feat(media): async diarization pipeline with speaker naming (D1-D4)

[Cyb3rb1ade]

Summary

Implement phases D1–D4 of the universal audio speaker attribution / diarization pipeline.

D1 — SpeakerSegment Schema / Discord Attribution

• Canonical SpeakerSegment type with source, speakerLabel, speakerDisplayName, timestamps, attributionSource, diarizationModel, asrModel.
• Discord voice per-user streams emit discord:<userId> segments with attributionSource: "discord_user_stream".
• Formatter supports speaker-wise output and plain-text fallback.

D2 — Batch / Non-Realtime Diarization

• AudioDiarizationConfig and SQLite-backed job queue/cache.
• Non-blocking hook in applyMediaUnderstanding: attaches a fallback segment and enqueues an async job.
• Providers: none, mock, sortformer stub.
• ASR transcripts are saved with a fallback speaker until diarization finishes.

D3 — ASR/Diarization Merge

• Separate merge_results cache keyed by audioHash + diarization provider/model/version + asrModel + asrTextHash + asrWordsHash + mergeConfigVersion.
• AudioDiarizationJob carries targetRef (source, sourceId, mediaOutputId, transcriptId, conversationId, messageId) and ASR text/words.
• Local merge only: word-timestamp assignment or proportional text split, overlap strategy, gap/short-segment smoothing.
• Idempotent enrichment via mediaOutputId; a D2 job stays completed even if D3 merge fails.

D4 — Speaker Naming / Manual Mapping / Contextual Proposals

• OpenClaw speaker_mappings table with manual / contextual_proposal / enrollment (reserved) sources.
• formatSpeakerSegments renders speakerDisplayName when a confirmed mapping exists.
• Hidden mediaOutputId token in formatted transcripts so PLUR1BUS can correlate.
• PLUR1BUS speaker-mapping-store bridges OpenClaw's diarization SQLite DB.
• PLUR1BUS speaker-proposer suggests names from direct address (German/English) and keeps proposals unconfirmed.
• /speaker chat commands: list, name <label> <name>, proposals, confirm <label>, reject <label>, clear <label>.
• agent_end pipeline extracts mediaOutputId tokens and persists new proposals.

Test Coverage

• src/media-understanding/diarization/db.test.ts
• src/media-understanding/diarization/diarization.test.ts
• src/media-understanding/diarization/mapping.test.ts
• src/media-understanding/apply.test.ts
• packages/media-understanding-common/src/format.test.ts
• extensions/discord/src/voice/prompt.test.ts
• PLUR1BUS: tests/speaker-segment-schema.test.js, tests/speaker-mapping-store.test.js, tests/speaker-proposer.test.js, tests/speaker-mapping-commands.test.js

Verification

• OpenClaw relevant Vitest suites: 6 files / 85 tests passed
• PLUR1BUS npm test: 1889 passed / 0 failed / 1 skipped

What is NOT included

• No D5 enrollment / voiceprints.
• No D6 streaming diarization.
• No D7 video-enhanced attribution.
• No automatic Christian-/Eva-style person recognition.
• pnpm typecheck is not defined in this repo; type safety is covered by Vitest runs.