fix(search): show similarity, not fused rank, as match

[chiruu12]

Part of #150, taking the two Python surfaces. The OpenClaw and shared-core TypeScript renderers and the hook pair are not in this PR.

Presentation only, as you asked. score and raw_score on the REST payload are untouched.

The three cases are decided in one place, palinode/core/scoring.py, rather than once per surface. You flagged the duplicated jq as a wart you did not want extended; two Python renderers making the same judgement separately is the same wart in a different language, so MCP and the CLI now both call describe_match.

Result	Before	After
cosine 0.421, fused 1.0	100% match	42% match
raw_score: null (BM25-only)	100% match	keyword match, rank 1.00
raw_score absent (pre-0.12)	100% match	rank 1.00

I kept the null and absent cases apart, as you asked. A null means the ranker told us this hit had no vector arm, so the surface says which arm did fire. An absent field means the server never sent one and the arm is unknown, so the rank is all it says. Neither claims a similarity.

On the naming: rank, used everywhere. For the BM25-only case I went with a word rather than silence. keyword match says what actually matched, and dropping the number entirely loses the ordering a reader is using the flag to see. It reads as a match without ever reading as a similarity, which was the property you wanted.

Validation

• pytest -q: 3263 passed, 10 skipped, 5 xfailed
• ruff check: clean
• The three surface tests fail on unfixed renderers and pass on the fix. Reverting just mcp.py and search.py gives [1.00] notes/a.md from the CLI and (100% match) from MCP.

The tests use the measured numbers from the issue: 0.421 as the median of the 32 searches, and each of the three points across the 0.402 to 0.459 range asserted to render under half.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[Paul-Kyle]

Merging. The shared helper is the right shape here: MCP and CLI make the same three-way decision once, numeric raw cosine is the only value presented as match percentage, present-null is labelled as keyword-only plus rank, and an absent legacy field is rank-only. I re-read the refreshed four-file diff at 5d591f3 after the branch update and all 9 required checks are green. This remains correctly partial, so #150 stays open for the TypeScript and hook renderers.