The master stage was a flat 0.5x attenuation into a static tanh soft-clip,
leaving ~9 dB of unused headroom: even the loudest single preset with its mix
bus maxed peaked around -7 dBFS, well below other media playback.
Replace it with a generous makeup gain (1.5x, ~+9.5 dB) feeding a feed-forward,
look-ahead brickwall limiter. A short delay line lets the peak detector pull the
gain down before a loud sample reaches the output and hold it across the
look-ahead window, so the output is mathematically guaranteed to stay under the
ceiling (0.97) - loudness without clipping, and none of the tanh grit a static
clipper adds to a hot single voice. Allocation-free (results go to scratch
fields, not a Pair) and NaN-guarded.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
5943040a2b