[Epic] First-class voice enrollment (active + passive + on-device)

[silverstein]

Tracking issue for making voice enrollment a first-class feature. Design + phasing: PR #495 (docs/plans/voice-enrollment.md).

Problem: the engine exists (embeddings, voices.db, blended profiles, L2/L3 attribution) but there's no product around it — no desktop UI, no MCP tool, no passive learning, no backfill. Surfaced by first-run testing:

• Recall chat panel text is not selectable or copyable (inherits body user-select: none) #490 — Recall panel text not copyable
• Configured "Your Name" is ignored for solo recordings — self-name normalization is gated behind calendar attendees #491 — configured "Your Name" ignored on solo captures
• Onboarding: "10-second test recording" reads like voice enrollment but isn't — no post-Stop feedback and no desktop enrollment path #492 — onboarding "test recording" reads like voice enrollment but isn't; no desktop enroll path

Approach (see PR #495 for the full design + codex adversarial review):

Phase 1 (next release) — safe foundation + active enrollment:

• Immutable provenance-rich voice_samples schema; one active profile per (person, model); never blend across models (fixes model-version-blind matching)
• Quality-producing embed_solo_clip primitive (16kHz, multi-window, reject inconsistent clips)
• Matching returns structured evidence + abstains on ambiguity; per-model calibrated thresholds
• Desktop active enrollment (multi-sample, quality-gated) + status/rename/delete + "delete all voice data"
• Granular privacy settings (store embeddings / restricted eligibility default off / retention)
• Onboarding rework (Onboarding: "10-second test recording" reads like voice enrollment but isn't — no post-Stop feedback and no desktop enrollment path #492) + self-anchoring for solo captures (Configured "Your Name" is ignored for solo recordings — self-name normalization is gated behind calendar attendees #491)
• Ship diarization on-by-default in the desktop bundle (~34MB first-run download)

Phase 2 (fast-follow, gated on an adversarial FAR/FRR eval):

• Passive candidate capture → explicit promotion (default off, multi-meeting agreement, no restricted/degraded)
• Confirmed-label backfill from .embeddings sidecars (dry-run report; no clustering-to-guess-self)
• MCP enroll tool; household/multi-user; merge duplicate profiles
• Adversarial FAR/FRR eval harness as the release gate

Sequencing: start after the in-flight conversation-trust work merges to avoid colliding on commands.rs/pipeline.rs.

[silverstein]

Phase 1 progress: the CLI-first slice is merged in #518 (builds on the #505 schema and #506 privacy config).

Shipped: quality-gated single-speaker enrollment (embed_solo_clip with multimodal and short-speech rejection), the minutes voice command family (enroll, status, list, test, remove, delete-all), evidence-based model-compatible matching, transactional delete-all sweep including sidecars and WAL, and idempotent legacy-profile migration.

Still open for Phase 1: the desktop enrollment UI, onboarding rework (#492), sidecar envelope, pipeline matching integration, and #491 self-anchoring. Those touch surfaces that are sequenced behind the conversation-trust landing. One follow-up flagged in the PR: the real-model wav-to-embedding integration test is ignored pending a model-equipped Mac run.

[silverstein]

Real-hardware QA of the CLI slice (macOS, Apple Silicon, real microphone): PASS.

Full loop verified end to end: minutes voice enroll produced a quality-evidenced profile (25.6 dB SNR, 0.00% clipping, 0.787 window consistency, wespeaker_en_voxceleb_CAM++_v0.3), and minutes voice test matched at 0.985 similarity against the 0.650 threshold with evidence output. This also covers the ignored real-model integration path flagged in #518.

The dogfood run surfaced four polish items for the desktop/onboarding phase, none release-blocking:

1. Enrollment spins up the full recording stack (loads the whisper sidecar, starts the screen-capture worker, logged dropped_chunks=384). Enrollment should be a mic-only capture path: faster start, no wasted transcription, and no screen worker during a biometric enrollment.
2. No speak-now affordance: the CLI does not tell the user to start reading aloud, so a quiet first run gets rejected (observed: 2.3s of speech in a 25s window). A visible countdown plus a read-this-sentence prompt fixes first-try conversion.
3. voice test is an 8s window with the same 5s speech floor, leaving under a second of slack after reaction time and VAD edge-trimming (observed: 4.4s rejection on a normal attempt). Lengthen the window or start the clock at first detected speech.
4. onnxruntime BFCArena INFO logging floods the terminal during embedding and buries the success output. Suppress ort logs below warn in CLI surfaces.

[silverstein]

v0.23.0 shipped with the CLI slice as the headline feature: https://github.com/silverstein/minutes/releases/tag/v0.23.0 - first-class voice enrollment is now in users' hands (enroll, status, list, test, remove, delete-all). Remaining Phase 1 surfaces (desktop UI, onboarding, pipeline integration) stay sequenced behind the conversation-trust landing, plus the four CLI polish items above.