[refactor] accounts tech debt: structured stance, sectioned briefing generation, speaker↔person binding UI

[Lanznx]

Three pieces of tech debt in accounts (the mini-CRM). All of them are the "the mechanism is there, but the quality ceiling is low" kind.

Scope

1. Structured stance output (replacing the keyword heuristic): stanceValueFromText in store.ts is a zh/en keyword regex with a poor hit rate → a person's stance dot is grey most of the time. The extraction schema (extract.ts) should have the LLM emit a structured {value: support|neutral|oppose, confidence} directly for stance-type claims, and the projection cache should read that structured field instead.
2. Sectioned briefing generation: once the claim library gets large, generating in a single pass blows the context budget (briefing.ts feeds in every claim at once). Generate per category/thread and then merge.
3. Live speaker↔person binding UI: the HistoryEntry.attendees[].speakerKey field exists, but there's no UI for binding a transcript speaker to a person. Wiring up "Speaker 2 = Director Wang" during the meeting is what gives the subjects from post-meeting extraction a reliable anchor.

Also of note

• The red-line guard (auto-injecting a critical eval when a meeting is linked to a company) is implemented but has never been tested on a real call. Report back on this issue once it has been verified.
• Conflict detection currently rests entirely on the LLM prompt; we're still watching how well it holds up.