio.length) fail('all raw WAV bindings required')
+  const samples: Record<string, number> = {}
+  const placed = input.rawAudio.map((raw, i) => {
+    const chunk = entries[i].chunk
+    if (raw.rawChunkIndex !== i || !HASH.test(raw.wavSHA256) || !HASH.test(raw.pcmSHA256) || chunk.audioSha256 && raw.wavSHA256 !== chunk.audioSha256 || !Number.isSafeInteger(raw.sampleCount) || raw.sampleCount <= 0 || raw.sampleCount > 480000 || !Number.isFinite(chunk.elapsed) || (chunk.elapsed as number) < 0) fail('raw WAV source binding')
+    samples[String(i)] = raw.sampleCount
+    return { rawChunkIndex: i, startSample: Math.floor((chunk.elapsed as number) * 16 + .5), sampleCount: raw.sampleCount, usedSampleCount: 0 }
+  }).sort((a, b) => a.startSample - b.startSample || a.rawChunkIndex - b.rawChunkIndex)
+  for (let i = 0; i < placed.length; i++) {
+    const row = placed[i]
+    row.usedSampleCount = i + 1 < placed.length ? Math.min(row.sampleCount, placed[i + 1].startSample - row.startSample) : row.sampleCount
+    if (row.usedSampleCount <= 0) fail('shadowed timeline row')
+  }
+  const total = Math.max(...placed.map(r => r.startSample + r.sampleCount))
+  if (total > 4 * 3600 * 16000 || !equal(placed, input.timeline.rows) || total !== input.timeline.total || !HASH.test(input.timeline.pcmSHA256)) fail('timeline geometry drift')
+  if (!HASH.test(input.cohort.modelSHA256) || !HASH.test(input.cohort.cliSHA256)) fail('model/CLI binding required')
+  // Validate every export row, including short confusers that decide will reject.
+  const native = boundedJSON(input.cohort.nativeJSON, 64 * 1024 * 1024), embeddings = boundedJSON(input.cohort.embeddingsJSON, 96 * 1024 * 1024)
+  if (!native || !Array.isArray(native.segments) || native.segments.length > 20000 || !Array.isArray(embeddings) || embeddings.length > 10000) fail('cohort schema bound')
+  const vector = (v: unknown) => Array.isArray(v) && v.length === 256 && v.every(x => typeof x === 'number' && Number.isFinite(x) && Math.abs(x) <= 1e10) && v.some(x => x !== 0)
+  for (const s of native.segments) {
+    if (!s || typeof s.speakerId !== 'string' || !s.speakerId || s.speakerId.length > 100 || !Number.isFinite(s.startTimeSeconds) || !Number.isFinite(s.endTimeSeconds) || s.startTimeSeconds < 0 || s.endTimeSeconds <= s.startTimeSeconds || s.endTimeSeconds > total / 16000 + 1) fail('native turn bound')
+    if (s.embedding !== undefined && !vector(s.embedding)) fail('native vector must be finite nonzero256d')
+  }
+  const excludedEmbeddingWindows: Array<{ originalIndex: number; startTime: number; endTime: number; windowSHA256: string; reason: 'window-crosses-capture-end' }> = []
+  const effectiveEmbeddings: ShadowEmbeddingRow[] = []
+  for (const [originalIndex, e] of embeddings.entries()) {
+    if (!e || !vector(e.embedding256) || !Number.isInteger(e.cluster) || e.cluster < 0 || e.cluster > 9999 || !Number.isFinite(e.startTime) || !Number.isFinite(e.endTime) || e.startTime < 0 || e.endTime <= e.startTime) fail('embedding must be bounded finite nonzero256d')
+    if (e.endTime > total / 16000 + 1e-6) {
+      // A vector covering source AND terminal model padding is not a vector for
+      // a cropped source window. Never clip its time or reuse its vector. The
+      // immutable-view protocol can exclude the ENTIRE terminal window while
+      // retaining the complete export in its source binding and audit receipt.
+      // The physical writer remains strict. Windows outside the capture or an
+      // overrun beyond the native export's one-second envelope remain fatal.
+      if (!immutableView || e.startTime >= total / 16000 || e.endTime > total / 16000 + 1) fail('embedding must be bounded finite nonzero256d')
+      excludedEmbeddingWindows.push({ originalIndex, startTime: e.startTime, endTime: e.endTime, windowSHA256: sha(JSON.stringify(e)), reason: 'window-crosses-capture-end' })
+    } else effectiveEmbeddings.push(e)
+  }
+  const { turns, dropped } = nativeTurns(native.segments, total / 16000)
+  if (dropped) fail('malformed/dropped native turn')
+  const ledgerRawProtection = ledgerProtection(input.ledgerJSONL, map, oldMap)
+  const currentRoots = new Map(map.map(m => [m.raw, m.chunk]))
+  // Historical ASR occasionally extends a generic Ext segment past its own WAV.
+  // An immutable view can quarantine that WHOLE row, never clip it or synthesize
+  // replacement support. Keep every native/confuser row and bind the unmodified
+  // original JSON; physical writers and malformed Voice/human rows still refuse.
+  const excludedProductionSegments: Array<{ rawChunkIndex: number; segmentIndex: number; rowSHA256: string; startSec: number; endSec: number; sourceSamples: number; overrunSamples: number; reason: 'generic-segment-crosses-source-end' }> = []
+  const productionEntries = entries.map(e => ({ ...e, chunk: { ...e.chunk, segments: ((e.chunk.segments as Doc[] | undefined) ?? []).filter((s, segmentIndex) => {
+    if (!s || typeof s.speaker !== 'string' || !Number.isFinite(s.startSec) || !Number.isFinite(s.endSec) || s.startSec < 0 || s.endSec <= s.startSec) fail('production segment coordinates')
+    const sourceSamples = samples[String(e.chunkIndex)], overrunSamples = s.endSec * 16000 - sourceSamples
+    if (overrunSamples <= 1e-6) return true
+    const currentRow = currentEntries[e.chunkIndex].chunk
+    const currentRoot = currentRoots.get(e.chunkIndex)
+    if (!immutableView || s.speaker !== 'Ext' || e.chunk.speaker !== 'Ext' || currentRow.speaker !== 'Ext'
+      || currentRoot && currentRoot.speaker !== 'Ext' || s.startSec * 16000 >= sourceSamples || overrunSamples > 16000
+      || ledgerRawProtection.has(e.chunkIndex)
+      || [e.chunk, currentRow, ...(currentRoot ? [currentRoot] : [])].some(c => hasMacSpeakerProtectedRows(c) || !equal(rows(c), rows(e.chunk)) && c !== e.chunk)) fail('production segment coordinates')
+    excludedProductionSegments.push({ rawChunkIndex: e.chunkIndex, segmentIndex, rowSHA256: sha(JSON.stringify(s)), startSec: s.startSec, endSec: s.endSec, sourceSamples, overrunSamples, reason: 'generic-segment-crosses-source-end' })
+    return false
+  }) } }))
+  const production = shadowProductionSegments(productionEntries, placed, total)
+  if (production.segs.length > 40000) fail('production segment bound')
+  const decision = decide(production.segs, total, turns, effectiveEmbeddings)
+  const protection = new Set(ledgerRawProtection)
+  for (const row of excludedProductionSegments) protection.add(row.rawChunkIndex)
+  for (const m of map) if ([m.chunk, m.entry].some(c => hasMacSpeakerProtectedRows(c) || rows(c).some(r => !VOICE.test(r.speaker)))) protection.add(m.raw)
+  const wordSpans = batchBindings(original, current, entries, samples, protection)
+  // Anonymous estimate aliases inherit the original machine localization; they
+  // must not confuse a structural Ext/Voice boundary with a human correction.
+  // Keep this separate from the stronger interval-projection exclusion above.
+  // No tail or mixed boundary admitted here becomes new acoustic support.
+  const humanProtection = new Set(ledgerRawProtection)
+  for (const row of excludedProductionSegments) humanProtection.add(row.rawChunkIndex)
+  const named = (row: Doc) => typeof row.speaker === 'string' && row.speaker.trim() !== ''
+    && !/^(?:Voice [1-9]\d?|Ext|Unknown|Unidentified(?: \d+)?|Speaker \d+)$/.test(row.speaker)
+  for (const [i, m] of map.entries()) {
+    const before = oldMap[i]
+    if (!equal(rows(before.chunk), rows(m.chunk)) || !equal(rows(before.entry), rows(m.entry))
+      || [m.chunk, m.entry].some(c => hasMacSpeakerProtectedRows(c) || rows(c).some(named))) humanProtection.add(m.raw)
+  }
+  for (const word of wordSpans) if (word.protected) for (const raw of word.protectedRaw) humanProtection.add(raw)
+  // A view keeps all base bytes and previous proof receipts intact. It may use
+  // this inspection, but it is not permission to feed an old receipt into the
+  // physical-patch writer below.
+  return { original, current, oldMap, map, entries, currentEntries, samples, placed, total, native, embeddings: effectiveEmbeddings, excludedEmbeddingWindows, excludedProductionSegments, turns, production, decision, protection, humanProtection, wordSpans }
+}
+
+/** Complete immutable-source inspection for the call-wide derived view. This
+ * accepts existing proof metadata because consumers MUST validate it against
+ * the unmodified base, and MUST NOT patch that base. No native execution is
+ * authenticated by a caller-supplied hash. */
+export function inspectMacSpeakerShadowCallSource(input: MacShadowRepairInput) {
+  return inspectSource(input, true)
+}
+
+export function buildMacSpeakerShadowRepair(input: MacShadowRepairInput): MacShadowRepairPreview {
+  const { sourceBindingSHA256 } = input
+  const { original, current, oldMap, map, entries, currentEntries, samples, placed, total, native, embeddings, turns, production, decision, protection, wordSpans } = inspectSource(input, false)
+  const withheld = new Map<number, string>(), chosen = new Map<number, { from: string; to: string }>()
+  let candidateChunks = 0
+  for (let i = 0; i < oldMap.length; i++) {
+    const before = oldMap[i], now = map[i], to = decision.remap.get(before.chunk.speaker)
+    if (!to) { withheld.set(now.raw, VOICE.test(before.chunk.speaker) ? 'no-qualified-merge' : 'non-Voice'); continue }
+    candidateChunks++
+    if (!VOICE.test(to) || !VOICE.test(before.chunk.speaker)) { withheld.set(now.raw, 'non-Voice'); continue }
+    if (protection.has(now.raw)) { withheld.set(now.raw, 'human-or-named-protected'); continue }
+    if (!equal(rows(before.chunk), rows(now.chunk)) || !equal(rows(now.chunk), rows(now.entry)) || rows(now.chunk).some(r => r.speaker !== before.chunk.speaker)) { withheld.set(now.raw, 'mixed-or-current-surfaces-differ'); continue }
+    const row = placed.find(r => r.rawChunkIndex === now.raw)!, segs = production.segs.filter((_, j) => production.sources[j].chunkIndex === now.raw)
+    const cluster = decision.rows.get(before.chunk.speaker)?.topCluster
+    if (row.usedSampleCount !== row.sampleCount || !segs.length || !cluster || segs.some(([lo, hi, label]) => label !== before.chunk.speaker || [...exactOverlap(lo, hi, decision.index)].some(([key]) => key !== cluster))) { withheld.set(now.raw, 'outside-fully-qualified-chunk'); continue }
+    chosen.set(now.raw, { from: now.chunk.speaker, to })
+  }
+  for (const raw of chosen.keys()) if (protection.has(raw)) { chosen.delete(raw); withheld.set(raw, 'batch-protected-or-crosses-unqualified-boundary') }
+  // Remove entire chunks to a fixed point; never write a partly unqualified word.
+  let changed = true
+  while (changed) {
+    changed = false
+    for (const word of wordSpans) {
+      const touched = word.raw.filter(r => chosen.has(r))
+      if (touched.length && (word.protected || word.raw.some(r => !chosen.has(r)) || new Set(touched.map(r => chosen.get(r)!.to)).size !== 1)) for (const raw of touched) { chosen.delete(raw); withheld.set(raw, 'batch-protected-or-crosses-unqualified-boundary'); changed = true }
+    }
+  }
+  const protectedRawChunkIndices = [...protection].sort((a, b) => a - b)
+  const assignments = [...chosen].map(([rawChunkIndex, pair]) => ({ rawChunkIndex, ...pair, evidenceId: `shadow256_${sourceBindingSHA256}`, batchWords: wordSpans.filter(w => w.raw[0] === rawChunkIndex).map(w => w.binding) }))
+  let plan: MacSpeakerRepairPlan | null = assignments.length ? { id: `shadow256_${sourceBindingSHA256.slice(0, 24)}`, evidenceHash: sourceBindingSHA256, expectedDocumentHash: hashMacSpeakerDocument(current), protectedRawChunkIndices, rawAudioSampleCounts: samples, assignments } : null
+  let after = current
+  if (plan) {
+    try { after = projectMacSpeakerRepair(current, plan).doc }
+    catch { for (const a of assignments) withheld.set(a.rawChunkIndex, 'strict-projector-refused'); plan = null }
+  }
+  const rootLabels = (d: Doc) => new Set(d.chunks.map((c: Doc) => c.speaker)).size
+  const labels = (d: Doc) => new Set([
+    ...d.chunks.flatMap((c: Doc) => rows(c).map(r => r.speaker)),
+    ...d.chunkEntries.flatMap((e: Doc) => rows(e.chunk).map(r => r.speaker)),
+    ...(d.batchSegments ?? []).flatMap((b: Doc) => b.speakerWords.map((w: Doc) => w.speaker)),
+  ].filter(v => typeof v === 'string' && v)).size
+  // An acoustic proposal alone is not the requested fragmentation repair. The
+  // CURRENT root display must improve without increasing any stored label set.
+  if (plan && (rootLabels(after) >= rootLabels(current) || labels(after) > labels(current))) {
+    for (const assignment of plan.assignments) withheld.set(assignment.rawChunkIndex, 'non-reducing-plan')
+    plan = null; after = current
+  }
+  return { plan, provenance: { productionQualified: false, nativeAuthenticated: false, sourceBindingSHA256, originalSHA256: sha(input.originalJSON), currentSHA256: sha(input.currentJSON), ledgerSHA256: sha(input.ledgerJSONL) }, protectedRawChunkIndices, withheld: [...withheld].map(([rawChunkIndex, reason]) => ({ rawChunkIndex, reason })), counts: { raw: entries.length, spoken: map.length, embeddings: embeddings.length, acousticAcceptedMerges: decision.accepted, candidateChunks, assignedChunks: plan?.assignments.length ?? 0, protectedChunks: protection.size, withheldChunks: withheld.size, beforeLabels: labels(current), afterLabels: labels(after), beforeRootLabels: rootLabels(current), afterRootLabels: rootLabels(after) } }
+}

package/server/lib/mac-voice-assignment.ts

@@ -77,6 +77,7 @@ import {
   greedyDiversitySelect,
 } from './voice-enrolment-selection.js'
 import { VOICEPRINT_MIN_AUDIO_SEC } from './voiceprint-floor.js'
+import { sameSpeakerName } from './speaker-name-nfc.js'
 
 /** A turn shorter than this is not sliced out of a mixed chunk. */
 export const MAC_ASSIGN_MIN_SLICE_SEC = 3.0
@@ -196,24 +197,55 @@ function emptyReport(person: string, skipped: MacAssignSkip): MacAssignReport {
   }
 }
 
-interface Turn { speaker: string; startSec?: number; endSec?: number }
-interface Unit { raw: number; slice: { startSec: number; endSec: number } | null }
+export interface MacVoiceTurn { speaker: string; startSec?: number; endSec?: number }
+export interface MacVoiceSlice { startSec: number; endSec: number }
+interface Unit { raw: number; slice: MacVoiceSlice | null }
 
-function turnsOf(chunk: Record<string, unknown> | undefined): Turn[] {
+export function macVoiceTurnsOf(chunk: Record<string, unknown> | undefined): MacVoiceTurn[] {
   if (!chunk) return []
   const final = chunk.finalSegments
   const live = chunk.segments
   const list = Array.isArray(final) && final.length > 0 ? final : Array.isArray(live) ? live : []
-  return list.filter((t): t is Turn => !!t && typeof t === 'object' && typeof (t as Turn).speaker === 'string')
+  return list.filter((t): t is MacVoiceTurn => !!t && typeof t === 'object' && typeof (t as MacVoiceTurn).speaker === 'string')
 }
 
-function spread<T>(items: T[], max: number): T[] {
+/** At most `max` items, evenly spread across the list. */
+export function spreadMacVoiceUnits<T>(items: T[], max: number): T[] {
   if (items.length <= max) return items
   const out: T[] = []
   for (let k = 0; k < max; k++) out.push(items[Math.floor((k * items.length) / max)])
   return out
 }
 
+/**
+ * What one chunk contributes for a voice: `whole` when every turn is that voice or the chunk has no turns,
+ * otherwise that voice's turns long enough to slice, inset at each end. Empty means nothing here is usable.
+ * Voice suggestions (mac-voice-suggestions.ts) use the same per-chunk rule. They do not pick the same units
+ * overall: this plan spreads to the cap before the 2 s filter, suggestions spread after it, as the study did.
+ */
+export function macVoiceChunkUnits(turns: MacVoiceTurn[], voice: string): 'whole' | MacVoiceSlice[] {
+  if (turns.length === 0 || turns.every(t => t.speaker === voice)) return 'whole'
+  return turns
+    .filter(t => t.speaker === voice
+      && Number.isFinite(t.startSec) && Number.isFinite(t.endSec)
+      && (t.endSec as number) - (t.startSec as number) >= MAC_ASSIGN_MIN_SLICE_SEC)
+    .map(t => ({ startSec: (t.startSec as number) + MAC_ASSIGN_SLICE_INSET_SEC, endSec: (t.endSec as number) - MAC_ASSIGN_SLICE_INSET_SEC }))
+}
+
+/** The audio a unit addresses and its length. Null when the WAV is unusable or the slice is empty. */
+export function macVoiceUnitAudio(wav: Buffer, slice: MacVoiceSlice | null): { audio: Buffer; sec: number } | null {
+  const pcm = readPcm16Mono(wav)
+  if (!pcm || pcm.sampleRate !== SAMPLE_RATE) return null
+  let audio = wav
+  if (slice) {
+    const from16 = Math.max(0, Math.round(slice.startSec * SAMPLE_RATE)) * 2
+    const to16 = Math.min(pcm.pcm.length, Math.round(slice.endSec * SAMPLE_RATE) * 2)
+    if (to16 <= from16) return null
+    audio = pcm16ToWav(pcm.pcm.subarray(from16, to16))
+  }
+  return { audio, sec: (audio.length - 44) / (SAMPLE_RATE * 2) }
+}
+
 function centroid(vectors: Float32Array[]): Float32Array | null {
   if (vectors.length === 0) return null
   const out = new Float32Array(vectors[0].length)
@@ -239,7 +271,7 @@ export async function planMacVoiceAssignment(input: MacAssignInput): Promise<Mac
   if (isPlaceholderLabel(to)) return none('unknown_profile')
   if (input.ownerLabel && sameName(to, input.ownerLabel)) return none('owner_profile')
   let profile: { name: string; embeddings: number[][]; sources?: string[] } | undefined
-  try { profile = readVoiceProfiles().profiles.find(p => p.name === to) } catch { return none('store_unavailable') }
+  try { profile = readVoiceProfiles().profiles.find(p => sameSpeakerName(p.name, to)) } catch { return none('store_unavailable') }
   if (!profile) return none('unknown_profile')
   if (!chunkEmbeddingsEnabled()) return none('disabled')
   if (!isEmbeddingAvailable()) return none('store_unavailable')
@@ -261,22 +293,16 @@ export async function planMacVoiceAssignment(input: MacAssignInput): Promise<Mac
   const report = emptyReport(to, null)
   const units: Unit[] = []
   for (const raw of [...new Set(rawIndices)].sort((a, b) => a - b)) {
-    const turns = turnsOf(chunksByRaw.get(raw))
-    if (turns.length === 0 || turns.every(t => t.speaker === from)) { units.push({ raw, slice: null }); continue }
-    const usable = turns.filter(t => t.speaker === from
-      && Number.isFinite(t.startSec) && Number.isFinite(t.endSec)
-      && (t.endSec as number) - (t.startSec as number) >= MAC_ASSIGN_MIN_SLICE_SEC)
-    if (usable.length === 0) { report.multiVoiceSkipped += 1; continue }
-    for (const t of usable) units.push({
-      raw,
-      slice: { startSec: (t.startSec as number) + MAC_ASSIGN_SLICE_INSET_SEC, endSec: (t.endSec as number) - MAC_ASSIGN_SLICE_INSET_SEC },
-    })
+    const chunkUnits = macVoiceChunkUnits(macVoiceTurnsOf(chunksByRaw.get(raw)), from)
+    if (chunkUnits === 'whole') { units.push({ raw, slice: null }); continue }
+    if (chunkUnits.length === 0) { report.multiVoiceSkipped += 1; continue }
+    for (const slice of chunkUnits) units.push({ raw, slice })
   }
 
   const embeddings: Float32Array[] = []
   const secs: number[] = []
   let foundAudio = 0
-  const picked = spread(units, MAC_ASSIGN_MAX_UNITS)
+  const picked = spreadMacVoiceUnits(units, MAC_ASSIGN_MAX_UNITS)
   stats.units = picked.length
   const deadline = started + (input.deadlineMs ?? deadlineMsFromEnv())
   for (const unit of picked) {
@@ -302,16 +328,9 @@ export async function planMacVoiceAssignment(input: MacAssignInput): Promise<Mac
     if (hit) { stats.reused += 1; embeddings.push(hit.vec); secs.push(hit.sec); continue }
     let wav: Buffer
     try { wav = readFileSync(path) } catch { report.missingAudio += 1; continue }
-    const pcm = readPcm16Mono(wav)
-    if (!pcm || pcm.sampleRate !== SAMPLE_RATE) continue
-    let audio = wav
-    if (unit.slice) {
-      const from16 = Math.max(0, Math.round(unit.slice.startSec * SAMPLE_RATE)) * 2
-      const to16 = Math.min(pcm.pcm.length, Math.round(unit.slice.endSec * SAMPLE_RATE) * 2)
-      if (to16 <= from16) continue
-      audio = pcm16ToWav(pcm.pcm.subarray(from16, to16))
-    }
-    const sec = (audio.length - 44) / (SAMPLE_RATE * 2)
+    const cut = macVoiceUnitAudio(wav, unit.slice)
+    if (!cut) continue
+    const { audio, sec } = cut
     if (sec < VOICEPRINT_MIN_AUDIO_SEC) { report.shortSkipped += 1; continue }
     const embedding = extractEmbedding(audio)
     stats.embedded += 1

package/server/lib/mac-voice-display.ts

@@ -0,0 +1,195 @@
+// 6.96.3: how a saved Mac recording's anonymous voices are DISPLAYED in speaker review.
+//
+// WHY. One real call returned 61 rows from GET /meeting/:sessionId/speakers: 58 `Voice N`, Ext and two
+// names. Speaker shadow already logged, for that call, which of those voices its gates would merge
+// (speaker-shadow.ts, proposals.accepted). Miles approved showing that at read time: merged groups
+// first, then one expandable row for voices under 15 s after merging. Display only.
+//
+// WHAT THIS IS NOT. It writes nothing and applies nothing: no label, name, sample, packet or sidecar
+// changes, and the words stay one click away under the same labels. There is no flag. The field exists
+// only when this Mac has a `proposed` shadow record for this session, so a Mac with shadow off answers
+// exactly what it answered before.
+//
+// THE RESPONSE WINS OVER THE RECORD. The record is a log from when the meeting finished; the labels in
+// the current response are the truth. A voice named since then leaves its group, and a current voice the
+// record never saw still gets a row. Every anonymous label in the response lands in exactly one bucket.
+//
+// FAILS TO ABSENT. Any malformed record, unreadable store or unexpected value omits the field, which
+// means Control shows what it shows today. Nothing here throws.
+
+import { readdir, stat } from 'node:fs/promises'
+import { join } from 'node:path'
+import { isVoiceLabel } from './mac-voice-labels.js'
+import { readSpeakerShadowRecords, SPEAKER_SHADOW_RECORDS } from './speaker-shadow.js'
+import { speakerShadowRoot } from './speaker-shadow-models.js'
+
+/** A voice with less known speech than this, after merging, folds into the one expandable row. */
+export const VOICE_DISPLAY_FOLD_UNDER_SECONDS = 15 as const
+/** Past this the store is not read at all. At 12 runs a day it is about two years of records. */
+const MAX_STORE_BYTES = 256 * 1024 * 1024
+const MAX_RECORDS_PER_SESSION = 100
+const EVIDENCE_CACHE_MAX = 32
+
+export interface VoiceDisplay {
+  source: 'speaker-shadow'
+  recordedAt: string
+  foldUnderSeconds: typeof VOICE_DISPLAY_FOLD_UNDER_SECONDS
+  /** Two or more labels, largest first. */
+  groups: Array<{ id: string; labels: string[]; seconds: number | null }>
+  /** 15 s or more, or unknown duration. */
+  singles: Array<{ label: string; seconds: number | null }>
+  /** Known duration under 15 s after merging. */
+  short: Array<{ label: string; seconds: number }>
+}
+
+/** The parts of one shadow record display needs. Validated once, so building a view cannot fail. */
+export interface VoiceDisplayEvidence {
+  recordedAt: string
+  groups: Array<{ id: string; labels: string[] }>
+  seconds: ReadonlyMap<string, number>
+}
+
+const isRecord = (value: unknown): value is Record<string, unknown> => !!value && typeof value === 'object' && !Array.isArray(value)
+const validSeconds = (value: unknown): value is number => typeof value === 'number' && Number.isFinite(value) && value >= 0
+
+/** Null on any problem at all. A label may appear once across the whole record. */
+export function voiceDisplayEvidence(record: unknown): VoiceDisplayEvidence | null {
+  if (!isRecord(record) || record.status !== 'proposed' || typeof record.recordedAt !== 'string' || !record.recordedAt) return null
+  const proposals = record.proposals
+  if (!isRecord(proposals)) return null
+  const { accepted, rejected, untouched } = proposals
+  if (!Array.isArray(accepted) || !Array.isArray(rejected) || !Array.isArray(untouched)) return null
+
+  const seconds = new Map<string, number>()
+  const seen = new Set<string>()
+  const claim = (label: unknown): label is string => {
+    if (typeof label !== 'string' || seen.has(label)) return false
+    seen.add(label)
+    return true
+  }
+  const groups: VoiceDisplayEvidence['groups'] = []
+  const ids = new Set<string>()
+  for (const group of accepted) {
+    if (!isRecord(group) || typeof group.cluster !== 'string' || !group.cluster || ids.has(group.cluster)) return null
+    ids.add(group.cluster)
+    if (!Array.isArray(group.merge) || !Array.isArray(group.voices)) return null
+    const labels: string[] = []
+    for (const label of [group.keepLabel, ...group.merge]) {
+      if (!claim(label)) return null
+      labels.push(label)
+    }
+    for (const row of group.voices) {
+      if (!isRecord(row) || typeof row.voice !== 'string' || !labels.includes(row.voice) || !validSeconds(row.seconds)) return null
+      seconds.set(row.voice, row.seconds)
+    }
+    groups.push({ id: `shadow-${group.cluster}`, labels })
+  }
+  for (const row of [...rejected, ...untouched]) {
+    if (!isRecord(row) || !claim(row.voice) || !validSeconds(row.seconds)) return null
+    seconds.set(row.voice, row.seconds)
+  }
+  return { recordedAt: record.recordedAt, groups, seconds }
+}
+
+const bySecondsThenLabel = (a: { seconds: number | null; key: string }, b: { seconds: number | null; key: string }): number => {
+  if (a.seconds === null && b.seconds !== null) return 1
+  if (a.seconds !== null && b.seconds === null) return -1
+  if (a.seconds !== null && b.seconds !== null && a.seconds !== b.seconds) return b.seconds - a.seconds
+  return a.key.localeCompare(b.key, 'en', { numeric: true })
+}
+
+/** Pure. `currentLabels` is the label of every row in the response being answered. */
+export function buildVoiceDisplay(evidence: VoiceDisplayEvidence, currentLabels: readonly string[]): VoiceDisplay {
+  const current = new Set(currentLabels.filter(isVoiceLabel))
+  const secondsOf = (label: string): number | null => evidence.seconds.get(label) ?? null
+  const groups: VoiceDisplay['groups'] = []
+  const singles: VoiceDisplay['singles'] = []
+  const short: VoiceDisplay['short'] = []
+  const place = (label: string): void => {
+    const seconds = secondsOf(label)
+    if (seconds !== null && seconds < VOICE_DISPLAY_FOLD_UNDER_SECONDS) short.push({ label, seconds })
+    else singles.push({ label, seconds })
+  }
+
+  const grouped = new Set<string>()
+  for (const group of evidence.groups) {
+    const labels = group.labels.filter(label => current.has(label))
+    for (const label of labels) grouped.add(label)
+    if (labels.length === 1) place(labels[0])
+    else if (labels.length >= 2) {
+      const known = labels.map(secondsOf)
+      const seconds = known.every((s): s is number => s !== null)
+        ? Math.round(known.reduce((sum, s) => sum + s, 0) * 100) / 100
+        : null
+      groups.push({ id: group.id, labels, seconds })
+    }
+  }
+  for (const label of current) if (!grouped.has(label)) place(label)
+
+  groups.sort((a, b) => bySecondsThenLabel({ seconds: a.seconds, key: a.labels[0] }, { seconds: b.seconds, key: b.labels[0] }))
+  singles.sort((a, b) => bySecondsThenLabel({ seconds: a.seconds, key: a.label }, { seconds: b.seconds, key: b.label }))
+  short.sort((a, b) => bySecondsThenLabel({ seconds: a.seconds, key: a.label }, { seconds: b.seconds, key: b.label }))
+  return { source: 'speaker-shadow', recordedAt: evidence.recordedAt, foldUnderSeconds: VOICE_DISPLAY_FOLD_UNDER_SECONDS, groups, singles, short }
+}
+
+/**
+ * Who a voice suggestion is for: one entry per group, single and short row when there is a display,
+ * otherwise one per anonymous label. Ext and names never appear, on either path.
+ */
+export function voiceDisplayScopes(display: VoiceDisplay | null, currentLabels: readonly string[]): string[][] {
+  if (!display) return [...new Set(currentLabels.filter(isVoiceLabel))].map(label => [label])
+  return [
+    ...display.groups.map(group => group.labels),
+    ...display.singles.map(row => [row.label]),
+    ...display.short.map(row => [row.label]),
+  ]
+}
+
+// Keyed on every proposal file's name, size and mtime. /speakers is opened again and again on one
+// meeting; a session with no record would otherwise scan the whole store each time.
+const evidenceCache = new Map<string, { signature: string; evidence: VoiceDisplayEvidence | null }>()
+
+export function __clearVoiceDisplayCacheForTests(): void { evidenceCache.clear() }
+
+async function storeSignature(root: string): Promise<string | null> {
+  const names = (await readdir(root)).filter(name => name === SPEAKER_SHADOW_RECORDS || /^proposals-[0-9TZ-]+\.jsonl$/.test(name)).sort()
+  if (names.length === 0) return null
+  let total = 0
+  const parts: string[] = []
+  for (const name of names) {
+    const info = await stat(join(root, name))
+    total += info.size
+    parts.push(`${name}:${info.size}:${info.mtimeMs}`)
+  }
+  return total > MAX_STORE_BYTES ? null : parts.join('|')
+}
+
+/** The latest `proposed` record for a session. Null when there is none or anything about it is off. Never throws. */
+export async function readVoiceDisplayEvidence(sessionId: string, root: string = speakerShadowRoot()): Promise<VoiceDisplayEvidence | null> {
+  try {
+    const signature = await storeSignature(root)
+    if (signature === null) return null
+    const key = `${root}\u0000${sessionId}`
+    const cached = evidenceCache.get(key)
+    if (cached?.signature === signature) return cached.evidence
+    const records = await readSpeakerShadowRecords(root, { sessionId, limit: MAX_RECORDS_PER_SESSION })
+    const latest = records.find(record => record.status === 'proposed')
+    const evidence = latest ? voiceDisplayEvidence(latest) : null
+    evidenceCache.delete(key)
+    evidenceCache.set(key, { signature, evidence })
+    while (evidenceCache.size > EVIDENCE_CACHE_MAX) evidenceCache.delete(evidenceCache.keys().next().value as string)
+    return evidence
+  } catch {
+    return null
+  }
+}
+
+/** The optional `voiceDisplay` field. Null means leave it out. Never throws. */
+export async function readVoiceDisplay(sessionId: string, currentLabels: readonly string[], root: string = speakerShadowRoot()): Promise<VoiceDisplay | null> {
+  try {
+    const evidence = await readVoiceDisplayEvidence(sessionId, root)
+    return evidence ? buildVoiceDisplay(evidence, currentLabels) : null
+  } catch {
+    return null
+  }
+}