feat: pick pairing from the ASR's segmentation instead of hardcoding 2

[ijuinryukichi]

pairing — how many lyric lines make one matching unit — was fixed at 2. That
was never a property of songs. It is a property of faster-whisper medium,
which merges about two sung lines into one segment on Japanese rap. Nothing in
the code or docs said so.

large-v3 segments far finer on the same track: 41 → 64 segments, mean segment
4.73 s → 2.78 s. A two-line unit then straddles a boundary.

model	segments	pairing	mean	≤0.5 s	worst	lines placed
medium	41	2	0.50 s	14/20	3.58 s	74/76
large-v3	64	1	0.31 s	15/20	0.76 s	49/76
large-v3	64	2	1.13 s	10/20	3.70 s	68/76

So upgrading the model alone makes things worse — and with the pairing it
deserves, large-v3 removes the 3.6 s outlier that five separate matcher
changes failed to fix earlier today (#13). Worst case 3.58 s → 0.76 s, mean down
39 %, outliers 1 → 0. The leverage was in the input, not the matcher.

The estimator

Lines-per-segment is exactly what pairing means, so it is also the estimate.
No audio, no ground truth, no new dependency — just the two counts we already
have.

	lines / segments	auto	vs. fixed 2
過ぎたるもの, medium	76 / 41	2	identical
過ぎたるもの, large-v3	76 / 64	1	0.50 s → 0.31 s
黒砂, medium	80 / 46	2	identical
黒砂, large-v3	80 / 46	2	identical
過ぎたるもの, full mix (ASR collapsed)	76 / 11	3 (capped)	no worse

Capped at 3 because past that the ASR has stopped producing line-like segments,
and the answer is to separate the vocal — which the CLI already tells you. Left
uncapped that case estimates 7 and is markedly worse.

What it costs, stated in the README

At pairing 1 there are only 64 segments for 76 lines, so a quarter cannot
place: 49/76 against medium's 74/76. Both large-v3 rows match the same 18 of
the 20 measured lines, and the extra lines pairing 2 places are ones the ground
truth cannot check — while that configuration also produces six measurable
outliers against pairing 1's none. medium stays the default model.

What it does not do

It does not rescue the repeated-hook track, which was predicted before the run:
large-v3 trails medium there at every pairing, because four identical hook
lines carry no information about which repetition they are, whatever transcribes
them. Better ASR fixes outliers made of garbled text, not outliers made of
identical text.

Default changed before the first PyPI release, deliberately. 7 tests added, 57
pass.

🤖 Generated with Claude Code