Overture

A clipper you can test against the circuit, not against itself. Because the Feedback voicing is a solver rather than a curve, its linear regime has a closed-form digital transfer function — so the measured response is checked against an absolute analytic prediction (±0.25 dB, 20 Hz–10 kHz), a pole/zero fit recovers the circuit's own pre-emphasis zero and its feedback pole to within 5% of the component-derived values, and its harmonic structure is compared against the same solver run at 32× the rate. The Newton iteration is safeguarded with a diode-aware bracket, so the 8-iteration cap is hard, not best-effort: swept over a wide voltage grid across four sample rates, three oversampling factors, five Drive settings and three asymmetries, it never exceeded eight iterations and never produced a non-finite sample.

A gate that closes like a fader. Per-sample, log-domain control path: a 5 ms mean-square detector behind a sidechain filter pair, a non-linear-capacitor smoother whose time constant collapses as the level step grows, 4 dB of hysteresis and a 20 ms retriggering hold. The closing ramp is dB-linear — a straight line in decibels, line-fitted in the test suite to better than R² = 0.99, which an exponential-in-linear-gain release cannot reach at the same endpoints. Zero added latency: no lookahead.

One Auto Release for two completely different jobs. Two envelopes race on the detector output. When the slow one pulls clear of the fast one by more than a few dB, the note stopped abruptly, so the gate dumps it and closes fast. Otherwise it releases at the note's own measured decay rate plus a small margin. Measured: −80 dB within 150 ms after a staccato stop, and never more than 1.5 dB down anywhere during a decaying chord — with the test asserting which branch fired, so passing by luck is not possible.

Anti-aliasing that is opt-in without being approximate. Enhanced wraps the three memoryless voicings in first-order antiderivative anti-aliasing — closed-form antiderivatives, an overflow-safe ln(cosh) branch, an input-scaled epsilon fallback — plus a 5 Hz DC blocker. It wraps the voicings and never edits them, which is why Classic remains the previous release's path bit for bit. On the plugin's worst case (a full-scale tone, 40 dB of Drive, a hard clamp) it removes 18.7 dB of alias energy at 2× oversampling, and Enhanced at 2× beats Classic at 4× by 9.3 dB. Harmonics move by less than 0.3 dB: this changes the aliases, not the tone.

Automation that stopped zippering, without changing a settled render. Filter coefficients now update every 32 samples instead of once per host block — roughly 1.5 kHz instead of 86 Hz — which measures as at least 12 dB less broadband non-harmonic spray when Bite Tilt is automated. And because a coefficient is only re-derived when its smoothed value has actually moved, a settled parameter's output is still bit-identical to the previous engine at sixteen times the cadence. A settled engine renders identically at block sizes of 512, 128, 64, 37 and 511.

Neutral is a regression gate, not a promise. Golden renders were recorded from the earlier engine before any of the new code existed, and are asserted for every legacy voicing at every oversampling factor. A session restored from that earlier state renders with a worst residual of exactly zero against one configured by hand. The older factory presets are byte-frozen by hash.

Oversampling adds a few samples of latency, honestly reported — the gate does not. The gate itself has no lookahead and adds nothing. Oversampling does: measured at 4 samples at 2× and 6 samples at both 4× and 8× at 48 kHz, empirically, non-decreasing as the factor rises. Reported latency is unchanged by every other new feature, at every oversampling factor, and the host compensates automatically.

Real-time discipline, proven under a replaced allocator. Zero heap allocations in processBlock() with the gate open and closing, the circuit solver running, Enhanced quality on and seven parameters sweeping their full ranges. A 30-second soak of silence, full-scale squares and decaying noise stays finite with no denormal stall. AU, VST3 and Standalone, built and tested on macOS and Windows on every push, then pluginval at strictness level 10 and auval -strict. 129 test cases, 24,213 assertions, Debug and Release.