fix: write single-note spanner starts before their stops

[webern]

Human Summary

When a spanner starts and stops on the same note, the start needs to be emitted before the stop, but mx had not considered single note spanners.

This looks like a solid fix for, at least the known and cited cases from #429. I will not close the issue so we can track future cases of out of order spanner start and stop problems.

The test file submitted in the issue has been added to the corpus and passes round trip now.

Summary

NotationsWriter wrote every spanner stop before every start. That is right for a note that chains two spanners (one ends where the next begins, #139), but backwards for a spanner contained in one note: a nested tuplet covering exactly one note, or a glissando or slide drawn from a note toward a rest, came out stop-then-start, so a reader saw a stop for a span that was never open followed by a start that never closes.

Tuplets are decided locally in NotationsWriter: matching numberLevels on one note always mean a single-note tuplet (two different tuplets cannot share a note, because a note carries only one time-modification), so each start is followed by its same-note stop and the remaining stops come last. The nested case now comes out as Finale writes it: start 2, stop 2, stop 1.

Glissando, slide, and wavy line cannot be decided locally: a chained note can carry the same number as a single-note span, and only the stream of earlier events tells them apart. The SpannerResolver's existing serialized walk now also records same-note pairs (a stop that closes nothing open pairs with a same-bucket start on its own note) and exposes them through sameNoteSpanPartner; the writer holds those stops back until right after their starts. Chained spanners keep the stop-before-start order of #139, covered by the existing GlissandoApiTest chain test.

The reporter's fixtures from the issue are added to the corpus as data/rpatters1/gliss_to_rest-ref.xml and tuplet_singletons-ref.xml (Finale's exports, both with the start-then-stop shape). The gliss fixture now survives the strict api roundtrip and is pinned in roundtrip-baseline.txt. The tuplet fixture is not pinned: its ordering is fixed, but the writer restates tuplet-actual/tuplet-normal on a start whose source omits them, which is a separate fidelity gap.

Testing

• New *_SingleNoteSpanner tests fail before the fix (7 of 8 cases, all start/stop inversions), pass after (36 assertions in 8 test cases)
• Existing chain test (stop before start on a chained note, same explicit number) still passes
• Full api suite passes (6411 assertions in 593 test cases) and all three examples run
• api-roundtrip regression gate passes, 409 of 409 pinned including the newly pinned gliss_to_rest-ref.xml
• core roundtrip passes with the corpus additions (840 files, pinned count updated)

References

• Improves NotationsWriter emits all spanner stops before all starts, breaking single-note spanners #429
• The ordering exception is carved out of Question about ordering of starts and stops for e.g. slurs #139's stop-before-start convention
• Related: NoteWriter infers <normal-type> from sibling notes instead of using DurationData::timeModificationNormalType #428 (accidental-mark half of the same nested-tuplet territory, independent), feat: unified spanner numbering with writer-side assignment #320 (tuplets still carry a raw numberLevel int rather than SpannerNumber; unchanged here)