Topic summary architecture — cost efficiency and freshness

[AndreRobitaille]

Problem

The current topic summary model generates one TopicSummary per topic-per-meeting during SummarizeMeetingJob. This creates several problems:

1. Waste — We pay for N LLM calls per topic but only display the most recent summary. All prior summaries are buried.
2. Narrow view — The visible summary only reflects what happened at the last meeting, not the full arc of the topic across months of discussion.
3. Staleness — The summary becomes outdated as soon as the next meeting happens. There's no mechanism to refresh it.
4. No narrative — A topic discussed across 8 meetings should tell the story of how it evolved. Instead, residents just see a snapshot of meeting Bump turbo-rails from 2.0.20 to 2.0.21 #8.

Current Architecture

SummarizeMeetingJob → for each topic in meeting:
  → Ai::OpenAiService#analyze_topic_summary (reasoning model)
  → Ai::OpenAiService#render_topic_summary (reasoning model)
  → TopicSummary.create!(topic:, meeting:, content:)

Topic show page displays: topic.topic_summaries.order(created_at: :desc).first

Possible Directions

• Rolling summary per topic — One summary record per topic, updated incrementally when new meeting data arrives. Cheaper (update vs. full generation), always reflects the full picture.
• Periodic regeneration — Regenerate the topic summary on a schedule (weekly?) using all available meeting data. Simpler than incremental updates.
• Layered approach — Keep per-meeting summaries for the meeting page, but generate a separate "topic overview" summary that synthesizes across all meetings.

Each has trade-offs around cost, freshness, and complexity. Needs design work before implementation.

Constraints

• Must align with TOPIC_GOVERNANCE.md (fact/framing/sentiment separation in generation, even if not in display)
• Must be cost-conscious — reasoning model calls are expensive
• Must not fabricate continuity that doesn't exist in the source documents

[AndreRobitaille]

Closing — the problems described here (waste, narrow view, staleness, no narrative) were solved by the TopicBriefing feature. TopicBriefing is a rolling briefing per topic (one record, updated in place) with a three-tier generation pipeline (headline_only → interim → full). This is the "rolling summary per topic" option from this issue. Remaining gap: Tier 2 (interim) trigger is tracked separately in #66.