Measurements

Every number here is printed by the host tests (make test), which run the same DSP core the firmware and the plugin use. CI runs them on every pull request.

1V/oct tracking, −3 V … +5 V	worst error 0.09 cents
Alias energy, heavy fold @ 2.96 kHz	−9 dB → −84 dB with 4×
Alias energy, FM index 1.0 @ C7	−40 dB → −84 dB with 4×
Self-feedback 0 → 0.5 → 1	pure sine (H2 −117 dB) → saw-like (H2 −10 dB) → noisy
60 s worst-case evil (fb, chaos, Lorenz, all max)	finite, alive, no DC lock-up, at 4 pitches
Output range	±5 V nominal, DC-free; heavy feedback/chaos peaks are soft-capped at 7.2 V
FLESH 0 → 1: spectral centroid	220 Hz → 6.8 kHz, rising at every step
MALICE: inharmonic share at 0.5 / 0.8 / 0.9	0.00 / 0.97 / 1.00
LEGION 0.4 → 1: beating	2.8× → 4.8×
Unison: 4 voices, no spread vs 1 voice	identical to within 0.05 mV
Unison: 4 voices spread, biased fold	beat-rate wobble 0.05 V (0.26 V without the envelope-aware DC fix)
patch.Init flash / RAM	106 KB of 128 KB / 36 KB (arm-gcc 13.3)

What the suite checks

Pitch accuracy, stability under abusive parameters and external FM, through-zero behaviour, the three macros (no jumps in their curves, and each one's sweep: FLESH's brightness, MALICE's path from harmonic to chaos, LEGION's sub and beating), the unison stack (one-voice equivalence, beating from spread and drift, the sub, DC under a biased fold, click-free voice changes), the feedback spectrum, that chaos actually breaks periodicity, Lorenz boundedness from 0.01 Hz to 20 kHz, 60-second worst-case runs, the shift-layer pickup logic, and output bounds. Any failed check makes make test exit non-zero.

make analyze also writes spectrograms of the fold and FM sweeps with and without oversampling to test/out/aliasing.png, and prints the alias energy of each steady tone:

non-harmonic (alias) energy in steady tones — lower is better
  fold_static OS1    -9.0 dB   OS4   -83.9 dB   improvement  74.9 dB
  fm_static   OS1   -39.7 dB   OS4   -83.9 dB   improvement  44.1 dB