fix: rogue "s" next to roman numeral harmony labels

[FelipeDefensor]

Harmony labels in roman-numeral display mode carry a stray s inside the numeral: a first-inversion triad reads Is6 instead of I6, a third-inversion dominant seventh Vss42 instead of V42.

Reported against 0.6.4 while importing a harmony CSV; this branch is off v0.6.4.

Cause

_figs_to_str emits "s" for every figured-bass slot without an accidental:

prefix = "%" if len(figures) > 2 else ""
accs = "".join(_fmt_mod(mod, "s") for _, mod in figures)

That placeholder is only consumed by MusAnalysis inside the three-figure stack introduced by % — the encoding is positional, so each of the three accidental slots needs a filler (%sbs765). The one- and two-figure forms have no blank slot, so the character reaches the glyph run and is drawn as a literal letter.

Verified against the font: Is6 really renders I + s + superscript 6. Two-figure forms accept accidentals positionally without any filler (#b65 draws ♯6 over ♭5), so dropping the placeholder outside the stack is the whole fix.

Fix

Emit the placeholder only in the stacked form. The % path is unchanged — V%sss432 is identical before and after.

Repro bundle

The fixture commit sits before the fix, so "before" is one checkout back.

One-time setup (re-running can discard local edits — it aborts when the head branch is checked out in another worktree, and --force resets the branch):

gh pr checkout <this PR>

After (branch tip):

uv run --python 3.12 tilia "$PWD/repro/harmony-roman-s.tla"

Before:

git checkout HEAD~1

uv run --python 3.12 tilia "$PWD/repro/harmony-roman-s.tla"

Back to the tip:

git checkout -

Acceptance criteria

• Do: look at the Harmony timeline — nine harmonies at 0–80 s, all in roman-numeral display mode.
• Was: stray s inside the numerals — Is6, Iss64, Vss65, Vss43, Vss42, Vss65/V, iiss65.
• Should be: I6, I64, V65, V43, V42, V65/V, ii65. The I at 0 s and the stacked V%sss432 at 80 s are unchanged controls.

Both states were checked by actually checking out the two commits and opening that file, not inferred from the diff.

Fixture lifecycle

repro/harmony-roman-s.tla and its generator tests/repro_harmony_roman_s.py are throwaway — delete both before merge. The generator is not auto-collected (no test_ prefix); run it with uv run --python 3.12 pytest tests/repro_harmony_roman_s.py -q. Harmony is outside the CLI's timelines add / components coverage, so it builds state through commands.execute(...) and saves via file.save, as CLAUDE.md prescribes.

The durable artifact is TestRomanNumeralDisplay::test_roman_label_has_no_blank_accidental_placeholder, which pins I6 / I64 / ii7 / ii65 / V65 / V43 / V42, plus a companion test pinning V%sss432 so the stacked form stays covered.

Out of scope, found while investigating

• Applied chords carry figure accidentals: viio65/ii renders viio#b65/II. The numeral is spelled against the applied key but _chord_fb computes the figures against the home key.
• A two-figure group whose lower figure alone is chromatic cannot be expressed — MusAnalysis fills accidentals from the top and has no blank slot outside %. Unchanged by this PR (before it, that case also printed the literal s).
• mode._get_music21_object_from_text returns None when the text does not start with a note name, so a key row with an invalid symbol raises TypeError: cannot unpack non-iterable NoneType object and aborts the whole CSV import instead of reporting one error.

[FelipeDefensor]

Superseded by the combined 0.6.5 release branch, which carries this fix plus three changes that came out of review:

• the % stack is now taken only for its exact shape (three figures, single-character accidental slots), so the literal s no longer survives for four-figure groups or double accidentals — over a 10500-label sweep, malformed stacks 1728 → 0 and stray s 988 → 0;
• the regression tests gained accidental-bearing cases, which the original parametrize had none of;
• the throwaway repro bundle was removed, as the commit message intended.

The remaining defect in this area — accidentals being emitted as a block before the numbers, so they can be drawn against the wrong figure — is pre-existing at v0.6.4 and tracked separately in #598.