test: pin the match threshold from both sides

[ijuinryukichi]

The tests looked comprehensive. Mutation testing says they were not covering the one contract this project is built on.

What was open

test_unmatched_line_is_honest passes window=1, so it withholds the segment entirely — it exercises the forward window, not the score. Deleting the threshold check that implements the honest-gap contract left it green.

Two mutations against main, 57 tests:

mutation	tests killed	the test named for the contract
CJK_THRESHOLD 0.25 → 0.60 (2.4x)	1 (test_latin_default_rejects_chance_similarity)	passed
drop the best_score > threshold gate	1 (test_interpolate_fills_gaps)	passed

Both survivors were indirect. Nothing was holding the decision boundary on purpose.

What this adds

Two tests, pinning that boundary from opposite sides.

test_a_reachable_segment_below_the_threshold_stays_unmatched — puts a candidate squarely in reach and lets the score alone reject it, so the gate is guarded rather than the window standing in for it.

test_the_cjk_floor_still_accepts_a_badly_heard_true_match — real ASR output for a real line, and the reason the floor sits at 0.25: 九月十五朝霧開戦 / 黒田細川矢刃交わし合戦 was heard as 9月15 朝霧開戦 来るだ 遅かえば 可視化せん and scores 0.256. The model got the moment right and still clears the floor by a hair. Raise the floor and this true match is lost — which is the evidence the README's 0.25 rests on, now enforced.

Verification

Each new test fails under its corresponding mutation — that is the point of adding them, and it was checked rather than assumed:

CJK floor 0.25 -> 0.60          2 failed (incl. cjk_floor_still_accepts_a_badly_heard_true_match)
threshold gate removed          2 failed (incl. reachable_segment_below_the_threshold_stays_unmatched)
clean                          59 passed

No source changes. Clear __pycache__ between mutation runs — an unchecked hash-based .pyc will keep a reverted constant alive and make git status look clean while the suite still fails.

🤖 Generated with Claude Code