docs: variable unit size places more lines, and some of them badly

[ijuinryukichi]

The next idea after --pairing auto: stop fixing the unit size up front and choose (lines, segment) jointly at each step. pairing is a rounded average, so it is wrong for part of any track — at 76 lines over 64 segments the true ratio is 1.19, and a fixed 1 leaves 27 lines unplaced while a fixed 2 straddles boundaries.

Measured worse. Documentation only, no behaviour change.

On the shipped metric it looks free

unit size	lines placed	mean |err|	within 0.5 s	worst	2nd lines outside their GT window
fixed, from the ASR (shipped)	49/76	0.31 s	15/20	0.8 s	0 of 16
variable, k ≤ 2	66/76	0.42 s	14/20	2.0 s	2 of 18
variable, k ≤ 2, only if it wins by 0.1	66/76	0.32 s	15/20	0.8 s	1 of 18 (by 13.3 s)

The last column is the point. The ground truth is marked per two-line pair and the metric scores only the first line of each — which is exactly where the extra placements do not land. Second lines have a free check: each must start inside its pair's window.

What the 17 "free" placements bought

The last verse line scores 0.000 against the hook segment on its own and 0.286 once the following hook line is absorbed into the same unit — over the 0.25 threshold. The breath split then drags that verse line 13.3 s forward into the hook, and displaces the hook line that had been placed correctly.

A unit chosen to maximise similarity will straddle a section boundary, and such a unit needs only its tail to match. pairing=1 cannot do this because a one-line unit has no tail. So variable pairing manufactures the very tail-match pathology that the dedicated veto in Known limits already failed to fix.

Of the two extra placements that can be checked, one is right and one is 13.3 s wrong.

It fails from the other side too

On a track where the ASR merges two lines consistently, deviating downward orphans the remaining line onto the next segment and shifts every later unit's phase: within-0.5 s 26/33 → 18/33, and the 4×-repeated hook lands a repetition early.

Restricting deviation to upward only appears to fix that — but only because pairing=2 with k ≤ 2 leaves upward no room. Allowing k ≤ 3 breaks the same track again (26/33 → 13/33).

Why the premise was wrong

The four earlier attempts all changed which segment a unit selects, so the plan here was to change how much a unit consumes and dodge that coupling. It does not dodge it: consumption sets how fast the segment cursor advances relative to the line cursor, so it moves idx as well — one step removed, same gains-and-losses-in-adjacent-pairs behaviour.

Also disclosed

The Accuracy section now states that the headline figures score first lines only, and that the shipped configuration passes the second-line containment check (19/19 and 29/33, the four misses being the already-documented repeated hook). The tables were not hiding a second failure mode — but the distinction decides any change judged on placement count.

Verification

57 tests pass. Harness in toryu-web/scratchpad/lyrics_align_bench: exp_var_pairing.py, exp_var_diag.py, exp_var_rows.py, exp_var_margin.py, exp_var_updown.py, exp_second_line.py. Every sweep endpoint (m=inf) reduces exactly to the shipped baseline on all four (track, model) configs — the first version did not, and that bug made the whole first sweep unreadable.