audiodynamics: a gain-smoothing option, so a fast release does not ride a 50 Hz waveform

[bdbarnett]

Stakes. Compressor's dossier fixed three "clean" traits — F5 at 0.5 %
THD, V6 at 0.2 %, M5 at 1 % — and asked one question at Station A: "Is gain
smoothing composable, or a fifth ask? If the shipped patches cannot reach the
three bars, the ask is a one-pole on the gain after the gain computer."
Measured 2026-09-07 on the shipped patches: they cannot, on
low-frequency material.

Read back off the instance, at 10 dB of gain reduction on a 50 Hz tone
(audiocomponents/tools/compressor_evidence.py patches, 48 kHz, CPython at
AUDIOIF_PIN 2f6cbc3):

  patch  name                    character   THD @50 Hz  THD @1 kHz
      0  Level Ride              optical      0.8323 %   0.0446 %
      1  Fast Peak Catch         fet          1.1289 %   0.0611 %
      2  Everything At Once      fet          4.0207 %   0.2668 %
      3  Bus Glue                vca          0.2415 %   0.0117 %
      5  Let The Stick Through   fet          0.9393 %   0.0462 %
      6  Parallel Squash         fet          1.0945 %   0.0735 %

Five of the fourteen shipped patches are over F5's 0.5 % bar at 50 Hz —
patch 0, the constructor's own defaults, among them — and patch 3 is over
V6's 0.2 %. Patch 2 is the 1176's all-buttons-in mode and is meant to be
dirty. At 1 kHz only patch 2 is over the bar, so this is specifically
low-frequency material: with a short release the per-sample gain follows the
waveform. Swept over the Release macro's whole travel at 10 dB of GR, patch 1
reads 0.0209 % at the slow end and 3.8909 % at the fast end (50 Hz);
1.1359 % at 200 Hz; 0.2397 % at 1 kHz.

The class cannot fix this with the nodes on the pin: the ripple is in the
gain the node computes per sample, and nothing downstream of Dynamics can
remove it without also removing the compression.

Ask. audiodynamics.Dynamics gets a one-pole smoothing on the computed
gain, after the gain computer and before it multiplies the audio —
gain_smooth_ms, default 0, so every figure already measured against the
pin stays exactly where it is. A class that wants the 1176's dirt keeps 0; a
class that wants the dbx's cleanliness at 50 Hz asks for a few milliseconds.

Related: the class's own evidence records the disconfirmation and the bound
(audiocomponents/docs/effects/Compressor-evidence.md §1, traits F5/V6, and
docs/effects/Compressor.md §8 question 7, which is the question this ask
answers).