Clean-code audio / DSP (depth)

Apply on top of the audio-plugin base rules (RT-audio non-negotiables, DSP-is-sacred, golden-buffer tests, pluginval). Also use clean-code-cpp.

The two-thread model

• The audio thread (processBlock) and the message thread (UI, params) never share mutable state via locks.
• Pass parameters across with std::atomic or a lock-free FIFO; the audio thread only reads smoothed snapshots.

Parameter smoothing

• Never apply a raw parameter jump in the audio path — zipper noise. Smooth it (juce::SmoothedValue or a one-pole) and read per-sample or per-block.

The inner loop

• Process in blocks; hoist invariants out of the per-sample loop.
• Prefer branch-free math in the hot loop (select/min/max over if); branches stall the pipeline and hurt determinism.
• Prefer float unless precision demands double. Flush denormals (ScopedNoDenormals), and design filters/feedback so denormals don't accumulate.

Testability & sacredness

• Keep DSP as pure functions/classes over buffers in Source/dsp/, decoupled from JUCE UI, so it's unit-testable with signal fixtures and golden buffers (impulse, sine sweep).
• processBlock stays a thin dispatcher: read the smoothed parameters, then call DSP functions that take the buffer and their state as arguments.
• Randomness (noise, dither, humanize) comes from a preallocated RNG the caller passes in and seeds. Tests use a fixed seed, so golden buffers stay reproducible.
• DSP is sacred: a golden-buffer test failing on a non-DSP change means something leaked into the algorithm — stop and fix the boundary, don't update the golden file.

Signals it's going wrong

• Allocation/lock/log inside processBlock → move it to the message thread or preallocate.
• A parameter read directly in the loop without smoothing → audible artifacts.
• DSP logic reaching into UI/JUCE component state → break the coupling.