@gotcos/glasses-server 6.92.0 → 6.93.0

package/server/lib/transcript-cleanup-spans.ts

@@ -16,10 +16,10 @@
 //    near_name        a word the engine was unsure of (below 0.9) that is a capitalised word
 //                     one edit past fuzzy-correct's limit from a known name, or that sounds
 //                     like one (same consonant skeleton and onset); two or three short
-//                     adjacent pieces joined ("U coma"), none a function word; or the
-//                     leading-y rule at any confidence ("Yukoma" for Ukaoma)
+//                     adjacent pieces joined ("U cmya"), none a function word; or the
+//                     leading-y rule at any confidence ("Yukmya" for Ukamya)
 //    oov_capitalized  a capitalised word mid-sentence that is no known name and no common
-//                     word, within two words of a known name ("Grace Eucoma")
+//                     word, within two words of a known name ("Faith Eucmya")
 //    repeat           a phrase of 2 to 4 words (or one long word) said three or more
 //                     times in a row: the classic decoder loop
 //    Flagged words one word apart merge; a span holds at most 6 words (12 for a repeat)
@@ -104,7 +104,7 @@ const POSSESSIVE = /(['’]s)$/i
  *     attendees as targets Slack still became Black twice; they reach the model as context
  *     and drive near-name spans instead), and the curated dictionary and the owner
  *     ("strong" targets) rewrite only a word the engine was unsure of. With those rules the
- *     68 meetings got 27 fixes, all real (Manuj to Manoj 9 times, MarketPOS to MarktPOS).
+ *     68 meetings got 27 fixes, all real (Mohun to Mohan 9 times, MarketPOS to MarktPOS).
  *   - a term from THIS meeting's screen or files (the per-meeting vocabulary) may rewrite a
  *     capitalised word at any confidence: it is what was on screen while it was said.
  * Never into an inflection of the word, and a word that only opens a sentence only when the
@@ -159,11 +159,11 @@ interface NearInput { key: string; capitalized: boolean; probability?: number; j
 
 /**
  * The known name `key` is plausibly a mishearing of, or null.
- *   - the leading-y rule (meeting-ask-owner.ts): "Yukoma" is one edit from "Ukaoma" once its y goes
+ *   - the leading-y rule (meeting-ask-owner.ts): "Yukmya" is one edit from "Ukamya" once its y goes
  *   - otherwise only for a word the engine was unsure of (below NEAR_NAME_UNSURE, or unscored):
  *       a capitalised single word one edit past fuzzy-correct's own limit, or
  *       any word (or short joined pieces) that sounds like the name (soundsAlike)
- *   - joined pieces may also be within fuzzy-correct's own limit ("U coma" read as one word)
+ *   - joined pieces may also be within fuzzy-correct's own limit ("U cmya" read as one word)
  */
 function nearName(input: NearInput, targets: readonly NameTarget[], known: ReadonlySet<string>): string | null {
   const { key } = input
@@ -174,7 +174,7 @@ function nearName(input: NearInput, targets: readonly NameTarget[], known: Reado
   for (const t of targets) {
     if (Math.abs(t.key.length - key.length) > 2) continue
     const d = levenshtein(key, t.key)
-    // A single word spelled exactly like a name is the name. Pieces that join into one exactly ("Uka oma") are the
+    // A single word spelled exactly like a name is the name. Pieces that join into one exactly ("Uka mya") are the
     // strongest sign of a split name there is.
     if (d === 0 && !input.join) return null
     // "setting" next to a "Settings" label is the same word, not a misheard name.
@@ -260,12 +260,12 @@ export function selectSpans(
       if (near) flag(t.i, 'near_name', near)
     }
   }
-  // Two or three adjacent words joined: "U coma", "Yu ko ma". One of them must be short (a split fragment).
+  // Two or three adjacent words joined: "U cmya", "Yu ko ma". One of them must be short (a split fragment).
   for (let i = 0; i + 1 < tokens.length; i++) {
     for (const width of [2, 3]) {
       const group = tokens.slice(i, i + width)
       if (group.length < width || group.some(g => g.line !== tokens[i].line || !g.key || fixed.has(g.i) || known.has(g.key))) continue
-      // A split name is short pieces ("U coma", "Yu ko ma"); a long word plus a short one is just two words,
+      // A split name is short pieces ("U cmya", "Yu ko ma"); a long word plus a short one is just two words,
       // and a function word is never a piece of a name.
       if (!group.some(g => g.key.length <= 3) || group.some(g => g.key.length > 5 || FUNCTION_WORDS.has(g.key))) continue
       const joined = group.map(g => g.key).join('')
@@ -303,7 +303,7 @@ export function selectSpans(
     while (k < flagged.length) {
       const next = flagged[k]
       if (tokens[next].line !== tokens[from].line || next - to > 1 || next + 1 - from > SPAN_LIMITS.maxSpanWords) break
-      // Never across a clause: "Yukoma and eucoma, then U coma" is two passages, not one.
+      // Never across a clause: "Yukmya and eucmya, then U cmya" is two passages, not one.
       if (tokens.slice(to - 1, next).some(t => /[,.;:!?]["')\]]*$/.test(t.raw))) break
       if ([...Array(next - to).keys()].some(g => repeatTokens.has(to + g))) break
       for (const r of flags.get(next)!.reasons) reasons.add(r)

package/server/lib/transcript-cleanup-text.ts

@@ -6,7 +6,7 @@
 //
 // Tokens are whitespace runs inside a line's text. Each carries its `core`: the token
 // without leading and trailing punctuation, which is what is compared, flagged and
-// replaced, so "Ukoma," becomes "Ukaoma," and the comma stays exactly where it was.
+// replaced, so "Ukmya," becomes "Ukamya," and the comma stays exactly where it was.
 //
 // No IO and no model here.
 
@@ -156,8 +156,8 @@ export function alignWordInfo(tokens: Token[], words: readonly WordInfo[]): numb
 }
 
 /**
- * A consonant skeleton: how a word sounds with its vowels unsure. Ukaoma, Yukoma,
- * eucoma and "U coma" all read "km". Only ever used together with an onset check and
+ * A consonant skeleton: how a word sounds with its vowels unsure. Ukamya, Yukmya,
+ * eucmya and "U cmya" all read "km". Only ever used together with an onset check and
  * a length bound, because a two-letter skeleton alone matches far too much.
  */
 export function skeleton(word: string): string {
@@ -168,7 +168,7 @@ export function skeleton(word: string): string {
   return s.replace(/(.)\1+/g, '$1')
 }
 
-/** Vowel onset, counting a leading y before a vowel ("Yukoma") as a vowel. */
+/** Vowel onset, counting a leading y before a vowel ("Yukmya") as a vowel. */
 export function vowelOnset(word: string): boolean {
   const f = fold(word)
   return /^[aeiou]/.test(f) || /^y[aeiou]/.test(f)

package/server/lib/transcript-cleanup.ts

@@ -405,7 +405,7 @@ export function recordSuggestions(applied: readonly AppliedEdit[], pack: Context
   for (const edit of applied) {
     const heardKeys = new Set([...edit.original.matchAll(WORDS)].map(m => fold(m[0])))
     const ws = [...edit.replacement.matchAll(WORDS)].map(m => m[0])
-    // Longest first: "Grace Ukaoma" is one suggestion, not that plus "Ukaoma".
+    // Longest first: "Faith Ukamya" is one suggestion, not that plus "Ukamya".
     const candidates = [...new Set<string>([edit.replacement, ...ws.slice(0, -1).map((w, i) => `${w} ${ws[i + 1]}`), ...ws])]
       .sort((a, b) => b.split(' ').length - a.split(' ').length)
     const covered = new Set<string>()

package/server/lib/unsaved-audio-quarantine.ts

@@ -1,3 +1,4 @@
+import { readPhoneRecording, PHONE_SOURCE_TYPE } from './phone-recording.js'
 // Quarantine for unsaved meeting audio.
 //
 // Until 6.19.0 the session-audio purge (boot sweep + 60s interval) DELETED any
@@ -54,6 +55,8 @@ export const DISCARDED_RETENTION_MS = 24 * 60 * 60 * 1000
 export interface QuarantineReceipt { outcome: 'no_speech'; discarded: true; chunkFiles?: number; note: string }
 
 export interface UnsavedCapture {
+  sourceType?: typeof PHONE_SOURCE_TYPE
+  provenanceState?: 'admitted' | 'unverified'
   sessionId: string
   dirName: string
   quarantinedAt: string | null
@@ -250,8 +253,14 @@ export function listUnsavedCaptures(
     const manifest = readManifest(dirPath)
     const atMs = quarantinedAtMs(dirPath, manifest)
     const ownRetentionMs = effectiveRetentionMs(manifest, retentionMs)
+    const sessionId = manifest?.sessionId ?? dir.replace(/\.\d+$/, '')
+    let phoneSource: { sourceType?: typeof PHONE_SOURCE_TYPE; provenanceState?: 'admitted' | 'unverified' } = {}
+    if (sessionId.startsWith('phone_')) {
+      try { phoneSource = readPhoneRecording(sessionId) ? { sourceType: PHONE_SOURCE_TYPE, provenanceState: 'admitted' } : { provenanceState: 'unverified' } } catch { phoneSource = { provenanceState: 'unverified' } }
+    }
     captures.push({
-      sessionId: manifest?.sessionId ?? dir.replace(/\.\d+$/, ''),
+      sessionId,
+      ...phoneSource,
       dirName: dir,
       quarantinedAt: atMs != null ? new Date(atMs).toISOString() : null,
       ageHours: atMs != null ? Math.round(((Date.now() - atMs) / 3_600_000) * 10) / 10 : null,

package/server/lib/voice-directory.ts

@@ -117,7 +117,7 @@ let building: Promise<VoiceDirectorySnapshot> | null = null
 function sourceCounts(profile: VoiceProfile): Record<string, number> {
   const counts: Record<string, number> = {}
   for (const source of profile.sources ?? []) {
-    const key = source.startsWith('auto:') ? 'auto' : source
+    const key = source.startsWith('auto:') ? 'auto' : source.startsWith('mac-assign:') ? 'mac-assign' : source
     counts[key] = (counts[key] ?? 0) + 1
   }
   return counts

package/server/lib/voice-evidence-guard.ts

@@ -29,7 +29,7 @@ export const IMPORTED_EVIDENCE_ID_PATTERN = /^(imported|blended):/
 /** The two sidecars of the imported library. Never chunk audio, never `.g2-chunks.json`. */
 export const IMPORTED_EVIDENCE_PATH_PATTERN = /\.(?:import|derived)\.json$/i
 
-export type VoiceEvidenceReason = 'imported_read_only' | 'blended_derived_record'
+export type VoiceEvidenceReason = 'imported_read_only' | 'blended_derived_record' | 'phone_anonymous_capture'
 
 export interface VoiceEvidenceRefusal {
   status: 409
@@ -66,6 +66,7 @@ export function isImportedVoiceEvidence(value: unknown): boolean {
  */
 export function assertVoiceEvidenceSource(values: ReadonlyArray<unknown>): VoiceEvidenceRefusal | null {
   for (const value of values) {
+    if (typeof value === 'string' && /(?:^|[:/])phone_/.test(value)) return { status: 409, body: { error: 'iPhone microphone recordings are anonymous and cannot name or train voices', reason: 'phone_anonymous_capture', source: value } }
     const kind = importedEvidenceKind(value)
     if (!kind) continue
     return {

package/server/lib/voice-profile-store.ts

@@ -15,7 +15,7 @@ import { checkSpeakerName, type SpeakerNameRejection } from './speaker-name.js'
 import { existsSync, mkdirSync, readdirSync, statSync, unlinkSync } from 'node:fs'
 import { basename, dirname, join, resolve } from 'node:path'
 import { durableAtomicWriteFileSync, loadJsonOrQuarantine } from './atomic-fs.js'
-import { EVICTION_ORDER, provenanceTier } from './embedding-eviction.js'
+import { EVICTION_ORDER, isMacAssign, macAssignCapFor, provenanceTier } from './embedding-eviction.js'
 
 export interface VoiceProfile {
   name: string
@@ -194,7 +194,7 @@ export function junkNameReason(name: string): SpeakerNameRejection | null {
   const check = checkSpeakerName(name, { ownerLabel: name })
   if (check.ok) return null
   if (check.reason === 'too_long' || check.reason === 'too_many_words') return check.reason
-  // Sentence punctuation alone is not enough: "Luke H." is a legitimate legacy
+  // Sentence punctuation alone is not enough: "Theo M." is a legitimate legacy
   // alias. Three or more words with punctuation is speech, not a name.
   if (check.reason === 'sentence_like' && name.trim().split(/\s+/).length >= 3) return check.reason
   return null
@@ -555,9 +555,15 @@ export function mergeProfilesInStore(
 
   // Preserve stronger label provenance. Within one
   // source tier retain newer evidence, matching weakest-first/oldest eviction.
-  const keep = embeddings.map((_, i) => i).sort((a, b) =>
-    EVICTION_ORDER.indexOf(provenanceTier(sources[b])) - EVICTION_ORDER.indexOf(provenanceTier(sources[a])) || b - a,
-  ).slice(0, cap).sort((a, b) => a - b)
+  // A Mac-assigned sample ranks BELOW every other tier, the same rule chooseEviction applies (embedding-eviction.ts).
+  const rank = (i: number): number => isMacAssign(sources[i]) ? -1 : EVICTION_ORDER.indexOf(provenanceTier(sources[i]))
+  const kept = embeddings.map((_, i) => i).sort((a, b) => rank(b) - rank(a) || b - a).slice(0, cap)
+  // The merged profile obeys the Mac cap too: no more than 10, and never more than its other samples. Surplus Mac
+  // samples go oldest first, which is what chooseEviction does, so a merge cannot hold what an enrolment would refuse.
+  const macKept = kept.filter(i => isMacAssign(sources[i])).sort((a, b) => a - b)
+  const macAllowed = macAssignCapFor(kept.length, macKept.length)
+  const trimmed = new Set(macKept.slice(0, Math.max(0, macKept.length - macAllowed)))
+  const keep = kept.filter(i => !trimmed.has(i)).sort((a, b) => a - b)
   outcome.droppedToCap = embeddings.length - keep.length
   target.embeddings = keep.map(i => embeddings[i])
   target.sources = keep.map(i => sources[i])

package/server/lib/whisper-local.ts

@@ -1398,7 +1398,7 @@ function buildCorrections(): Array<[RegExp, string]> {
  * The profile's order, except that a key goes ahead of any earlier key whose words it
  * contains (6.83.1): "forward slash build and price" runs before "forward slash build", or
  * the shorter key would cut the longer phrase short. Chains the person relies on
- * ("Myles" then "Miles Ukoma") keep their order, since neither key contains the other.
+ * ("Myles" then "Miles Ukmya") keep their order, since neither key contains the other.
  */
 export function correctionOrder(corrections: Record<string, string>): Array<[string, string]> {
   const norm = (key: string) => ` ${key.trim().toLowerCase().split(/[\s-]+/).join(' ')} `

package/server/routes/health.ts

@@ -1,3 +1,5 @@
+import { phoneMicrophoneMeetingsCapability } from '../lib/phone-recording.js'
+import { macLiveVoicesSnapshot } from '../lib/mac-window-replay.js'
 import { DICTATION_CLEAN_MODELS } from '../lib/dictation-clean-models.js'
 import { customInstructionsCapability, learnedStyleCapability, polishOptionsCapability } from './dictation-polish-settings.js'
 import { PAIRING_PROTOCOL_VERSION } from '../lib/glasses-pairing.js'
@@ -550,6 +552,9 @@ healthRouter.get('/health', async (_req, res) => {
         finalization: getMeetingFinalizationSnapshot(),
       },
       ...(localFirstMeetings ? { localFirstMeetings } : {}),
+      phoneMicrophoneMeetings: phoneMicrophoneMeetingsCapability(getServerInstanceId()),
+      // 6.93.0 W8: which live Mac Voice N path is on (window by default since 6.93.0; anchors; off) and the window path's counts.
+      macLiveVoices: macLiveVoicesSnapshot(),
       ...(morningBrief ? { morningBrief } : {}),
       tasks: {
         gate: taskBridgeAvailable() ? 'ready' : 'disabled',
@@ -741,6 +746,9 @@ healthRouter.get('/models', async (req, res) => {
         },
       },
       ...(localFirstMeetings ? { localFirstMeetings } : {}),
+      phoneMicrophoneMeetings: phoneMicrophoneMeetingsCapability(getServerInstanceId()),
+      // 6.93.0 W8: which live Mac Voice N path is on (window by default since 6.93.0; anchors; off) and the window path's counts.
+      macLiveVoices: macLiveVoicesSnapshot(),
       tasks: {
         gate: taskBridgeAvailable() ? 'ready' : 'disabled',
       },