Feedback-driven keyword refinement

[toejough]

Problem

When a memory is marked irrelevant, its effectiveness drops but its keywords don't change. It will keep surfacing in the same wrong contexts. The feedback loop measures impact but doesn't fix the root cause.

Desired behavior

When --irrelevant feedback is given:

1. Record the query context (tool name/input or user message)
2. Identify which keywords on the memory matched the query
3. If those keywords are generic (high corpus frequency), remove them
4. Memory stops surfacing in that context, retains relevance where it matters

Dependencies

Requires #343 (irrelevant_count tracking).

Priority

High effort, very high impact. The truly self-correcting loop.

See investigation doc for full analysis.