fix: avoid audio-thread stalls on model eviction and AU bypass toggle

[Ahorix]

Problem

Disabling the NAM plugin in Ableton Live (and likely other AU hosts) via the device power button causes an audible dropout/crackle, even on completely silent tracks. This was reproducible and isolated to NAM — simpler plugins toggle cleanly.

Two separate real-time safety violations were found and fixed.

---

Fix 1: Deferred model deletion

_ApplyDSPStaging() is called from ProcessBlock (audio thread). When a model is removed or replaced, the outgoing unique_ptr destructor was freeing several MB of Eigen weight matrices inline on the audio thread. On macOS, free() is non-deterministic on large blocks and can stall the audio callback long enough to miss the buffer deadline.

Fix: evict the old model into mModelPendingDeletion (an std::atomic<ResamplingNAM*>) instead of destroying it in place. OnIdle() (UI thread) performs the actual delete on the next tick (~20 ms later).

Fix 2: Skip _ResetModelAndIR() on AU bypass toggles

iPlug2 calls OnReset() in response to kAudioUnitProperty_BypassEffect (the AU power-button toggle). This triggered _ResetModelAndIR() which calls DSP::Reset() → SetMaxBufferSize() + prewarm(). prewarm() runs the model forward for thousands of samples to settle LSTM state — all on the audio thread — causing a guaranteed dropout on every toggle.

Fix: cache the last sample rate and block size seen by OnReset(). Skip _ResetModelAndIR() when neither has changed, which is always the case for a bypass toggle. Legitimate format changes (sample rate, buffer size) still trigger a full reset.

Note: the iPlug2 AUv2 handler for kAudioUnitProperty_BypassEffect already has a // TODO: should the following be called here? comment on the OnReset() call, suggesting this was a known uncertainty upstream. The answer for NAM is: no, it should not.

---

Testing

• Toggle power button off/on repeatedly → no dropout
• Load a model, swap to a different model → no dropout, new model sounds correct
• Change DAW sample rate with model loaded → model resets correctly
• Change buffer size → model resets correctly
• Multiple instances → independent, no interference