Clarify the lossless recall prompt and tool metadata so agents treat
lcm_expand_query(query=...) the same way as lcm_grep full-text search.
Add explicit guidance that FTS5 defaults to AND matching, that query
strings should stay short and distinctive, and that prompt carries the
natural-language question while query is only for recall matching.

Regeneration-Prompt: |
  Improve the system prompt guidance for lcm_expand_query in the
  lossless-claw codebase. Agents were over-stuffing query with lots of
  keywords because they did not realize lcm_expand_query(query=...) runs
  through the same FTS5 full-text search path as lcm_grep, where terms
  are ANDed by default and extra words can force zero matches.

  Update the static recall policy and any session-specific compacted
  context guidance so it explicitly says query follows the same FTS5
  rules as lcm_grep, that more terms make matching stricter rather than
  broader, that query should usually be 1-3 distinctive terms or a
  quoted phrase without synonym padding, and that prompt is where the
  natural-language answer request belongs.

  Tighten the lcm_grep and lcm_expand_query tool description/schema text
  so query construction is harder to misuse, add regression tests for
  the prompt and metadata strings, and include a patch changeset because
  this changes user-visible tool guidance.