Investigate ACE-Step arrangement repetition before changing candidate ranker

[a-github-name]

Problem

In the controlled 30-second ACE-Step comparison, the Base and SFT renders were perceived as repetitive, especially in the drums. This did not track one checkpoint or one planner size.

The current candidate ranker includes a positive periodicity contribution. That metric is measured as short-window waveform autocorrelation (lags 80...1200 samples), so it describes local tonal/rhythmic periodicity—not arrangement-level repetition. The captured runs often selected the candidate with the highest local periodicity, but that alone does not prove the ranker selected the most repetitive arrangement.

Required controlled test

1. Render a fixed 60–90 second prompt with --keep-candidates, fixed checkpoint/planner/metadata/seed family, and at least four candidates.
2. Blind-listen and label section development, drum-pattern change, transition count, and obvious repeated blocks for every candidate.
3. Compare those labels against overall rank and each technical metric.
4. Repeat on at least Turbo and SFT with the same independently resolved 1.7B planner.

Acceptance criteria

• If the selected candidate is consistently less varied than alternatives, replace or cap the local-periodicity reward and add an arrangement-scale repetition metric/test fixture.
• If candidates are similarly repetitive, keep the ranker unchanged and investigate prompt/duration/model behavior instead.
• Preserve technical rejection of silence, clipping, non-finite output, and broken tails.

This issue intentionally does not treat the earlier raw LM duration diagnostic as evidence: recipes showed effective duration and semantic-code counts were already bounded, and PR #253 makes that precedence/reporting explicit.