fix: narrow an oversized octave shift to the closest ottava line

[webern]

Human Summary

Some corpus files existed with bad octave-shift size values. lysuite/ly33e_Spanners_OctaveShifts_InvalidSize.xml was apparently created for this purpose. synthetic/octave-shift.3.0.xml and synthetic/octave-shift.3.1.xml were created by an LLM that might not have known better (just going off of the schema).

Either way, the right thing to do is to test the expected behavior of mx which is to coerce these to the next lower, valid, values of 22, 15 or 8.

Summary

MusicXML lets octave-shift/@size be any positive integer, and three corpus files use that room for a number that names no line a performer could read: 27, 11 and 1 (lysuite/ly33e_Spanners_OctaveShifts_InvalidSize.xml is named for it). mx::api narrows each of those to an ottava it can draw, so the round-trip saw a 27 go in and a different number come out.

Narrowing is the behavior we want — mx::api models the six ottava lines music notation has, and everything downstream is allowed to assume an ottava is one of them. Two changes, one to the rule and one to the round-trip.

The rule now picks the closest line. DirectionReader::parseOctaveShift matched 22 exactly, so a source asking for 27 fell through to the 15th line even though the 22nd was both available and closer to what it asked for. It matches 22 and up instead. Sizes of exactly 8, 15 and 22 are unaffected, so this only moves sources that were already asking for something notation has no name for.

The round-trip expects the narrowed size. narrowOctaveShiftSizes() in CorpusRoundtripMain.cpp rewrites octave-shift/@size on the expected document by that same rule, rather than the api learning to echo the original number back out. That is the move the harness already makes for mx's attribution stamp: add it to the expected side instead of stripping it from the output.

• Only the expected side is narrowed, so a write that produced a size other than 8, 15 or 22 still fails.
• A stop's size is taken there from the stop's own attribute, while mx takes it from the start the stop closes. A source whose stop contradicts its start still fails, which is what OttavaStopSizeContradictionIsNormalizedToTheStart covers from the api side.
• The api already round-tripped the attribute's presence — OttavaStart::writeDefaultSize and OttavaStop::writeSize from feat: derive octave-shift stop size from its start #415 — so only the value ever differed. No public api change.

Round-trip corpus

Adds 3 files to roundtrip-baseline.txt (407 -> 410). value:octave-shift@size was the sixth entry on the classifier worklist.

Testing

Three new OttavaSizeApiTest cases state the narrowing away from the corpus: a size between the lines taking the 15th, a size past the 22nd taking the 22nd, and a size below 8 becoming a plain octave line.

The gates below were run against the first commit. The second commit (the >= 22 rule) landed after, at the author's direction, and is covered by CI on this PR.

• make api-test (6348 assertions in 588 test cases)
• make api-roundtrip (410 passed, 0 failed of 410 pinned)
• make api-roundtrip-discover: the PASS set is exactly the 410 pinned files — the 3 new passes and no regressions
• make core-roundtrip-test (839 test cases) and make core-unit (221 assertions in 44 test cases)
• make fmt and make fmt-check

References

• Progresses the round-trip corpus work tracked in api: expand corpus coverage of the api round-trip pass-list #208
• Builds on feat: derive octave-shift stop size from its start #415, which introduced the ottava stop-size resolver and the size-presence fields
• Last of six stacked PRs; stacked on feat: preserve source voice names in mx::api #422, feat: round-trip a redundant staves element in mx::api #421, feat: support the direction directive attribute in mx::api #420, feat: expose the identification source in mx::api #419, feat: model note playback velocity in mx::api #417. The base branch is feat: preserve source voice names in mx::api #422's branch, so this diff shows only the octave-shift change.