GP vocal import drops lyrics when the first bar voice is silent #103

Target

lib/gp2rs_gpx.py, convert_vocal_track(), multi-voice loop around lines 2290–2360 (commit 6246269).

Category

Workflow / data-loss regression

Flaw analysis

The new unconditional break runs after processing the first valid GP voice object in a bar, regardless of whether that voice emitted any lyric token. A common GP layout can place a silent/alternate-rhythm voice first and the actual lyric-bearing voice second. In that case the converter drops every lyric in the second voice.

Minimal reproduction against current HEAD:

• one 4/4 master bar
• <Bar><Voices>silent lyric</Voices></Bar>
• silent contains a quarter-note beat with no <Lyrics>
• lyric contains a quarter-note beat with <Lyrics><Line>hello</Line></Lyrics> and a pitched note

convert_vocal_track() returns:

<?xml version="1.0" ?>
<vocals count="0"/>

The targeted parser suite currently passes because tests/test_gp2rs_gpx.py tests convert_vocal_track_to_pitch_sidecar() but has no multi-voice tests for convert_vocal_track().

There is also a consistency gap: convert_vocal_track_to_pitch_sidecar() still walks all voices and leaves tie durations uncapped, so callers following its documented “call both” workflow can produce a lyric XML and pitch ribbon with different voice membership and sustain lengths.

Remediation

Select the primary vocal voice by content, not position: inspect voices for lyric-bearing beats and choose the first coherent lyric-bearing voice for that bar (or define a deterministic track-level voice policy). Apply the same voice-selection and tie-duration policy to the pitch-sidecar helper. Add regressions for:

1. first voice silent, second voice lyrical;
2. two simultaneous lyrical voices;
3. XML and pitch-sidecar timelines having matching selected onsets and capped durations.