feat(workspace): topic clustering and generated TOC/SKILL docs

[RichardHightower]

The last open item — and the only one that adds product surface rather than repairing something.

Flagged before building: its own ticket calls it a deferred go/no-go, and the approved plan scoped it out. The goal said all 24, so it is built. The concern is on the record, not in the way.

What it does

Synthesize reads every note, summarizes each, clusters them by topic, and writes TOC.md and SKILL.md back into the workspace.

This connects the last three orphans. TopicRefiner, TOCGenerator and SkillGenerator were written months ago, tested, and completely unreachable — nothing imported them, and they called Bun.spawn so they could not have run in either shipped mode regardless.

The client-bundle guard now walks 23 modules instead of 19 and passes. That increment is the proof they are genuinely reachable and genuinely browser-safe, not merely edited.

Three things it deliberately does not do

• Read its own output. A second run would otherwise summarize the index it wrote on the first, and topic labels would drift toward describing the index rather than the notes.
• Process an unbounded workspace. One LLM round trip per note; capped at 40.
• Truncate silently. When it skips notes it says so — a silent cap reads as "covered everything" when it did not.

Testing

Six unit tests against an in-memory workspace (LLM stubbed in code) and two E2E specs that stub only the /api/llm network boundary, so everything above it is real. The second asserts the failure path: a synthesis that dies quietly is worse than one that refuses.

🤖 Generated with Claude Code

https://claude.ai/code/session_01TYuTJt2hh7k8mg8FXCKrw1