fix: keep WebVTT cue text around inline markup

[Kevin Turcios (KRRT7)]

Summary

parse_voice_tags (used by ingest_vtt_transcript and get_transcript_speakers) lost or garbled caption text whenever a cue contained WebVTT markup:

Cue raw_text	Before	After
<v Alice>I <i>really</i> mean it</v>	('Alice', 'I')	('Alice', 'I really mean it')
<v Bob>Use the <c.code>make</c> command	('Bob', 'Use the')	('Bob', 'Use the make command')
I <i>really</i> mean it (no voice tag)	(None, 'I <i>really</i> mean it')	(None, 'I really mean it')
<v.loud Alice>Hi <b>there</b></v>	(None, '<v.loud Alice>Hi <b>there</b></v>')	('Alice', 'Hi there')
Intro <v Bob>hello</v>	('Bob', 'hello')	(None, 'Intro'), ('Bob', 'hello')

• Each voice segment now runs until the next voice tag, and the remaining cue markup (</v>, <i>, <b>, <c.class>, timestamp tags) is stripped while its text is kept, matching what webvtt's caption.text does.
• Voice tags with classes (<v.class Speaker>) are recognized, as webvtt's own voice pattern allows.
• Cues without markup are unchanged: old and new output are identical on all 908 cues in tests/testdata/*.vtt.

Validation

• New parametrized parse_voice_tags tests (the table above, plus timestamp tags and a plain multi-voice regression case), and a test that parses a real .vtt file through webvtt.read. All markup cases fail without the fix.
• make format check test

[Bernhard Merkle (bmerkle)]

LGTM — verified locally.

• The new tests: 7 cases fail against main's transcript_ingest.py and pass with the fix; full make is green (820 passed).
• Differential against webvtt: for all 908 cues in tests/testdata/*.vtt the joined segment text equals caption.text (whitespace-normalized), and the same holds for <v> without annotation, <V>, multi-word speakers, tab after v, multi-line cues, and <lang>/<ruby> spans.
• The second commit's _VOICE_TAG_PATTERN is linear; CodeQL alert #43 shows as fixed.

Non-blocking suggestions:

1. _CUE_TAG_PATTERN = r"<[^>]*>" is quadratic on cue text with many unclosed < (each < scans to the end): about 0.09 s at 20k chars and 2 s at 100k. Not realistic for well-formed VTT, but transcripts are external input. r"<[^<>]*>" makes it linear and strips identically on everything above.
2. Text after </v> and before the next voice tag is attributed to the preceding speaker: <v Alice>a</v> mid <v Bob>b</v> → ('Alice', 'a mid'). Strictly it's outside any voice span, so (None, 'mid') would be more faithful, and consistent with how this PR now treats text before the first voice tag. Rare in practice; fine as a follow-up or a docstring note.

(FYI, unchanged by this PR: character references such as &/< stay undecoded, same as webvtt's caption.text; html.unescape after stripping tags would be safe if you ever want plain text in the index.)

[Kevin Turcios (KRRT7)]

Thanks! Took suggestion 1: _CUE_TAG_PATTERN is now <[^<>]*> (7885b38). Besides being linear (50k <: 0.76 s → 0.001 s), it also stops eating text after a stray <: a < b <i>c</i> used to strip to a c, now a < b c. Added that as a test case. Merged the latest main; make format check test is green.

I'll leave the text-after-</v> attribution for a follow-up.