[webern]

AI Summary

musuite/testVoiceMapper1.xml writes three notes labelled voice 4, 5 and 4 one after another, with no backup between them. mx's reader only started a new voice at a backup, so all three notes went into one voice with one name, and the middle note came back as voice 4.

Reader: between two backups, each staff's notes are now split into voices by the text of their voice element. A chord member stays with the note it sounds with, and a note without a voice element stays with the note before it. A stretch that names at most one voice is one voice, as before, and the same name after a backup is still a separate voice.

Writer: when a voice only plays in the rests of the voice before it on the same staff, it is written in those rests with its own voice text, instead of after a forward and a backup. This happens only when the notes of the voices form one line in time with no two voices starting on the same tick, each voice first appears in voice order, every voice on the staff has its own voice text, and every note has a position. Otherwise voices are written one after another as before, and a voice that starts after the one before it ends is written as before. A staff with one voice is written exactly as before.

The writer rule also applies to scores built by hand through mx::api, so their output can change. The new output is an equivalent encoding that reads back to the same voices. The spanner number resolver now walks notes in the order the writer writes them, so slurs and tuplets given an identity get different numbers when they overlap in the shared stream.

Four other corpus files are grouped differently by the reader: Berlioz Le Corsaire, Black Note Study Op 10 no 5, Moments Musicaux Op16 No4 and Nocturne No. 20 in C Minor. Their written output differs from main only in voice text, and every change brings back the source's text (326 notes in Berlioz, one in each of the others). None of them has a voice written in another voice's rests.

Testing

• VoiceInterleaveApiTest: 15 test cases, 2549 assertions. One test writes and reads back 400 random one-staff measures of one to three voices with chords, grace notes and identity slurs, and checks the voices come back unchanged and no slur number is reused while open
• The slur, repeated-name and random tests fail before the follow-up fix (a write with an identity slur in a merged voice failed outright)
• make api-test: 9173 assertions in 799 test cases
• make api-roundtrip: 459 of 459 pinned files pass, with musuite/testVoiceMapper1.xml added to the baseline
• make api-roundtrip-discover: 460 pass, 381 fail (main: 459 pass); the only new pass is testVoiceMapper1.xml and no file stops passing
• Unity build of mx (batch size 0) and make wasm-test
• make fmt-check

Not covered:

• A cross-staff beam group in a voice whose rests another voice fills. The reader finds cross-staff beam groups per backup stretch, so a group broken by another voice's note is treated like one broken by a forward, as on main.
• Two voices with the same name on one staff that meet without a backup are still read as one voice. The writer does not merge voices on such a staff.

References

• Closes Possible problem with backup multi-voice logic found with musuite/testVoiceMapper1.xml. #425
• Related: feat: preserve source voice names in mx::api #422 (voice names)