the audio thread

A glitch in audio is a missed deadline. So the audio thread does only what is absolutely necessary, and never anything whose time is not bounded.

what never happens there

• No allocation and no freeing. Not directly, and not hidden behind a container’s push_back or resize, a std::function copied, a shared_ptr whose count may reach zero, or a lambda capturing by value into a callback.
• No locks, including one reached through a helper.
• No logging, no file access, no juce::String.
• No shared memory. The link between instances is the message thread’s; the audio thread hands its state to a publisher there, which costs about a nanosecond.

Atomics are fine, with the reasoning for their ordering stated beside them in the code.

how state reaches it

The message thread owns the patch. Every change becomes an immutable snapshot, which the audio thread adopts by pointer at the start of a block. The old snapshot goes back to the message thread through a retire queue and is freed there. A state load returns in a fraction of a millisecond however busy the audio thread is, and nothing is ever freed inside processBlock.

the control grid

Modulation runs on a fixed grid of 256 samples, whatever block size the host uses, so the result never depends on the host’s buffer. New modulation or feature work joins that grid rather than inventing a rate of its own.

what may be modulated

A parameter can be a modulation target only if its effect is recomputed on the audio thread, in bounded work, with everything it needs sized in advance — and, if it is a gain, interpolated across the grid step rather than stepped at its edge. A value that needed the message thread to rebuild something would arrive at a sample the scheduler picked, and an offline render would pick a different one. So a room’s size, which changes hundreds of delay lengths, is a setting and not a target.

memory is sized at prepare

Anything the audio thread might need is allocated when the plugin is prepared, at its worst case. A buffer that grew on demand would grow at a moment the message thread’s timing decides, which breaks determinism as surely as it breaks the deadline.

Denormals are flushed for the length of each block by one scoped guard, the same one the render tools use.