- **KNOWN BUG**: Single `filter` instance processes both L and R sequentially — state gets contaminated between channels, causing crashes when filter env amount > 0. FIX: add `Filter filterR` member, process left through `filter`, right through `filterR`.
`SubtractiveVoice` now has separate `filter` (L) and `filterR` (R) instances; `startNote()` prepares both. Resolves the state-sharing crash when filter env amount > 0.
- New class `DSP/Arpeggiator.h` — tempo-synced, pure note-logic (no audio). Tracks held notes (press order for the As Played pattern), builds a pitch pool from held notes × octave range (1–3, clamped to MIDI range), and sequences it per pattern.
- Patterns: Up, Down, UpDown, DownUp, Random, As Played, Chord.
-`arpDirection` (Up/Down) controls whether octave copies extend above or below the root; sequential patterns always order by ascending pitch.
- Routing (`PluginProcessor.cpp::processBlock`): when `arpEnabled`, MIDI note events feed the arp instead of the synth; the arp drives `synth.noteOn/noteOff` (with transpose) on each tick via `arpSampleCount` accumulation. Piano-roll `noteOn/noteOff` route through the same logic. When disabled, `arp.reset()` stops any sounding arp note and notes play normally.
- UI: bottom-right section at (860, 746) titled ARPEGGIATOR with 5 combos (ON, PATTERN, OCTAVES, DIR, RATE).
-`WaveformDisplay` (bottom-left, 60×746, 790×128) draws the scope ring buffer waveform AND a semi-transparent `juce::dsp::FFT` spectrum overlay behind it (30 fps). Requires `juce::juce_dsp` linked and `#include <juce_dsp/juce_dsp.h>`.
- Standalone `SpectrumAnalyzer` component was removed.
- Standalone JUCE console tool that renders every preset offline (C4, 8s, 44.1 kHz) and derives a new `masterLevel` for consistent loudness: peak normalized to 0.80, RMS-normalized to 0.15, final value = `min(gainPeak, gainRms) * current`, clamped to 0..1.
- Build target: `cmake --build build --target level_normalizer`; run `./build/level_normalizer --apply` to rewrite the `masterLevel` entries in `Source/PresetManager.h` (reports a count mismatch and aborts if preset order/count ever drifts).
- Applied 2026: all 136 presets re-balanced; `masterLevel` now ranges 0.325..1.0. Sustained/washy presets (heavy reverb/delay) get the largest cuts; quiet transient presets clamp at 1.0.
- Caveat: noise-based presets render with slight run-to-run variance (`juce::Random` reseeds per process), so re-running can shift borderline values by a few percent.
1.**ComboBox items**: Must manually call `addItemList()` BEFORE creating `ComboBoxAttachment` — the attachment alone does NOT populate items.
2.**AffineTransform::scale**: Use `AffineTransform::scale(float, float)`, NOT `scaled()` — the method name in JUCE is `AffineTransform::scale`.
3.**Filter state sharing**: Never share a single `Filter` instance between L/R or between voices — each needs its own filter state.
4.**JUCE modules**: If you use `juce::dsp::FFT` or any `juce_dsp` functionality, you must link `juce::juce_dsp` in CMakeLists.txt AND add `#include <juce_dsp/juce_dsp.h>`.
5.**Denormals**: `processBlock()` wraps with `juce::ScopedNoDenormals` — always keep this.
6.**Voice parameter update**: `updateVoiceParameters()` is called every `processBlock()` — this reads all 27 params and pushes to all 16 voices. This is intentional for simplicity but is CPU-expensive.
7.**Oscillator pan**: The oscillator's `process()` adds directly to output (`*leftOut += ...`), so it's additive between the two oscs in each voice.
8.**Drive math**: `std::tanh(left * drive) / std::tanh(drive)` — when `drive = 1.0`, division by `tanh(1.0) ≈ 0.762` normalizes. `drive` is clamped to min 1.001 in `setParameters()`.
