feat: summary quality — speaker-aware input, grounding/citeability, structured decisions & action items

[fmasi]

Why this matters more than it looks

Summaries are becoming the retrieval index for the agent-queryable context layer (#152): agents scan cheap summaries to decide which full transcripts to open. That makes summary quality load-bearing — a weak summary sends the agent to the wrong transcript. This is not cosmetic polish.

What to verify / fix

• Speaker-aware input — confirm the summariser is fed the speaker-attributed transcript, not a flat text blob. Without it, the summary can't express "who committed to what" and we throw away our best signal (diarization). Likely the highest-value change.
• Grounding / citeability — summary claims should cite back to transcript timestamps. For a canonical-truth product, a hallucinated decision in the summary is a provenance violation; the signed transcript is truth, the summary is derived and can drift.
• Structured output — extract owner-attributed decisions and action items (who owns what), not just prose. More useful to an agent and to a human glancing at the map.
• Accuracy eval — a small labelled set to measure hallucination / omission, across the local and remote summary paths.

Relates to #37 (on-device summary provider), #36 (map-reduce long transcripts), #152 (summaries-as-index).