feat: --dump-segments, so a lyrics edit doesn't re-run the ASR

[ijuinryukichi]

--segments could read a segments file, but nothing could write one — -f json
emits aligned lines, and Segment had from_dict without its mirror. The cheap
path existed and was unreachable from the CLI.

What this adds

--dump-segments FILE writes the transcription in exactly the shape --segments
reads back, and Segment.to_dict / Word.to_dict to go with it.

lyric-align song.wav lyrics.txt --separate --dump-segments segs.json -o out.lrc
# fix a typo in lyrics.txt
lyric-align --segments segs.json lyrics.txt -o out.lrc      # seconds, not minutes

The write lands before alignment, not after: the expensive half is done, the
cheap half is the one you re-run, and a run that dies in the cheap half must not
cost the audio work. That is the case #28 came from. When --segments was the
input it no-ops with a note rather than copying the input under a second name.

Measured

Real end-to-end on the public-domain demo (tiny model, no Demucs):

	wall clock
transcribe + align + dump	7.02 s
re-run off the dump	0.04 s

Output byte-identical between the two. With medium and --separate the first
number is minutes.

The file is one segment per line instead of a plain indent=, which puts every
word timing on three lines of its own — 443 lines vs 8 for a 6-segment song.
Floats are left exactly as the ASR gave them; rounding would make the documented
from_dict(to_dict(x)) == x a lie for a cosmetic gain.

Tests

Six new tests (3 model, 3+1 CLI). Five mutations run against them, each caught by
the test that claims to guard it:

mutation	caught by
to_dict drops word timings	round-trip, word-timing, CLI word-timing (3)
dump moved after alignment	the crash-survival test (+1)
no-op guard removed	the --segments no-op test
hand-written brackets broken	every test that parses the file (4)
words re-expanded with indent=	the one-line-per-segment test

Full suite: 67 passed on 3.9 and 3.14.

Version bumped to 0.4.0, matching how #6 handled a feature. Publishing is a
separate decision.

Closes #28

🤖 Generated with Claude Code