SEMANTIFX_EQ_EXACT=1 (or exact=True) applies the dasp RBJ sections with a
float64 torchaudio lfilter cascade instead of the frequency-sampling
approximation, matching the plugin's double-state biquads. Off by default:
CPU cost is 2.3x, but torchaudio's lfilter is ~1400x slower on CUDA.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>