Memory: end-of-session tidy pass (vague notes -> precise)

Approved (C5) but deferred from the memory-polish batch — it's an LLM call + a dedup design and deserves its own careful pass.

**Problem.** The deterministic writer stores a correction as the user's raw words: `for rent totals — actually, for rent totals count HOUSING and mortgage as rent too`. A capable model applies that fine; on the `gemma4` floor it took the schema-level case-sensitivity flag (#46-era work) *plus* the note to get the right query. A crisper note — `rent → lower(cat) IN ('rent','housing','mortgage')` — would carry better.

**Shape.**
- One **optional** LLM call at end of a session (not per turn), gated behind a setting, `FELLA_MEMORY` still the master off switch.
- Input: the session's episode log + the current `memory.md`. Output: a tidied `memory.md` — vague correction notes rewritten as precise mappings where the categories/columns are unambiguous, near-duplicate recipes folded, recipes unused in N sessions dropped.
- Never invents a fact; only rewrites/removes what's already there. The user can always see the diff (it's a plain file) and revert.
- Measure on `agent_eval memory` cross-session: does the tidied note close the `gemma4` gap without a `--compare` regression elsewhere?

Design notes in `docs/FOLDER-MEMORY.md` §"Writing" and §Implementation (deferred list).

Start with docs/FOLDER-MEMORY.md, especially the “Writing” and “Implementation” sections and the deferred list. Then inspect the existing session episode log, memory.md handling, and the `agent_eval memory` entry point. Done means an optional end-of-session tidy pass is gated by a setting, preserves existing facts, and improves the gemma4 cross-session result without a `--compare` regression.

Written by the indexing model from the issue text.