Triptych

A detector, in three orthogonal layers. Peak is a thin wrapper around the exact envelope-follower call the previous version made — deliberately not a re-implementation. RMS runs a mean-square one-pole at max(Attack, 5 ms) into that same ballistics chain, reading the textbook 3.01 dB below peak on a sine. Auto Release runs a fast branch and a slow reservoir in parallel and takes the maximum, which is what guarantees the tail decays monotonically instead of bouncing. Stereo Link is a max-law blend applied to the detector inputs, so at 100% both channels integrate the same value sequence and their gains cannot differ by even a rounding step.

A ceiling with a proof, not just a test. The lookahead brickwall pairs an amortised-O(1) sliding minimum with a cascaded box smoother whose FIR support never exceeds the lookahead window. Because every value averaged into the final gain is bounded by the request at the delayed sample, and the release pole only lowers it further, the output magnitude can never exceed the threshold — for any program material. Verified over 10,000 randomised segments per setting, driven well above full scale, with worst-case overshoot below 1e-6.

A sidechain that matches the band it keys. Selecting External runs the key through a second, independent crossover pair at the same frequencies and slope, so the low band's detector follows the sidechain's low band — not its cymbals. Nothing connected? It falls back to internal keying sample-exactly, not silently quiet. SC Listen plays a band's detector key so you set thresholds against what the detector actually hears.

Crossovers built the hard way. 12 and 48 dB/oct are genuine Linkwitz-Riley designs — a squared 1st-order and a squared 4th-order Butterworth prototype — because cascading the existing 24 dB/oct crossover with itself is a different transfer function, not a steeper one. The 24 dB/oct path is byte-untouched and asserted sample-exactly against the original. Measured: 12 ±1.5, 24 ±2.0, 48 ±3.0 dB per octave, with the band sum flat to ±0.1 dB at the default.

A gate that cannot click. Hold lives in the envelope domain as a shadow held envelope that decays with the gate's own release coefficient; Hysteresis moves the effective threshold through the 50 ms smoother that already owns it. Neither ever steps the output gain by more than 0.5 dB per sample — asserted across hold expiry and every hysteresis transition, worst case included.

Zero latency is the default, not a universal property. Every band's crossover is minimum-phase and every detector is causal, so a fresh instance reports 0 samples of latency — and every session saved before Lookahead existed still does. Engage Lookahead on any band and that stops being true: the plugin reports exactly the lookahead length, in samples, and the host re-runs delay compensation to match.

Compatibility as an engineering contract. 23 new parameters, all neutral by default, all appended after the 59 shipped ones so AU automation lanes keep their indices. A fresh instance renders sample-exactly like the previous version — proven by a same-binary A/B against a reference chain rebuilt from the very components still linked, with bit-exact equality rather than a tolerance. Zero audio-thread allocations with the entire feature matrix engaged, under a replaced allocator that counts.