feat(perf): PWM-phase histogram of zero-crossings (v4) — hump mechanism identified

[AlexKlimaj]

The diagnostic proposed after #23's refutation: a 32-bin histogram of the PWM phase (TIM1->CNT/ARR) at which each accepted zero-cross fired. Captured at ISR entry (pre-confirm), committed on accept only, monotonic u16 bins differenced by the host per steady tail. Perf struct v4 (148 B); host decodes v1–v4; includes DWARF array-member sizing support and a frozen v3 compat fixture. 137 offline tests, both build variants clean.

First capture answers the question (runs/phase-hist-probe, jitter_probe)

The phase distributions are massively non-uniform — and the structure identifies the hump mechanism:

point	duty	off-time	detections in bins 0-2 (uniform: 9.4%)	jitter
t30	24%	76%	27.1%	4.9%
t60	61%	39%	35.9%	23.5%
t70	71%	29%	30.8%	18.6%
t90	90%	10%	25.8%	7.2%
t100	100%	0%	12.8% (≈uniform)	2.9%

Two features, consistent at every point:

1. A large pile-up in bins 0–2 — immediately after PWM turn-on — scaling with the off-time fraction, and
2. Heavy depletion of the bins covering the PWM off-period (at t60, bins ~17–31 are nearly empty out of 45k events; the off-edge sits at bin ~19.5). At t100 there is no off-time and the distribution is nearly flat — and jitter is at its floor.

The mechanism

Zero-crossings that occur during the PWM off-time cannot be detected; they surface as comparator edges at the next turn-on. Detection is quantized to the PWM grid by up to the full off-window (~19 µs at t60, on a 66 µs commutation interval). Those quantized timestamps feed commutation_interval → advance/waitTime, so the control loop schedules the next window off a corrupted measurement and hunts — which is why measured jitter (23%) is several times the direct quantization error (~4%), and why the hump lives where commutation frequency approaches PWM frequency (the walk through the off-window beats slowly). It also explains every prior negative result: coarse/fine sampling, glitch tolerance, and #23's blanking all operate at the detection instant — none of them can see a crossing the comparator was blind to.

Proposed next experiment (cheap, concrete)

Turn-on pile-up timestamp compensation: for detections landing within the first ~2 bins after turn-on when duty < 100%, the true crossing occurred somewhere in the preceding off-window — subtract half the off-window duration (duty and ARR are both known) from thiszctime. On average this halves the quantization error, and more importantly decouples the commutation loop from the PWM grid, which should collapse the hunting amplification. Fully measurable with this histogram + the jitter metric via the interleaved A/B protocol.

Merging this PR adds the instrument (report-only metrics: zc_phase_peak_ratio/peak_bin + full histogram per steady point); the compensation experiment would be its own PR.

🤖 Generated with Claude Code