Silentium

An expander whose neutral setting is the old code, not an approximation of it. Ratio runs a continuous downward-expander law between Threshold and Range — (Ratio − 1) dB per dB below Threshold, measured to within 0.25 dB at 2:1, 4:1 and 8:1, with a soft knee that is continuous in value and slope at both edges so widening it never introduces a corner. Range is the law's floor, not a second stage. And at the top of the range the gain computer branches to the literal pre-existing binary path, so "unchanged by default" is a code-level fact rather than a numerical coincidence.

A 0 ms attack that genuinely does not click. Smooth Open shapes the opening with a moving-max plus two cascaded box filters, all in the dB domain and all inside the existing lookahead window. A backward moving-max delays falls, not rises — so a rising edge passes it instantly and the box cascade spreads it linearly across the window, finishing exactly as the delayed transient leaves the delay line. Steepest step: 59.5 → 0.50 dB per sample, a 119× reduction, within 0.01% of the triangular kernel's theoretical peak. Reported latency at Lookahead 0, 2.5, 5 and 20 ms is identical with Smooth Open on and off. An earlier max(ballistic, smoothed) design was tried and rejected — at 0 ms attack the raw step wins the max and the click survives — which is why the smoother runs in series instead.

Lookahead you can move while the music is playing. The delay crossfades equal-power to its new length over 10 ms; the impulse verifiably moves from 240 to 480 samples with no re-prepare, and the largest sample-to-sample step through the transition is smaller than a fifth of a decibel. The new latency reaches the host from a 25 ms message-thread poll of a relaxed atomic — deliberately not an async update posted from the audio thread, which would take a lock and might grow an array. processBlock() allocates zero bytes even on the block where the lookahead moves.

Two detectors and two sidechain slopes, both always warm. The RMS detector (5 ms mean-square) and the 24 dB/oct Butterworth-Q-paired filter chain run every block, whether selected or not, so switching is a 5 ms or 10 ms crossfade between two live signals instead of a jump to a cold filter with a settling transient. Measured: 12.30 dB and 24.11 dB of rejection one octave below cutoff, against a theory of 12.3 and 24.1. On material with isolated spikes over a quiet bed, the peak detector opens the gate 99 times a second where RMS opens it not at all.

Neutrality proven against the previous version's own code. The test suite holds renders captured from a build of the prior engine before a single line of the new release was written, with full provenance — commit, compiler, host, flags, toolchain tag. A two-tier comparison keeps that honest across platforms: an always-asserted windowed level fingerprint at 2e-3 dB (cross-toolchain float noise lands at 1e-5; a real voicing change moves a window past 1e-2), and an exact SHA-256 check on the generating configuration where the toolchain matches. And the policy is proved non-vacuous: the same test asserts that a genuinely different Ratio setting must fail it.

Honest engineering, enforced by CI. 114 test cases and 9,156 assertions on macOS and Windows, plus pluginval at strictness 10 and auval -strict: zero heap allocations on the audio thread under a replaced allocator with every feature engaged, renders independent of host block size within 1e-6 RMS, exactly 0.0 output after ten seconds of silence, and an open gate that is still a pure delay to better than −120 dBFS with every new stage switched on at once.