Retrospective save: write an already-gathered research context to disk after the fact

Today the save flow only works if the user asks up front: SKILL.md redirects the engine's stdout to `~/.last30flames//-.md` at run time. If the user only realizes a run was valuable after reading the brief ("actually, save that"), there is no documented path.

But the numbered research context is still in the agent's conversation after a run, so retrospective save is just a SKILL.md instruction change - no engine work:

- On a retrospective save request, write the research context the agent already read (verbatim, not the brief) to the normal thread folder, picking ``/`` exactly as in the existing save flow, and report the path.
- Caveat to document: if the conversation has since been compacted/summarized and the verbatim context is gone, the honest move is to re-run the engine fresh rather than reconstruct it from memory.

No new sources, no keys, no engine changes - fits all design constraints.

Start with SKILL.md and compare the retrospective-save instruction with the existing run-time save flow, including its thread, slug, date, and normal folder conventions. Done means the documented flow saves the already-read research context verbatim, reports the path, and tells users to re-run when compaction has removed that context.

Written by the indexing model from the issue text.