Whalesong: rhythm forensics

whalesong_rhythm bins a trace's onset train into 1,024 samples and runs the instrument's real DSP: centered variance-normalized autocorrelation and a Hann-windowed one-sided periodogram. Frequencies here are cycles/second of the event train — not audio pitch, not the sonogram's semantic rows.

Reading the output

• dominantPeriodSeconds — the strongest repetition interval. ~4s means something fired every four seconds; match it against loop findings.
• spectralEntropy 0–1 — low = one dominant rhythm (metronomic), high = arrhythmic activity. Use it to separate "stuck on a timer" from "busy".
• strongestAutocorrLagSeconds + autocorrAtBestLag — self-similarity at its best offset; a high value confirms periodic structure the spectrum might split into harmonics.
• Sparse onset trains produce harmonics — the dominant peak may be a multiple of the true period. Cross-check with the autocorr lag.

Windows and categories

Default is the whole trace. Narrow with start_ms/end_ms (finding atMs values are relative to trace start — feed them straight in) or one category (e.g. tool for just the call cadence, reasoning for model cadence). A stalled session's gap boundaries make good window edges.

Workflow

1. whalesong_analyze or the sonogram flags a repeating pattern.
2. whalesong_rhythm on the flagged window → period + confidence.
3. whalesong_observations on that window, full: true → what the metronome was actually calling.
4. State it plainly: "retried bash <cmd> every ~8.5s for 40 min", not "exhibited anomalous periodicity".

Cadence is a property of the record — it still doesn't prove intent.