Relevance-scoped retrieval from memory.md as it grows

From the same context-engineering principle as the schema JIT issue, applied to per-folder memory (#43): "retrieve notes when relevant, not everything." Fella's memory recipes are already the "record & replay" pattern from agent-reliability research — a query that passed self-check becomes reusable guidance — the missing piece is *relevance-scoped* replay.

**Mechanism.** Once `memory.md` has enough entries to matter, filter what's included in a given turn's prompt to the recipes/notes whose associated tables or columns overlap with the current question's vocabulary (the same column-name matching the schema digest already does), instead of including the whole file every time. Small memory files (the common case today) see no change.

**Not urgent as filed** — worth checking whether `memory.md` is even large enough yet for this to matter before investing; file now so it's tracked for when it does.

Related: #43, #53 (the other open memory issue — a different angle: tidying content rather than scoping retrieval).

Start by checking how per-folder memory.md is loaded and where the existing schema digest performs column-name matching. First verify whether current memory files are large enough to justify filtering; done means larger files replay only notes whose tables or columns overlap the question, while small files behave unchanged.

Written by the indexing model from the issue text.