pi-multifx-pedal/src/dsp/pipeline.py

"""FX/Audio pipeline — the main real-time signal chain.

Runs on RPi 4B under JACK, connecting:
  Guitar -> Gate -> Comp -> Boost -> NAM Amp -> IR Cab -> EQ -> Mod -> Delay -> Reverb -> Volume -> Out

Each block can be bypassed per-preset. The pipeline manages
block-level audio routing using numpy arrays for zero-copy
inter-block communication.

All DSP state is stored per-block-instance in self._state,
keyed by chain index. This allows multiple instances of the
same effect type at different positions in the chain.
"""

from __future__ import annotations

import logging
import threading
from dataclasses import dataclass, field
from typing import Optional

import numpy as np
from scipy.signal import lfilter

from .nam_router import NAMEngineRouter
from .ir_loader import IRLoader, IRFile
from ..presets.types import FXBlock, FXType, Preset

logger = logging.getLogger(__name__)

BLOCK_SIZE = 256       # Samples per JACK callback
SAMPLE_RATE = 48000    # Standard guitar audio rate

# ── Biquad coefficient helpers ─────────────────────────────────────

_EPS = 1e-10


def _compute_lowshelf_coeffs(freq: float, gain_db: float, q: float, sr: float) -> tuple:
    """RBJ low-shelf biquad coefficients."""
    a = 10 ** (gain_db / 40.0)
    omega = 2 * np.pi * freq / sr
    sn = np.sin(omega)
    cs = np.cos(omega)
    beta = np.sqrt(a) / q  # sqrt(A) / Q
    if gain_db >= 0:
        b0 = a * (a + 1 - (a - 1) * cs + beta * sn)
        b1 = 2 * a * (a - 1 - (a + 1) * cs)
        b2 = a * (a + 1 - (a - 1) * cs - beta * sn)
        a0 = a + 1 + (a - 1) * cs + beta * sn
        a1 = -2 * a * (a - 1 + (a + 1) * cs)
        a2 = a + 1 + (a - 1) * cs - beta * sn
    else:
        b0 = a * (a + 1 + (a - 1) * cs + beta * sn)
        b1 = -2 * a * (a - 1 + (a + 1) * cs)
        b2 = a * (a + 1 + (a - 1) * cs - beta * sn)
        a0 = a + 1 - (a - 1) * cs + beta * sn
        a1 = 2 * a * (a - 1 - (a + 1) * cs)
        a2 = a + 1 - (a - 1) * cs - beta * sn
    return (b0 / a0, b1 / a0, b2 / a0, a1 / a0, a2 / a0)


def _compute_highshelf_coeffs(freq: float, gain_db: float, q: float, sr: float) -> tuple:
    """RBJ high-shelf biquad coefficients."""
    a = 10 ** (gain_db / 40.0)
    omega = 2 * np.pi * freq / sr
    sn = np.sin(omega)
    cs = np.cos(omega)
    beta = np.sqrt(a) / q
    if gain_db >= 0:
        b0 = a * (a + 1 + (a - 1) * cs + beta * sn)
        b1 = -2 * a * (a - 1 + (a + 1) * cs)
        b2 = a * (a + 1 + (a - 1) * cs - beta * sn)
        a0 = a + 1 - (a - 1) * cs + beta * sn
        a1 = 2 * a * (a - 1 - (a + 1) * cs)
        a2 = a + 1 - (a - 1) * cs - beta * sn
    else:
        b0 = a * (a + 1 - (a - 1) * cs + beta * sn)
        b1 = 2 * a * (a - 1 - (a + 1) * cs)
        b2 = a * (a + 1 - (a - 1) * cs - beta * sn)
        a0 = a + 1 + (a - 1) * cs + beta * sn
        a1 = -2 * a * (a - 1 + (a + 1) * cs)
        a2 = a + 1 + (a - 1) * cs - beta * sn
    return (b0 / a0, b1 / a0, b2 / a0, a1 / a0, a2 / a0)


def _compute_peaking_coeffs(freq: float, gain_db: float, q: float, sr: float) -> tuple:
    """RBJ peaking biquad coefficients."""
    a = 10 ** (gain_db / 40.0)
    omega = 2 * np.pi * freq / sr
    sn = np.sin(omega)
    cs = np.cos(omega)
    alpha = sn / (2 * q)
    b0 = 1 + alpha * a
    b1 = -2 * cs
    b2 = 1 - alpha * a
    a0 = 1 + alpha / a
    a1 = -2 * cs
    a2 = 1 - alpha / a
    return (b0 / a0, b1 / a0, b2 / a0, a1 / a0, a2 / a0)


def _compute_hpf_coeffs(freq: float, q: float, sr: float) -> tuple:
    """RBJ high-pass biquad coefficients."""
    omega = 2 * np.pi * freq / sr
    sn = np.sin(omega)
    cs = np.cos(omega)
    alpha = sn / (2 * q)
    b0 = (1 + cs) / 2
    b1 = -(1 + cs)
    b2 = (1 + cs) / 2
    a0 = 1 + alpha
    a1 = -2 * cs
    a2 = 1 - alpha
    return (b0 / a0, b1 / a0, b2 / a0, a1 / a0, a2 / a0)


def _compute_lpf_coeffs(freq: float, q: float, sr: float) -> tuple:
    """RBJ low-pass biquad coefficients."""
    omega = 2 * np.pi * freq / sr
    sn = np.sin(omega)
    cs = np.cos(omega)
    alpha = sn / (2 * q)
    b0 = (1 - cs) / 2
    b1 = 1 - cs
    b2 = (1 - cs) / 2
    a0 = 1 + alpha
    a1 = -2 * cs
    a2 = 1 - alpha
    return (b0 / a0, b1 / a0, b2 / a0, a1 / a0, a2 / a0)


def _compute_bpf_coeffs(freq: float, q: float, sr: float) -> tuple:
    """RBJ band-pass biquad coefficients."""
    omega = 2 * np.pi * freq / sr
    sn = np.sin(omega)
    cs = np.cos(omega)
    alpha = sn / (2 * q)
    b0 = sn / 2
    b1 = 0.0
    b2 = -sn / 2
    a0 = 1 + alpha
    a1 = -2 * cs
    a2 = 1 - alpha
    return (b0 / a0, b1 / a0, b2 / a0, a1 / a0, a2 / a0)


def _compute_notch_coeffs(freq: float, q: float, sr: float) -> tuple:
    """RBJ notch biquad coefficients."""
    omega = 2 * np.pi * freq / sr
    sn = np.sin(omega)
    cs = np.cos(omega)
    alpha = sn / (2 * q)
    b0 = 1.0
    b1 = -2 * cs
    b2 = 1.0
    a0 = 1 + alpha
    a1 = -2 * cs
    a2 = 1 - alpha
    return (b0 / a0, b1 / a0, b2 / a0, a1 / a0, a2 / a0)


# ── Circular delay line (block-vectorised) ─────────────────────────

class _DelayLine:
    """Vectorised circular buffer with linear interpolation."""

    __slots__ = ("buf", "max_len", "write_idx")

    def __init__(self, max_delay_samples: int):
        self.buf = np.zeros(max_delay_samples, dtype=np.float32)
        self.max_len = max_delay_samples
        self.write_idx = 0

    def write_block(self, block: np.ndarray) -> None:
        n = len(block)
        pos = 0
        while pos < n:
            if self.write_idx >= self.max_len:
                self.write_idx = 0
            space = self.max_len - self.write_idx
            chunk = min(n - pos, space)
            self.buf[self.write_idx:self.write_idx + chunk] = block[pos:pos + chunk]
            self.write_idx += chunk
            pos += chunk
        # Keep type: numpy automatically promotes on write into float32

    def read_block_varying(self, delay_samples: np.ndarray) -> np.ndarray:
        """Read a block with different (fractional) delay per sample.

        Fully vectorized using numpy advanced indexing.
        ``delay_samples`` must have shape (N,) or broadcastable to it.
        """
        delays = np.asarray(delay_samples, dtype=np.float64)
        int_delays = np.floor(delays).astype(np.int32)
        frac = delays - int_delays
        read_start = (self.write_idx - int_delays) % self.max_len
        read_next = (read_start + 1) % self.max_len
        return (self.buf[read_start] * (1.0 - frac)
                + self.buf[read_next] * frac)

    def read_block(self, delay_samples: float, n_samples: int) -> np.ndarray:
        """Read n_samples with linear interpolation at a fractional delay."""
        n_delay = int(delay_samples)
        frac = delay_samples - n_delay
        read_start = (self.write_idx - n_delay) % self.max_len
        indices = (read_start + np.arange(n_samples)) % self.max_len
        next_indices = (indices + 1) % self.max_len
        return self.buf[indices] * (1.0 - frac) + self.buf[next_indices] * frac

    def add_to_block(self, block: np.ndarray, delay_samples: float,
                     gain: float) -> np.ndarray:
        """Add delayed + gained signal to block (for feedback loops)."""
        n_delay = int(delay_samples)
        frac = delay_samples - n_delay
        read_start = (self.write_idx - n_delay) % self.max_len
        indices = (read_start + np.arange(len(block))) % self.max_len
        next_indices = (indices + 1) % self.max_len
        delayed = self.buf[indices] * (1.0 - frac) + self.buf[next_indices] * frac
        return delayed * gain

    def read_all(self) -> np.ndarray:
        """Return the full buffer (for debugging / IR export)."""
        return self.buf.copy()


# ── Schroeder reverb helpers ───────────────────────────────────────

class _CombFilter:
    """Comb filter for Schroeder reverb."""

    __slots__ = ("delay", "feedback", "damping", "damp_filt", "buf", "_buf_size")

    def __init__(self, delay_samples: int, block_size: int = 256):
        line_len = max(block_size * 2, delay_samples + 1)
        self.delay = _DelayLine(line_len)
        self.feedback: float = 0.5
        self.damping: float = 0.5      # low-pass damping coefficient
        self.damp_filt: float = 0.0    # state variable for damping
        self.buf = np.zeros(block_size, dtype=np.float32)
        self._buf_size = block_size

    def process(self, block: np.ndarray) -> np.ndarray:
        # Resize internal buffer if block size changed (e.g. JACK period switch)
        n = len(block)
        if n != self._buf_size:
            self.buf = np.zeros(n, dtype=np.float32)
            self._buf_size = n
        self.buf[:] = block
        # Write with feedback: out[n] = in[n] + feedback * damped_delayed
        delayed = self.delay.add_to_block(self.buf, self.delay.max_len - 1, self.feedback)
        # One-pole low-pass on feedback path (vectorised)
        b = np.array([1.0 - self.damping], dtype=np.float64)
        a = np.array([1.0, -self.damping], dtype=np.float64)
        damped, _ = lfilter(b, a, delayed.astype(np.float64), zi=np.atleast_1d(self.damp_filt))
        self.damp_filt = float(damped[-1])
        self.buf[:] = block + damped.astype(np.float32)
        self.delay.write_block(self.buf)
        return self.buf


class _AllpassFilter:
    """Allpass filter for Schroeder reverb."""

    __slots__ = ("delay", "gain", "buf", "_buf_size")

    def __init__(self, delay_samples: int, block_size: int = 256):
        line_len = max(block_size * 2, delay_samples + 1)
        self.delay = _DelayLine(line_len)
        self.gain: float = 0.5
        self.buf = np.zeros(block_size, dtype=np.float32)
        self._buf_size = block_size

    def process(self, block: np.ndarray) -> np.ndarray:
        # Resize internal buffer if block size changed (e.g. JACK period switch)
        n = len(block)
        if n != self._buf_size:
            self.buf = np.zeros(n, dtype=np.float32)
            self._buf_size = n
        # out[n] = -gain * in[n] + delay[n - D] + gain * delay_output[n - D]
        # Standard allpass: out = -g * in + delayed + g * delayed_out
        # But block-wise: read delayed, write in + g * delayed, output = -g * in + delayed
        self.buf[:] = block
        delayed = self.delay.add_to_block(self.buf, self.delay.max_len - 1, self.gain)
        # Write: buf + gain * delayed
        self.buf[:] = block + delayed * self.gain
        self.delay.write_block(self.buf)
        # Output: -gain * block + delayed
        return -self.gain * block + delayed


# ── Audio Pipeline ─────────────────────────────────────────────────

class AudioPipeline:
    """Orchestrates the real-time audio FX chain.

    The pipeline processes audio block-by-block, chaining
    effect modules in order. Each module receives a numpy
    array of audio samples and returns processed samples.
    Effect state (delay buffers, LFO phases, envelope followers,
    filter memory) is stored per-instance in self._state.
    """

    def __init__(
        self,
        nam_host: Optional[NAMEngineRouter] = None,
        ir_loader: Optional[IRLoader] = None,
    ):
        self.nam = nam_host or NAMEngineRouter()
        self.ir = ir_loader or IRLoader()

        # Signal chain — list of (FXType, enabled, bypass, params)
        self._chain: list[dict] = []
        self._master_volume: float = 0.8
        self._tuner_enabled: bool = False
        self._bypassed: bool = False      # Global bypass

        # 4-Cable Method routing
        self._routing_mode: str = "mono"        # "mono" or "4cm"
        self._routing_breakpoint: int = 7       # chain index where pre/post split occurs

        # Per-block DSP state: {f"fx_{idx}": {state_dict}}
        self._state: dict[str, dict] = {}

        # Cached filter coefficients per block
        self._coeffs: dict[str, tuple] = {}

        # Runtime audio params (may be updated via set_audio_profile)
        self._block_size: int = 256
        self._sample_rate: int = 48000

        # VU meter level tracking — updated on every process() call
        # Smoothed RMS levels (0.0–1.0) read by web server for live VU meters
        self._input_level: float = 0.0
        self._output_level: float = 0.0
        # Smoothing factor: ~50ms time constant — recomputed on profile change
        self._vu_alpha: float = np.exp(-256 / (0.05 * 48000))

        # ── Tuner / pitch detection state ────────────────────────────────
        self._tuner_frequency: float = 0.0      # detected fundamental freq (Hz)
        self._tuner_note: str = "--"            # closest note name
        self._tuner_cents: float = 0.0          # cent deviation from closest note
        self._tuner_string: int = -1            # string number (1-6) or -1 if not matched
        self._tuner_confidence: float = 0.0     # 0.0 to 1.0
        # Pitch detection buffer (keep last N samples for analysis)
        self._pitch_buffer: np.ndarray = np.array([], dtype=np.float32)
        self._pitch_buffer_max: int = 2048      # ~43ms at 48kHz

        # Thread-safety lock — protects config state swapped by load_preset()
        # and read by process(). process() snapshots under this lock briefly;
        # load_preset() holds it only for the atomic swap (not during model I/O).
        self._lock = threading.Lock()

        logger.info("Audio pipeline initialized (block=%d, sr=%d)",
                     self._block_size, self._sample_rate)

    @property
    def sample_rate(self) -> int:
        """Current sample rate in Hz."""
        return self._sample_rate

    @property
    def block_size(self) -> int:
        """Current audio block size (frames per callback)."""
        return self._block_size

    def load_preset(self, preset: Preset) -> None:
        """Load a complete preset (NAM, IR, and FX chain).

        Builds the new chain list off-thread, then swaps atomically under
        ``_lock`` so that the audio thread's ``process()`` never sees a
        half-constructed chain.
        """
        new_chain: list[dict] = []
        new_routing_mode = preset.routing_mode
        new_routing_breakpoint = preset.routing_breakpoint
        new_master_volume = preset.master_volume
        new_tuner_enabled = preset.tuner_enabled

        for block in preset.chain:
            entry = {
                "fx_type": block.fx_type,
                "enabled": block.enabled,
                "bypass": block.bypass,
                "params": dict(block.params),
                "subtype": block.subtype,
            }

            # Load NAM model if needed (may do I/O — don't hold pipeline lock)
            if block.fx_type == FXType.NAM_AMP and block.nam_model_path:
                self.nam.load_model(block.nam_model_path)

            # Load IR if needed
            if block.fx_type == FXType.IR_CAB and block.ir_file_path:
                self.ir.load_ir(block.ir_file_path)

            new_chain.append(entry)

        # Atomic swap under lock — keep the window as short as possible
        with self._lock:
            self._chain = new_chain
            self._state = {}
            self._coeffs = {}
            self._master_volume = new_master_volume
            self._tuner_enabled = new_tuner_enabled
            self._routing_mode = new_routing_mode
            self._routing_breakpoint = new_routing_breakpoint

        logger.info("Preset '%s' loaded: %d blocks, routing=%s breakpoint=%d",
                     preset.name, len(self._chain),
                     self._routing_mode, self._routing_breakpoint)

    def process(self, audio_in: np.ndarray) -> np.ndarray:
        """Process a block of audio through the entire FX chain.

        Args:
            audio_in: numpy array of PCM samples (float32 [-1, 1]).
                Mono mode: shape (N,) — single audio channel.
                4CM mode:  shape (2, N) — two channels, [guitar_in, return_in].

        Returns:
            Processed audio block.
                Mono mode: shape (N,) — processed output.
                4CM mode:  shape (2, N) — [send_out, return_out].
        """
        # ── Snapshot config under lock (brief — no I/O or heavy work) ──
        with self._lock:
            tuner_enabled = self._tuner_enabled
            bypassed = self._bypassed
            routing_mode = self._routing_mode
            routing_breakpoint = self._routing_breakpoint
            master_volume = self._master_volume
            # Shallow-copy the chain list so iteration is isolated
            chain = list(self._chain)
            # Snapshot the state dict — process() mutates entries inside it
            # but load_preset() replaces the whole dict under lock
            state = self._state

        # ── Tuner mode: mute output, keep input tracking for pitch detection ──
        if tuner_enabled:
            # Still track input level for tuner display
            if audio_in.ndim == 1:
                in_rms = np.sqrt(np.mean(audio_in ** 2) + _EPS)
                self._input_level = (
                    self._input_level * self._vu_alpha
                    + in_rms * (1.0 - self._vu_alpha)
                )
            else:
                ch0 = audio_in[0, :]
                in_rms = np.sqrt(np.mean(ch0 ** 2) + _EPS)
                self._input_level = (
                    self._input_level * self._vu_alpha
                    + in_rms * (1.0 - self._vu_alpha)
                )
            # Run pitch detection on input
            self._detect_pitch(audio_in)
            return np.zeros_like(audio_in)

        if bypassed:
            return audio_in * master_volume

        if routing_mode == "4cm":
            return self._process_4cm(audio_in, master_volume, chain, state,
                                     routing_breakpoint)
        else:
            return self._process_mono(audio_in, master_volume, chain, state)

    # ── Pitch detection for tuner ──────────────────────────────────────────────

    # Standard tuning frequencies (E2, A2, D3, G3, B3, E4) — guitar strings
    _STRING_FREQS = [82.41, 110.0, 146.83, 196.0, 246.94, 329.63]
    _STRING_NAMES = ["E", "A", "D", "G", "B", "e"]

    # Note names in chromatic order (C = 0)
    _NOTE_NAMES = ["C", "C#", "D", "D#", "E", "F", "F#", "G", "G#", "A", "A#", "B"]

    def _detect_pitch(self, audio_in: np.ndarray) -> None:
        """Run pitch detection on the input buffer using autocorrelation.

        Updates ``self._tuner_frequency``, ``self._tuner_note``,
        ``self._tuner_cents``, and ``self._tuner_string``.

        Args:
            audio_in: Input audio block — mono (N,) or stereo (2, N).
        """
        # Extract mono channel
        if audio_in.ndim == 2:
            signal = audio_in[0, :].copy()
        else:
            signal = audio_in.copy()

        # Append to rolling pitch buffer
        self._pitch_buffer = np.concatenate([self._pitch_buffer, signal])
        if len(self._pitch_buffer) > self._pitch_buffer_max:
            self._pitch_buffer = self._pitch_buffer[-self._pitch_buffer_max:]

        # Need enough signal for meaningful analysis
        if len(self._pitch_buffer) < 512:
            self._tuner_confidence = 0.0
            return

        # Simple autocorrelation pitch detection
        buf = self._pitch_buffer
        # Remove DC offset
        buf = buf - np.mean(buf)

        # Check if there's enough amplitude
        rms = np.sqrt(np.mean(buf ** 2))
        if rms < 0.002:  # Silence threshold
            self._tuner_confidence = 0.0
            self._tuner_frequency = 0.0
            self._tuner_note = "--"
            self._tuner_cents = 0.0
            self._tuner_string = -1
            return

        # Autocorrelation: find the fundamental period
        # Search lag range: 30 to 1024 samples (46.9Hz to 1600Hz at 48kHz)
        min_lag = int(self._sample_rate / 1600)  # ~30
        max_lag = min(int(self._sample_rate / 50), len(buf) // 2)  # ~960

        if max_lag <= min_lag:
            self._tuner_confidence = 0.0
            return

        corr = np.correlate(buf, buf, mode='full')
        # Take only the second half (positive lags)
        corr = corr[len(corr) // 2:]

        # Simple autocorrelation
        lag_slice = corr[min_lag:max_lag + 1]

        # Find the first peak in the autocorrelation
        diffs = np.diff(lag_slice)
        # Look for zero crossings in diff (peaks: positive→negative)
        peaks = []
        for i in range(1, len(diffs)):
            if diffs[i-1] > 0 and diffs[i] <= 0:
                peaks.append((min_lag + i, lag_slice[i]))

        if not peaks:
            self._tuner_confidence = 0.0
            return

        # Pick the strongest peak
        best_lag, best_val = max(peaks, key=lambda x: x[1])

        # Confidence based on relative peak strength
        noise_floor = np.mean(np.abs(corr[min_lag:max_lag + 1]))
        confidence = best_val / (noise_floor + 1e-10)
        # Parabolic interpolation for sub-sample accuracy
        if best_lag > min_lag and best_lag < max_lag:
            idx = best_lag - min_lag
            if 0 < idx < len(lag_slice) - 1:
                y0, y1, y2 = lag_slice[idx-1], lag_slice[idx], lag_slice[idx+1]
                if y0 + y2 - 2 * y1 != 0:
                    correction = (y0 - y2) / (2 * (y0 + y2 - 2 * y1))
                    best_lag = best_lag + correction

        # Fundamental frequency
        freq = self._sample_rate / best_lag if best_lag > 0 else 0

        # Clip confidence to 0-1 range
        self._tuner_confidence = min(1.0, max(0.0, confidence / 10.0))

        if freq < 30 or freq > 1600:
            self._tuner_confidence = 0.0
            return

        self._tuner_frequency = freq

        # ── Convert frequency to note name ──
        # A4 = 440Hz, MIDI note 69
        midi_note = 12 * np.log2(freq / 440.0) + 69
        midi_rounded = round(midi_note)
        cents = int(100 * (midi_note - midi_rounded))

        # Clamp to valid MIDI range
        if midi_rounded < 0 or midi_rounded > 127:
            self._tuner_note = "--"
            self._tuner_confidence = 0.0
            return

        octave = (midi_rounded // 12) - 1
        note_idx = midi_rounded % 12
        note_name = self._NOTE_NAMES[note_idx]
        self._tuner_note = f"{note_name}{octave}"
        self._tuner_cents = cents

        # ── Guess guitar string ──
        self._tuner_string = -1
        for si, sf in enumerate(self._STRING_FREQS):
            # +/- 3 semitones from the string's fundamental
            if abs(freq - sf) / sf < 0.2:
                self._tuner_string = si + 1
                break

    def _process_mono(self, audio_in: np.ndarray,
                      master_volume: float,
                      chain: list[dict],
                      state: dict[str, dict]) -> np.ndarray:
        """Process a mono block through the full chain (all blocks)."""
        # Update input VU level (RMS with envelope smoothing)
        in_rms = np.sqrt(np.mean(audio_in ** 2) + _EPS)
        self._input_level = (
            self._input_level * self._vu_alpha
            + in_rms * (1.0 - self._vu_alpha)
        )

        buf = audio_in.copy()
        for idx, entry in enumerate(chain):
            if entry["bypass"] or not entry["enabled"]:
                continue
            buf = self._process_single_block(buf, idx, entry, state)
        out = np.clip(buf * master_volume, -1.0, 1.0)

        # Update output VU level
        out_rms = np.sqrt(np.mean(out ** 2) + _EPS)
        self._output_level = (
            self._output_level * self._vu_alpha
            + out_rms * (1.0 - self._vu_alpha)
        )

        return out

    def _process_4cm(self, audio_in: np.ndarray,
                     master_volume: float,
                     chain: list[dict],
                     state: dict[str, dict],
                     routing_breakpoint: int) -> np.ndarray:
        """Process stereo block with 4CM split routing.

        audio_in has shape (2, N):
            ch0 = guitar input (Input 1)
            ch1 = FX loop return (Input 2)

        Splits at routing_breakpoint:
            pre blocks  → ch0 processed through [0..breakpoint)
            post blocks → ch1 processed through [breakpoint..]

        Returns (2, N):
            ch0 = Send output (to amp input)
            ch1 = Return output (to amp FX return)
        """
        ch0 = audio_in[0, :].copy()
        ch1 = audio_in[1, :].copy()

        # Update input VU level from the guitar input channel (ch0)
        in_rms = np.sqrt(np.mean(ch0 ** 2) + _EPS)
        self._input_level = (
            self._input_level * self._vu_alpha
            + in_rms * (1.0 - self._vu_alpha)
        )

        bp = routing_breakpoint

        for idx, entry in enumerate(chain):
            if entry["bypass"] or not entry["enabled"]:
                continue

            if idx < bp:
                # Pre-amp block — process on guitar (ch0)
                ch0 = self._process_single_block(ch0, idx, entry, state)
            else:
                # Post-amp block — process on return (ch1)
                ch1 = self._process_single_block(ch1, idx, entry, state)

        out = np.zeros_like(audio_in)
        out[0, :] = np.clip(ch0 * master_volume, -1.0, 1.0)
        out[1, :] = np.clip(ch1 * master_volume, -1.0, 1.0)

        # Update output VU level from the processed effect return (ch1)
        out_rms = np.sqrt(np.mean(out ** 2) + _EPS)
        self._output_level = (
            self._output_level * self._vu_alpha
            + out_rms * (1.0 - self._vu_alpha)
        )

        return out

    def _process_single_block(self, buf: np.ndarray, idx: int,
                              entry: dict,
                              state: dict[str, dict]) -> np.ndarray:
        """Process a single mono audio block through one FX block.

        Args:
            buf: Mono audio block (N,) to process.
            idx: Chain index for state lookup.
            entry: Chain entry dict with fx_type, params.
            state: Per-block DSP state dict (snapshotted by ``process()``).

        Returns:
            Processed mono block (N,).
        """
        fx_type = entry["fx_type"]
        params = dict(entry["params"])  # copy so dispatchers can safely inject subtype
        # Inject entry-level subtype into params (only when set on FXBlock),
        # so existing dispatchers that read params.get("subtype", ...) get it.
        # This is backward-compatible: if FXBlock.subtype is empty (default),
        # any legacy subtype set in params is preserved.
        subtype = entry.get("subtype", "")
        if subtype:
            params["subtype"] = subtype
        fx_state = state.setdefault(f"fx_{idx}", {})

        match fx_type:
            case FXType.NOISE_GATE:
                return self._apply_gate(buf, params, fx_state)
            case FXType.COMPRESSOR:
                return self._apply_compressor(buf, params, fx_state)
            case FXType.BOOST:
                return self._apply_boost(buf, params, fx_state)
            case FXType.OVERDRIVE:
                subtype = params.get("subtype", "ts808")
                match subtype:
                    case "ts808":
                        return self._apply_overdrive(buf, params, fx_state)
                    case "klon":
                        return self._apply_klon(buf, params, fx_state)
                    case "bd2":
                        return self._apply_bd2(buf, params, fx_state)
                    case _:
                        logger.warning("Unknown OVERDRIVE subtype '%s', falling back to ts808", subtype)
                        return self._apply_overdrive(buf, params, fx_state)
            case FXType.DISTORTION:
                subtype = params.get("subtype", "rat")
                match subtype:
                    case "rat":
                        return self._apply_distortion(buf, params, fx_state)
                    case _:
                        logger.warning("Unknown DISTORTION subtype '%s', falling back to rat", subtype)
                        return self._apply_distortion(buf, params, fx_state)
            case FXType.FUZZ:
                subtype = params.get("subtype", "fuzz")
                match subtype:
                    case "fuzz":
                        return self._apply_fuzz(buf, params, fx_state)
                    case "muff":
                        return self._apply_muff(buf, params, fx_state)
                    case _:
                        logger.warning("Unknown FUZZ subtype '%s', falling back to fuzz", subtype)
                        return self._apply_fuzz(buf, params, fx_state)
            case FXType.EQ:
                return self._apply_eq(buf, params, fx_state)
            case FXType.CHORUS:
                return self._apply_chorus(buf, params, fx_state)
            case FXType.FLANGER:
                return self._apply_flanger(buf, params, fx_state)
            case FXType.PHASER:
                return self._apply_phaser(buf, params, fx_state)
            case FXType.TREMOLO:
                return self._apply_tremolo(buf, params, fx_state)
            case FXType.VIBRATO:
                return self._apply_vibrato(buf, params, fx_state)
            case FXType.DELAY:
                return self._apply_delay(buf, params, fx_state)
            case FXType.REVERB:
                return self._apply_reverb(buf, params, fx_state)
            case FXType.VOLUME:
                return np.clip(self._apply_volume(buf, params, fx_state), -1.0, 1.0)
            case FXType.NAM_AMP:
                # Use C++ NeuralAudio engine when a .nam file is loaded.
                if self.nam.is_loaded:
                    # Apply input gain as pre-amp drive (level 0.0-1.0 maps to 0-1.0x gain)
                    level = params.get("level", 0.75)
                    drive = np.clip(level, 0.0, 1.0)
                    if drive != 1.0:
                        buf = buf * drive

                    # Single process call through the C++ engine
                    processed = self.nam.process(buf)

                    # Crossfade on preset switch
                    if self.nam._crossfade_buf is not None:
                        processed = self.nam.apply_crossfade(processed)

                    # Clip output to prevent digital distortion
                    return np.clip(processed, -1.0, 1.0)

                logger.debug("NAM_AMP: engine not loaded, passing through")
                return buf
            case FXType.IR_CAB:
                if self.ir.is_loaded:
                    return self._apply_ir_cab(buf, params, fx_state)
                return buf
            # ── Pitch & Frequency ──────────────────────────────────────
            case FXType.OCTAVER:
                return self._apply_octaver(buf, params, fx_state)
            case FXType.PITCH_SHIFTER:
                return self._apply_pitch_shifter(buf, params, fx_state)
            case FXType.HARMONIZER:
                return self._apply_harmonizer(buf, params, fx_state)
            case FXType.WHAMMY:
                return self._apply_whammy(buf, params, fx_state)
            case FXType.DETUNE:
                return self._apply_detune(buf, params, fx_state)
            # ── Modulation ─────────────────────────────────────────────
            case FXType.RING_MODULATOR:
                return self._apply_ring_modulator(buf, params, fx_state)
            case FXType.AUTO_WAH:
                return self._apply_auto_wah(buf, params, fx_state)
            case FXType.ENVELOPE_FILTER:
                return self._apply_envelope_filter(buf, params, fx_state)
            case FXType.ROTARY_SPEAKER:
                return self._apply_rotary_speaker(buf, params, fx_state)
            case FXType.UNI_VIBE:
                return self._apply_uni_vibe(buf, params, fx_state)
            case FXType.AUTO_PAN:
                return self._apply_auto_pan(buf, params, fx_state)
            case FXType.STEREO_WIDENER:
                return self._apply_stereo_widener(buf, params, fx_state)
            # ── Drive & Saturation ─────────────────────────────────────
            case FXType.BITCRUSHER:
                return self._apply_bitcrusher(buf, params, fx_state)
            case FXType.WAVEFOLDER:
                return self._apply_wavefolder(buf, params, fx_state)
            case FXType.RECTIFIER:
                return self._apply_rectifier(buf, params, fx_state)
            # ── Dynamics ───────────────────────────────────────────────
            case FXType.EXPANDER:
                return self._apply_expander(buf, params, fx_state)
            case FXType.DE_ESSER:
                return self._apply_de_esser(buf, params, fx_state)
            case FXType.TRANSIENT_SHAPER:
                return self._apply_transient_shaper(buf, params, fx_state)
            case FXType.SIDECHAIN_COMPRESSOR:
                return self._apply_sidechain_compressor(buf, params, fx_state)
            # ── Filters & EQ ──────────────────────────────────────────
            case FXType.PARAMETRIC_EQ:
                return self._apply_parametric_eq(buf, params, fx_state)
            case FXType.HIGH_PASS_FILTER:
                return self._apply_hpf(buf, params, fx_state)
            case FXType.LOW_PASS_FILTER:
                return self._apply_lpf(buf, params, fx_state)
            case FXType.BAND_PASS_FILTER:
                return self._apply_bpf(buf, params, fx_state)
            case FXType.NOTCH_FILTER:
                return self._apply_notch(buf, params, fx_state)
            case FXType.FORMANT_FILTER:
                return self._apply_formant_filter(buf, params, fx_state)
            # ── Time-Based ─────────────────────────────────────────────
            case FXType.PING_PONG_DELAY:
                return self._apply_ping_pong_delay(buf, params, fx_state)
            case FXType.MULTI_TAP_DELAY:
                return self._apply_multi_tap_delay(buf, params, fx_state)
            case FXType.REVERSE_DELAY:
                return self._apply_reverse_delay(buf, params, fx_state)
            case FXType.TAPE_ECHO:
                return self._apply_tape_echo(buf, params, fx_state)
            case FXType.SHIMMER_REVERB:
                return self._apply_shimmer_reverb(buf, params, fx_state)
            case FXType.LOOPER:
                return self._apply_looper(buf, params, fx_state)
            # ── Ambience ──────────────────────────────────────────────
            case FXType.EARLY_REFLECTIONS:
                return self._apply_early_reflections(buf, params, fx_state)
            case _:
                return buf

    # ── LFO helpers ─────────────────────────────────────────────────

    def _lfo_phase(self, rate_hz: float, state: dict, block_size: int) -> np.ndarray:
        """Generate LFO phase ramp (0->1), update state."""
        phase = state.get("phase", 0.0)
        delta = rate_hz / self._sample_rate
        t = np.arange(block_size, dtype=np.float64) * delta + phase
        t %= 1.0
        state["phase"] = float(t[-1] + delta) % 1.0
        return t

    @staticmethod
    def _lfo_wave(phase: np.ndarray, shape: str = "sine") -> np.ndarray:
        """Generate LFO waveform from phase array."""
        match shape:
            case "sine":
                return 0.5 + 0.5 * np.sin(2 * np.pi * phase)
            case "triangle":
                return 2.0 * np.abs(2.0 * phase - 1.0) - 1.0
                # Returns in [-1, 1]; normalise below
            case "square":
                return np.where(phase < 0.5, 1.0, 0.0)
            case _:
                return 0.5 + 0.5 * np.sin(2 * np.pi * phase)

    # ── Effect implementations ──────────────────────────────────────

    # ── 1. Noise Gate ───────────────────────────────────────────────

    def _apply_gate(self, buf: np.ndarray, params: dict,
                    state: dict) -> np.ndarray:
        """Noise gate with adjustable threshold and release envelope."""
        threshold = params.get("threshold", 0.01)
        release_ms = params.get("release", 100.0)

        envelope = state.get("envelope", 0.0)
        rms = np.sqrt(np.mean(buf ** 2) + _EPS)

        if rms >= threshold:
            # Instant attack
            envelope = rms
        else:
            # Exponential release — time constant per block
            release_coeff = np.exp(-self._block_size / (release_ms * self._sample_rate / 1000.0))
            envelope = envelope * release_coeff + rms * (1.0 - release_coeff)

        state["envelope"] = envelope

        if envelope < threshold:
            return np.zeros_like(buf)
        return buf

    # ── 2. Compressor ───────────────────────────────────────────────

    def _apply_compressor(self, buf: np.ndarray, params: dict,
                          state: dict) -> np.ndarray:
        """Compressor with threshold (dB), ratio, attack, release, make-up gain."""
        threshold_db = params.get("threshold", -20.0)  # dB
        ratio = params.get("ratio", 3.0)
        attack_ms = params.get("attack", 5.0)
        release_ms = params.get("release", 100.0)
        makeup = params.get("gain", 1.0)

        # RMS envelope with attack/release shaping
        rms = np.sqrt(np.mean(buf ** 2) + _EPS)
        envelope = state.get("envelope", 0.0)

        if rms > envelope:
            alpha = np.exp(-self._block_size / (attack_ms * self._sample_rate / 1000.0))
        else:
            alpha = np.exp(-self._block_size / (release_ms * self._sample_rate / 1000.0))

        envelope = envelope * alpha + rms * (1.0 - alpha)
        state["envelope"] = envelope

        # Compute gain reduction in dB domain
        if envelope > 1e-10:
            env_db = 20.0 * np.log10(envelope)
        else:
            env_db = -120.0

        if env_db > threshold_db:
            # gain_db = threshold + (env - threshold) / ratio - env
            gain_db = threshold_db + (env_db - threshold_db) / ratio - env_db
        else:
            gain_db = 0.0

        gain_lin = 10 ** (gain_db / 20.0)
        return np.clip(buf * gain_lin * makeup, -1.0, 1.0)

    # ── 3. Boost / Overdrive / Distortion / Fuzz ────────────────────

    def _apply_boost(self, buf: np.ndarray, params: dict,
                     state: dict) -> np.ndarray:
        """Clean boost with linear gain."""
        gain_db = params.get("gain_db", 6.0)
        gain_linear = 10 ** (gain_db / 20.0)
        return np.clip(buf * gain_linear, -1.0, 1.0)

    def _apply_overdrive(self, buf: np.ndarray, params: dict,
                         state: dict) -> np.ndarray:
        """Tube-style overdrive with asymmetric soft clipping."""
        drive = params.get("drive", 0.5)
        tone = params.get("tone", 0.5)
        gain = params.get("gain", 1.0)

        drive_scaled = drive * 15.0 + 1.0
        shaped = buf * drive_scaled

        # Asymmetric soft clipping (tube-like)
        # Positive half clips softer than negative (tube asymmetry)
        pos = np.where(shaped > 0, shaped / (1.0 + shaped * 0.3), shaped)
        neg = np.where(pos < 0, pos / (1.0 - pos * 0.5), pos)
        out = np.tanh(neg)  # Final polish with tanh

        return np.clip(out * gain, -1.0, 1.0)

    def _apply_distortion(self, buf: np.ndarray, params: dict,
                          state: dict) -> np.ndarray:
        """Harder asymmetric clipping with diode-style transfer."""
        drive = params.get("drive", 0.7)
        tone = params.get("tone", 0.5)
        gain = params.get("gain", 1.0)

        drive_scaled = drive * 30.0 + 1.0
        shaped = buf * drive_scaled

        # Diode-style asymmetric clipping
        clipped = np.where(
            shaped > 0,
            np.clip(shaped, 0, 0.8) / (1.0 + np.abs(np.clip(shaped, 0, 0.8)) * 0.5),
            np.clip(shaped, -0.6, 0) / (1.0 + np.abs(np.clip(shaped, -0.6, 0)) * 0.3),
        )
        return np.clip(clipped * gain, -1.0, 1.0)

    def _apply_fuzz(self, buf: np.ndarray, params: dict,
                    state: dict) -> np.ndarray:
        """Octave-fuzzy hard clipping with gated sustain."""
        drive = params.get("drive", 0.8)
        tone = params.get("tone", 0.5)
        gain = params.get("gain", 1.0)

        drive_scaled = drive * 50.0 + 1.0
        shaped = buf * drive_scaled

        # Hard square-wave clip with asymmetric gate
        clipped = np.sign(shaped) * (1.0 - np.exp(-np.abs(shaped) * 2.0))
        # Foldover for octave effect
        folded = np.abs(clipped) * 0.3 + clipped * 0.7
        return np.clip(folded * gain, -1.0, 1.0)

    # ── 3a. Klon Centaur (subtype: klon) ────────────────────────────

    def _apply_klon(self, buf: np.ndarray, params: dict,
                    state: dict) -> np.ndarray:
        """Klon Centaur-style transparent overdrive with clean blend.

        The Klon's hallmark is its ability to mix a clean (buffered) signal
        with a lightly overdriven signal, preserving pick attack and dynamics.
        Uses symmetrical soft clipping (germanium diode-style) with a
        high-cut tone control on the drive path only.

        Parameters:
            drive (0-1): Overdrive gain.
            tone  (0-1): Treble cut — 0=bright, 1=dark.
            gain  (0-1): Output level recovery.
            blend (0-1): Dry/wet mix — 0=clean only, 1=drive only.
                         Default 0.5 (~50/50 blend, original Klon character).
        """
        drive = params.get("drive", 0.5)
        tone = params.get("tone", 0.5)
        gain = params.get("gain", 1.0)
        blend = params.get("blend", 0.5)

        # Pre-gain (Klon uses moderate gain staging)
        drive_scaled = drive * 10.0 + 1.0
        drive_path = buf * drive_scaled

        # Symmetrical soft clipping (germanium diode character)
        clipped = np.tanh(drive_path * 2.0)

        # One-pole LPF on drive path for tone control (treble cut)
        tone_cut = 1.0 - tone  # 1 = max cut
        if tone_cut > 0.001:
            fc = 2000.0 + (1.0 - tone_cut) * 18000.0  # 2kHz to 20kHz
            omega = 2.0 * np.pi * fc / self._sample_rate
            a0 = 1.0 + omega  # one-pole approximation
            b0 = omega / a0
            a1 = (1.0 - omega) / a0
            # State tracking for the one-pole
            lp_key = "klon_lp_zi"
            zi = state.get(lp_key, 0.0)
            clipped_filtered = np.zeros_like(clipped)
            for i in range(len(clipped)):
                clipped_filtered[i] = b0 * clipped[i] + a1 * zi
                zi = clipped_filtered[i]
            state[lp_key] = zi
            clipped = clipped_filtered

        # Clean blend: mix clean and overdrive signals
        out = (1.0 - blend) * buf + blend * clipped

        # Output level
        out = out * gain
        return np.clip(out, -1.0, 1.0)

    # ── 3b. Blues Driver (subtype: bd2) ─────────────────────────────

    def _apply_bd2(self, buf: np.ndarray, params: dict,
                   state: dict) -> np.ndarray:
        """Blues Driver-style soft clipping with responsive tone stack.

        The Boss BD-2 uses a two-stage asymmetric clipping topology with an
        active tone control that can both cut bass/boost treble, giving it
        a very wide tonal range from warm low-gain to biting high-gain.

        Parameters:
            drive (0-1): Drive gain.
            tone  (0-1): Active tone control — 0=warm (bass), 1=bright (treble).
            gain  (0-1): Output level recovery.
        """
        drive = params.get("drive", 0.5)
        tone = params.get("tone", 0.5)
        gain = params.get("gain", 1.0)

        # Pre-gain boost
        drive_scaled = drive * 12.0 + 1.0
        shaped = buf * drive_scaled

        # Two-stage asymmetric clipping
        # Stage 1: moderate soft clip (asymmetric — positive softer)
        pos = np.where(shaped > 0, np.tanh(shaped * 1.2), shaped)
        stage1 = np.where(pos < 0, pos / (1.0 - pos * 0.2), pos)

        # Stage 2: harder clip on positive side for BD-2 character
        stage2 = np.tanh(stage1 * 2.0)

        # Active tone control: shelving EQ shaped by tone param
        # tone=0: bass boost / treble cut
        # tone=0.5: flat
        # tone=1.0: treble boost / bass cut
        if abs(tone - 0.5) > 0.01:
            # Map tone to gain: -6dB to +6dB
            shelf_gain_db = (tone - 0.5) * 12.0  # -6 to +6 dB
            # Treble shelf (3.5kHz, Q=0.7)
            coeffs = state.get("bd2_tshelf_coeffs")
            tag = round(shelf_gain_db, 2)
            if coeffs is None or state.get("bd2_tshelf_tag") != tag:
                coeffs = _compute_highshelf_coeffs(3500.0, shelf_gain_db, 0.7, self._sample_rate)
                state["bd2_tshelf_coeffs"] = coeffs
                state["bd2_tshelf_tag"] = tag

            b0, b1, b2, a1, a2 = coeffs
            b = np.array([b0, b1, b2], dtype=np.float64)
            a = np.array([1.0, a1, a2], dtype=np.float64)
            zi = state.get("bd2_tshelf_zi", np.zeros(2, dtype=np.float64))
            sig, zf = lfilter(b, a, stage2.astype(np.float64, copy=False), zi=zi)
            state["bd2_tshelf_zi"] = zf
            stage2 = np.clip(sig, -1.0, 1.0).astype(np.float32)

        # Output level
        out = np.clip(stage2 * gain, -1.0, 1.0)
        return out

    # ── 3c. Big Muff Pi (subtype: muff) ─────────────────────────────

    def _apply_muff(self, buf: np.ndarray, params: dict,
                    state: dict) -> np.ndarray:
        """Big Muff Pi-style fuzz with tone stack.

        The Big Muff uses three cascaded gain stages with clipping diodes
        between each stage, creating massive sustain, followed by a
        passive tone stack (bass cut + mid scoop + treble cut) and a
        volume recovery stage.

        Parameters:
            sustain (0-1): Gain/fuzz amount (the "Sustain" knob).
            tone    (0-1): Tone stack position —
                           0=dark (bass), 0.5=mid-scoop (classic), 1=bright (treble).
            volume  (0-1): Output volume recovery.
        """
        sustain = params.get("sustain", 0.5)
        tone = params.get("tone", 0.5)
        volume = params.get("volume", 0.7)

        # Three-stage gain with inter-stage soft-clipping (1N4148 diode style)
        # Stage 1
        g1 = sustain * 20.0 + 1.0
        s1 = np.tanh(buf * g1)
        # Stage 2
        g2 = sustain * 15.0 + 1.0
        s2 = np.tanh(s1 * g2)
        # Stage 3
        g3 = sustain * 10.0 + 1.0
        s3 = np.tanh(s2 * g3)

        # Big Muff tone stack: three-band passive EQ
        # Maps tone param (0-1) across the classic tone sweep:
        # 0.0 = bass-heavy (low-pass dominant)
        # 0.5 = mid-scoop (notch around 1kHz)
        # 1.0 = treble-heavy (high-pass dominant)
        if tone < 0.5:
            # Bass range: low-pass emphasis
            blend = tone * 2.0  # 0.0 -> 1.0
            bass_gain = 6.0 * (1.0 - blend)
            treble_gain = -6.0 * blend
        elif tone > 0.5:
            # Treble range: high-pass emphasis
            blend = (tone - 0.5) * 2.0  # 0.0 -> 1.0
            bass_gain = -6.0 * blend
            treble_gain = 6.0 * (1.0 - blend)
        else:
            # Flat response (tone at 0.5 = minimal EQ)
            bass_gain = 0.0
            treble_gain = 0.0

        sig = s3.astype(np.float64, copy=False)

        # Bass shelf (200Hz)
        if abs(bass_gain) > 0.5:
            coeffs = state.get("muff_bass_coeffs")
            tag = (round(bass_gain, 1), round(tone, 2))
            if coeffs is None or state.get("muff_bass_tag") != tag:
                coeffs = _compute_lowshelf_coeffs(200.0, bass_gain, 0.707, self._sample_rate)
                state["muff_bass_coeffs"] = coeffs
                state["muff_bass_tag"] = tag
            b0, b1, b2, a1, a2 = coeffs
            b = np.array([b0, b1, b2], dtype=np.float64)
            a = np.array([1.0, a1, a2], dtype=np.float64)
            zi = state.get("muff_bass_zi", np.zeros(2, dtype=np.float64))
            sig, zf = lfilter(b, a, sig, zi=zi)
            state["muff_bass_zi"] = zf

        # Treble shelf (3kHz)
        if abs(treble_gain) > 0.5:
            coeffs = state.get("muff_treb_coeffs")
            tag = (round(treble_gain, 1), round(tone, 2))
            if coeffs is None or state.get("muff_treb_tag") != tag:
                coeffs = _compute_highshelf_coeffs(3000.0, treble_gain, 0.707, self._sample_rate)
                state["muff_treb_coeffs"] = coeffs
                state["muff_treb_tag"] = tag
            b0, b1, b2, a1, a2 = coeffs
            b = np.array([b0, b1, b2], dtype=np.float64)
            a = np.array([1.0, a1, a2], dtype=np.float64)
            zi = state.get("muff_treb_zi", np.zeros(2, dtype=np.float64))
            sig, zf = lfilter(b, a, sig, zi=zi)
            state["muff_treb_zi"] = zf

        out = np.clip(np.clip(sig, -1.0, 1.0).astype(np.float32) * volume, -1.0, 1.0)
        return out

    # ── 4. Three-band EQ ────────────────────────────────────────────

    def _apply_eq(self, buf: np.ndarray, params: dict,
                  state: dict) -> np.ndarray:
        """3-band EQ: bass shelf, mid peaking, treble shelf.

        Uses scipy.signal.lfilter with persistent state for
        zero-crosstalk between blocks.
        """
        bass_gain = params.get("bass", 0.0)       # dB
        mid_gain = params.get("mid", 0.0)          # dB
        treble_gain = params.get("treble", 0.0)    # dB
        bass_freq = params.get("bass_freq", 200.0)
        mid_freq = params.get("mid_freq", 1000.0)
        treble_freq = params.get("treble_freq", 3500.0)
        q = params.get("q", 0.707)

        sig = buf.astype(np.float64, copy=False)

        for band_name, freq, gain_db, compute_fn in [
            ("bass", bass_freq, bass_gain, _compute_lowshelf_coeffs),
            ("mid", mid_freq, mid_gain, _compute_peaking_coeffs),
            ("treble", treble_freq, treble_gain, _compute_highshelf_coeffs),
        ]:
            if gain_db == 0.0:
                continue
            key = f"eq_{band_name}"
            coeffs = state.get(f"{key}_coeffs")
            param_tag = (bass_freq, mid_freq, treble_freq, bass_gain, mid_gain, treble_gain, q)
            if coeffs is None or state.get(f"{key}_tag") != param_tag:
                coeffs = compute_fn(freq, gain_db, q, self._sample_rate)
                state[f"{key}_coeffs"] = coeffs
                state[f"{key}_tag"] = param_tag

            b0, b1, b2, a1, a2 = coeffs
            b = np.array([b0, b1, b2], dtype=np.float64)
            a = np.array([1.0, a1, a2], dtype=np.float64)
            zi = state.get(f"{key}_zi", np.zeros(2, dtype=np.float64))

            sig, zf = lfilter(b, a, sig, zi=zi)
            state[f"{key}_zi"] = zf

        return np.clip(sig, -1.0, 1.0).astype(np.float32)

    # ── 5. Chorus ───────────────────────────────────────────────────

    def _apply_chorus(self, buf: np.ndarray, params: dict,
                      state: dict) -> np.ndarray:
        """Chorus with LFO-driven modulated delay line (stereo-ish)."""
        rate = params.get("rate", 0.5)        # Hz
        depth = params.get("depth", 0.5)      # 0.0-1.0
        mix = params.get("mix", 0.5)          # wet/dry
        delay_base = params.get("delay", 20.0)  # ms (typical chorus: 15-30ms)

        base_samples = delay_base * self._sample_rate / 1000.0
        mod_range = depth * 5.0 * self._sample_rate / 1000.0

        if "delay" not in state:
            max_d = int(base_samples + mod_range + 10.0 * self._sample_rate / 1000.0) + 1
            state["delay"] = _DelayLine(max_d)
            state["delay"].write_block(np.zeros(max_d))

        delay_line: _DelayLine = state["delay"]
        phase = self._lfo_phase(rate, state, len(buf))
        lfo = self._lfo_wave(phase, "sine")
        mod_delay = base_samples + lfo * mod_range

        # Vectorised read
        wet = delay_line.read_block_varying(mod_delay)

        delay_line.write_block(buf)

        return np.clip(buf * (1.0 - mix) + wet * mix, -1.0, 1.0)

    # ── 6. Flanger ──────────────────────────────────────────────────

    def _apply_flanger(self, buf: np.ndarray, params: dict,
                       state: dict) -> np.ndarray:
        """Flanger with swept comb filter and feedback."""
        rate = params.get("rate", 0.25)       # Hz (slower than chorus)
        depth = params.get("depth", 0.7)      # 0.0-1.0
        feedback = params.get("feedback", 0.3)
        mix = params.get("mix", 0.5)          # wet/dry
        delay_base = params.get("delay", 5.0)  # ms (typical flanger: 1-10ms)

        base_samples = delay_base * self._sample_rate / 1000.0
        mod_range = depth * 5.0 * self._sample_rate / 1000.0

        if "delay" not in state:
            max_d = int(base_samples + mod_range + 10.0 * self._sample_rate / 1000.0) + 1
            state["delay"] = _DelayLine(max_d)
            state["delay"].write_block(np.zeros(max_d))

        delay_line: _DelayLine = state["delay"]
        phase = self._lfo_phase(rate, state, len(buf))
        lfo = self._lfo_wave(phase, "sine")
        mod_delay = base_samples + lfo * mod_range

        # Feedback buffer
        feedback_buf = state.get("fb_buf", np.zeros(len(buf), dtype=np.float32))

        # Blend feedback into input
        fb_input = buf + feedback_buf * feedback

        # Vectorised read
        wet = delay_line.read_block_varying(mod_delay)

        delay_line.write_block(fb_input)

        # Store feedback for next block
        state["fb_buf"] = wet * 0.5

        return np.clip(buf * (1.0 - mix) + wet * mix, -1.0, 1.0)

    # ── 7. Phaser ───────────────────────────────────────────────────

    def _apply_phaser(self, buf: np.ndarray, params: dict,
                      state: dict) -> np.ndarray:
        """Phaser with allpass filter cascade and feedback."""
        rate = params.get("rate", 0.4)        # Hz
        depth = params.get("depth", 0.5)      # 0.0-1.0
        feedback = params.get("feedback", 0.3)
        mix = params.get("mix", 0.5)
        stages = int(params.get("stages", 4))

        # Map LFO to centre frequency sweep: 200-2000 Hz
        phase = self._lfo_phase(rate, state, len(buf))
        lfo = self._lfo_wave(phase, "sine")
        freq_range = 200.0 + lfo * depth * 1800.0

        fb_buf = state.get("fb_buf", np.zeros(len(buf), dtype=np.float64))
        fb_input = buf.astype(np.float64, copy=False) + fb_buf * feedback

        sig = fb_input.copy()
        for stage in range(stages):
            # Allpass as first-order IIR: H(z) = (coeff + z^-1) / (1 + coeff * z^-1)
            # Which is lfilter(b=[coeff, 1], a=[1, coeff], ...)
            # But coeff varies per sample (LFO-driven)! Can't use lfilter directly.
            # Use block-constant approximation: one coeff per block at LFO centre.
            freq = np.mean(freq_range)
            w = 2.0 * np.pi * freq / self._sample_rate
            tan_half_w = np.tan(w / 2.0)
            coeff = (1.0 - tan_half_w) / (1.0 + tan_half_w)

            b = np.array([coeff, 1.0], dtype=np.float64)
            a = np.array([1.0, coeff], dtype=np.float64)
            zi = state.get(f"ap_zi_{stage}", np.zeros(1, dtype=np.float64))
            sig, zf = lfilter(b, a, sig, zi=zi)
            state[f"ap_zi_{stage}"] = zf

        state["fb_buf"] = sig * 0.5
        sig = np.clip(sig, -1.0, 1.0)
        return (buf * (1.0 - mix) + sig.astype(np.float32) * mix).astype(np.float32)

    # ── 8. Tremolo ──────────────────────────────────────────────────

    def _apply_tremolo(self, buf: np.ndarray, params: dict,
                       state: dict) -> np.ndarray:
        """Tremolo with configurable LFO shape."""
        rate = params.get("rate", 4.0)        # Hz
        depth = params.get("depth", 0.7)
        shape = params.get("shape", "sine")   # sine / triangle / square

        phase = self._lfo_phase(rate, state, len(buf))
        lfo = self._lfo_wave(phase, shape)

        # LFO is 0-1; tremolo scales between full volume and attenuated
        mod = 1.0 - depth * (1.0 - lfo)
        return buf * mod

    # ── 9. Vibrato ──────────────────────────────────────────────────

    def _apply_vibrato(self, buf: np.ndarray, params: dict,
                       state: dict) -> np.ndarray:
        """Vibrato — modulated delay with 100% wet (pitch modulation)."""
        rate = params.get("rate", 3.0)        # Hz
        depth = params.get("depth", 0.5)      # cents equivalent

        base_samples = 2.0 * self._sample_rate / 1000.0  # fixed ~2ms base
        mod_range = depth * 3.0 * self._sample_rate / 1000.0

        if "delay" not in state:
            max_d = int(base_samples + mod_range + 5.0 * self._sample_rate / 1000.0) + 1
            state["delay"] = _DelayLine(max_d)
            state["delay"].write_block(np.zeros(max_d))

        delay_line: _DelayLine = state["delay"]
        phase = self._lfo_phase(rate, state, len(buf))
        lfo = self._lfo_wave(phase, "sine")
        mod_delay = base_samples + lfo * mod_range

        wet = delay_line.read_block_varying(mod_delay)
        delay_line.write_block(buf)
        return wet

    # ── 10. Delay ───────────────────────────────────────────────────

    def _apply_delay(self, buf: np.ndarray, params: dict,
                     state: dict) -> np.ndarray:
        """Delay dispatcher — routes to subtype-specific implementation.

        Subtype is read from ``params["subtype"]``, defaulting to ``"digital"``.
        Available subtypes: ``digital``, ``analog``, ``ping_pong``, ``tape``.
        """
        subtype = params.get("subtype", "digital")
        match subtype:
            case "analog":
                return self._apply_analog_delay(buf, params, state)
            case "ping_pong":
                return self._apply_ping_pong_delay(buf, params, state)
            case "tape":
                return self._apply_tape_echo(buf, params, state)
            case _:
                return self._apply_digital_delay(buf, params, state)

    def _apply_digital_delay(self, buf: np.ndarray, params: dict,
                             state: dict) -> np.ndarray:
        """Digital delay with feedback and tap-tempo support."""
        time_ms = params.get("time", 400.0)
        feedback = params.get("feedback", 0.3)
        mix = params.get("mix", 0.4)
        tap_tempo = params.get("tap_tempo", 0.0)

        # Tap tempo overrides time_ms when > 0
        if tap_tempo > 0:
            time_ms = tap_tempo

        delay_samples = int(time_ms * self._sample_rate / 1000.0)

        if "delay" not in state:
            # Allocate 2x requested delay for headroom
            max_d = max(delay_samples * 2, self._sample_rate)  # at least 1s
            state["delay"] = _DelayLine(max_d + 1)
            state["delay"].write_block(np.zeros(max_d // 2))

        delay_line: _DelayLine = state["delay"]

        # Read delayed signal
        wet = delay_line.read_block(float(delay_samples), len(buf))

        # Write dry + clipped feedback to prevent delay-line runaway
        fb_gain = min(feedback, 0.98)
        write_sig = np.clip(buf + wet * fb_gain, -1.0, 1.0)
        delay_line.write_block(write_sig)

        return np.clip(buf * (1.0 - mix) + wet * mix, -1.0, 1.0)

    def _apply_analog_delay(self, buf: np.ndarray, params: dict,
                            state: dict) -> np.ndarray:
        """BBD-style analog delay with low-pass filtering in the feedback path.

        Each repeat gets progressively darker (less treble) — the classic
        warm, murky analog delay sound. Uses a one-pole low-pass filter
        in the feedback loop with subtle BBD-style saturation on the wet path.

        Params:
            time (float): delay time in ms (default 400.0)
            feedback (float): feedback amount 0.0-1.0 (default 0.3)
            mix (float): wet/dry blend 0.0-1.0 (default 0.4)
            tone (float): feedback brightness 0.0-1.0 (default 0.5)
                0.0 = very dark (heavy LPF), 1.0 = brighter (less LPF)
        """
        time_ms = params.get("time", 400.0)
        feedback = params.get("feedback", 0.3)
        mix = params.get("mix", 0.4)
        tone = params.get("tone", 0.5)  # 0.0=dark, 1.0=bright

        delay_samples = int(time_ms * self._sample_rate / 1000.0)

        if "delay" not in state:
            max_d = max(delay_samples * 2, self._sample_rate)
            state["delay"] = _DelayLine(max_d + 1)
            state["delay"].write_block(np.zeros(max_d // 2))

        delay_line: _DelayLine = state["delay"]

        # Read delayed signal
        wet = delay_line.read_block(float(delay_samples), len(buf))

        # ── Low-pass filter on feedback path (darker repeats) ────────
        # Map tone to cutoff: 0.0 → ~500Hz (very dark), 1.0 → ~12kHz (bright)
        # cutoff_factor is the one-pole coefficient (0.1-0.99)
        cutoff_factor = 0.1 + tone * 0.89

        lp_z = state.get("lp_z", 0.0)
        lp_out = np.zeros_like(wet)
        for i in range(len(wet)):
            lp_z = wet[i] * (1.0 - cutoff_factor) + lp_z * cutoff_factor
            lp_out[i] = lp_z
        state["lp_z"] = float(lp_z)

        # ── BBD-style subtle saturation on feedback path ─────────────
        # Soft-clip the filtered feedback (BBD companding characteristic)
        lp_out = np.tanh(lp_out * 0.5)

        # Write with clipped feedback
        fb_gain = min(feedback, 0.98)
        write_sig = np.clip(buf + lp_out * fb_gain, -1.0, 1.0)
        delay_line.write_block(write_sig)

        return np.clip(buf * (1.0 - mix) + wet * mix, -1.0, 1.0)

    # ── 11. Reverb (subtype dispatch) ───────────────────────────────

    def _apply_reverb(self, buf: np.ndarray, params: dict,
                      state: dict) -> np.ndarray:
        """Reverb dispatcher — selects algorithm by ``params['subtype']``.

        Supported subtypes:
            'hall'   — classic Schroeder reverb (8 comb + 4 allpass) [default]
            'spring' — multi-spring delay-line model (metallic / boingy)
            'plate'  — dense plate reverb (smooth, rich tail)
            'room'   — room reverb (early reflections + diffuse late tail)

        Backward-compatible: existing presets without ``subtype`` default to 'hall'.
        """
        subtype = params.get("subtype", "hall")
        mix = params.get("mix", 0.3)

        match subtype:
            case "spring":
                wet = self._reverb_spring(buf, params, state)
            case "plate":
                wet = self._reverb_plate(buf, params, state)
            case "room":
                wet = self._reverb_room(buf, params, state)
            case _:
                wet = self._reverb_hall(buf, params, state)

        return np.clip(buf * (1.0 - mix) + wet * mix, -1.0, 1.0)

    def _reverb_hall(self, buf: np.ndarray, params: dict,
                     state: dict) -> np.ndarray:
        """Hall reverb — classic Schroeder with 8 comb + 4 allpass."""
        decay = params.get("decay", 0.5)
        damping = params.get("damping", 0.4)
        predelay_ms = params.get("predelay", 30.0)

        if "combs" not in state:
            comb_delays = [29, 37, 44, 50, 31, 39, 47, 53]  # ms
            ap_delays = [5, 7, 11, 13]  # ms
            state["combs"] = [
                _CombFilter(int(d * self._sample_rate / 1000.0), block_size=self._block_size)
                for d in comb_delays
            ]
            state["allpasses"] = [
                _AllpassFilter(int(d * self._sample_rate / 1000.0), block_size=self._block_size)
                for d in ap_delays
            ]
            state["predelay"] = _DelayLine(
                int(predelay_ms * self._sample_rate / 1000.0) + 1
            )
            state["predelay"].write_block(np.zeros(self._block_size))
            state["_computed"] = False

        combs: list[_CombFilter] = state["combs"]
        allpasses: list[_AllpassFilter] = state["allpasses"]
        predelay_line: _DelayLine = state["predelay"]

        param_tag = (decay, damping)
        if state.get("_param_tag") != param_tag:
            scaled_fb = 0.3 + decay * 0.6
            scaled_damp = 0.1 + damping * 0.7
            for comb in combs:
                comb.feedback = min(scaled_fb, 0.95)
                comb.damping = min(scaled_damp, 0.85)
            for ap in allpasses:
                ap.gain = 0.3 + damping * 0.3
            state["_param_tag"] = param_tag

        delayed = predelay_line.read_block(
            float(predelay_ms * self._sample_rate / 1000.0), len(buf))
        predelay_line.write_block(buf)

        wet = np.zeros_like(buf, dtype=np.float64)
        for comb in combs:
            wet += comb.process(delayed)
        wet /= len(combs)

        for ap in allpasses:
            wet = ap.process(wet)

        return np.clip(wet, -1.0, 1.0).astype(np.float32)

    # ── Spring reverb ──────────────────────────────────────────────

    def _reverb_spring(self, buf: np.ndarray, params: dict,
                       state: dict) -> np.ndarray:
        """Spring reverb — multi-spring delay-line model.

        Emulates 3-4 parallel springs with resonant bandpass feedback,
        producing the characteristic metallic / boingy spring-tank sound.

        Params:
            decay   — 0.0-1.0, spring tension / sustain length
            damping — 0.0-1.0, high-frequency absorption per spring
            colour  — 0.0-1.0, spring resonant peak emphasis
        """
        decay = params.get("decay", 0.5)
        damping = params.get("damping", 0.3)
        colour = params.get("colour", 0.5)
        predelay_ms = params.get("predelay", 10.0)

        if "spring_lines" not in state:
            # Spring delay times in ms — prime-ish to avoid comb filtering
            spring_delays = [18, 23, 30, 27]  # ms — 4 springs
            spring_q = [4.0, 6.0, 3.0, 5.0]  # resonance Q per spring
            state["spring_lines"] = [
                {
                    "delay": _DelayLine(int(d * self._sample_rate / 1000.0
                                            + self._block_size + 1)),
                    "q": q,
                    "prev": 0.0,
                    "filt_prev": np.zeros(2, dtype=np.float64),
                    "lp_prev": 0.0,
                }
                for d, q in zip(spring_delays, spring_q)
            ]
            state["predelay"] = _DelayLine(
                int(predelay_ms * self._sample_rate / 1000.0) + 1)
            state["predelay"].write_block(np.zeros(self._block_size))

        predelay_line: _DelayLine = state["predelay"]
        delayed = predelay_line.read_block(
            float(predelay_ms * self._sample_rate / 1000.0), len(buf))
        predelay_line.write_block(buf)

        # Feedback / damping scaling
        fb = 0.4 + decay * 0.5          # 0.4-0.9
        damp = 0.05 + damping * 0.6     # 0.05-0.65 (one-pole LP coeff)
        res_gain = 0.3 + colour * 0.6   # spring resonance emphasis 0.3-0.9

        wet = np.zeros(len(buf), dtype=np.float64)
        for spring in state["spring_lines"]:
            dl: _DelayLine = spring["delay"]
            q_val = spring["q"]

            # Read delayed signal
            delay_samps = dl.max_len - self._block_size - 1
            spring_out = dl.read_block(float(delay_samps), len(buf))

            # Bandpass filter per spring — emphasises resonant frequency
            coeff = _compute_bpf_coeffs(
                self._sample_rate / (delay_samps + 1), q_val, self._sample_rate)
            b0, b1, b2, a1, a2 = coeff
            b_arr = np.array([b0, b1, b2], dtype=np.float64)
            a_arr = np.array([1.0, a1, a2], dtype=np.float64)
            f_prev = spring["filt_prev"]
            res, f_zf = lfilter(b_arr, a_arr,
                                spring_out.astype(np.float64), zi=f_prev)
            spring["filt_prev"] = f_zf

            # Apply resonance gain
            res = res * res_gain

            # High-frequency absorption (one-pole LP on feedback path)
            lp_prev = spring["lp_prev"]
            damped = np.zeros_like(res)
            for i in range(len(res)):
                lp_prev = lp_prev * (1.0 - damp) + res[i] * damp
                damped[i] = lp_prev
            spring["lp_prev"] = float(lp_prev)

            # Write back with feedback
            write_sig = delayed.astype(np.float64) + damped * fb
            dl.write_block(np.clip(write_sig, -1.0, 1.0).astype(np.float32))

            wet += damped

        wet /= len(state["spring_lines"])
        return np.clip(wet, -1.0, 1.0).astype(np.float32)

    # ── Plate reverb ───────────────────────────────────────────────

    def _reverb_plate(self, buf: np.ndarray, params: dict,
                      state: dict) -> np.ndarray:
        """Plate reverb — dense comb bank + allpass cascade + modulation.

        Emulates a large metal plate using 12 parallel comb filters with
        modulated delay times for realism, followed by a two-stage allpass
        diffuser for smooth, dense decay.

        Params:
            decay   — 0.0-1.0, plate sustain length
            damping — 0.0-1.0, high-frequency absorption
            density — 0.0-1.0, diffusion density (controls allpass gain)
        """
        decay = params.get("decay", 0.5)
        damping = params.get("damping", 0.4)
        density = params.get("density", 0.6)
        predelay_ms = params.get("predelay", 15.0)

        if "combs" not in state:
            # 12 combs — more = denser plate sound
            comb_delays_ms = [5, 11, 19, 23, 34, 39, 42, 48, 53, 57, 62, 68]
            ap_delays_ms = [2, 5, 9, 13]
            state["combs"] = [
                _CombFilter(int(d * self._sample_rate / 1000.0), block_size=self._block_size)
                for d in comb_delays_ms
            ]
            state["allpasses"] = [
                _AllpassFilter(int(d * self._sample_rate / 1000.0), block_size=self._block_size)
                for d in ap_delays_ms
            ]
            state["predelay"] = _DelayLine(
                int(predelay_ms * self._sample_rate / 1000.0) + 1)
            state["predelay"].write_block(np.zeros(self._block_size))
            # Plate modulation LFO phase
            state["mod_phase"] = 0.0

        combs: list[_CombFilter] = state["combs"]
        allpasses: list[_AllpassFilter] = state["allpasses"]
        predelay_line: _DelayLine = state["predelay"]

        # Update comb parameters
        param_tag = (decay, damping, density)
        if state.get("_param_tag") != param_tag:
            fb = 0.3 + decay * 0.6          # 0.3-0.9
            damp = 0.05 + damping * 0.7     # 0.05-0.75
            ap_gain = 0.3 + density * 0.4   # 0.3-0.7
            for comb in combs:
                comb.feedback = min(fb, 0.94)
                comb.damping = min(damp, 0.85)
            for ap in allpasses:
                ap.gain = ap_gain
            state["_param_tag"] = param_tag

        # Predelay
        delayed = predelay_line.read_block(
            float(predelay_ms * self._sample_rate / 1000.0), len(buf))
        predelay_line.write_block(buf)

        # Plate modulation: very slow LFO (0.15 Hz) for subtle detuning
        mod_phase = state.get("mod_phase", 0.0)
        delta = 0.15 / self._sample_rate
        t = np.arange(len(buf), dtype=np.float64) * delta + mod_phase
        t %= 1.0
        state["mod_phase"] = float((t[-1] + delta) % 1.0)

        # Comb filters (parallel)
        wet = np.zeros(len(buf), dtype=np.float64)
        for comb in combs:
            comb_out = comb.process(delayed)
            wet += comb_out.astype(np.float64, copy=False)
        wet /= len(combs)

        # Allpass diffuser
        for ap in allpasses:
            wet = ap.process(wet)

        # Apply modulation interpolation for subtle pitch wobble
        mod = 0.5 + 0.5 * np.sin(2.0 * np.pi * t)  # 0-1 wobble
        mod_idx = np.arange(len(buf), dtype=np.float64) + (mod - 0.5) * 0.3
        mod_idx = np.clip(mod_idx, 0, len(buf) - 1)
        int_idx = mod_idx.astype(np.int32)
        frac = mod_idx - int_idx
        nxt = np.minimum(int_idx + 1, len(buf) - 1)
        wet = wet[int_idx] * (1.0 - frac) + wet[nxt] * frac

        return np.clip(wet, -1.0, 1.0).astype(np.float32)

    # ── Room reverb ────────────────────────────────────────────────

    def _reverb_room(self, buf: np.ndarray, params: dict,
                     state: dict) -> np.ndarray:
        """Room reverb — early reflection taps + late diffuse tail.

        First 30-80 ms: distinct early reflections (scaled by room size).
        After: a small Schroeder-style late reverb for the diffuse tail.

        Params:
            size    — 0.0-1.0, room dimensions (scales all delays)
            decay   — 0.0-1.0, reverb tail length
            damping — 0.0-1.0, high-frequency absorption
        """
        size = params.get("size", 0.5)
        decay = params.get("decay", 0.4)
        damping = params.get("damping", 0.4)
        predelay_ms = params.get("predelay", 5.0)

        size_factor = 0.3 + size * 1.7  # 0.3-2.0

        if "er_delay" not in state:
            # Early reflection tap times (ms) — room-appropriate spacing
            base_taps_ms = [3, 7, 12, 18, 26, 36, 48, 62]
            # Small late comb/allpass for diffuse tail
            tail_comb_ms = [21, 29, 37, 44]
            tail_ap_ms = [4, 8, 13]
            max_tap = int(max(base_taps_ms) * size_factor
                          * self._sample_rate / 1000.0)
            state["er_delay"] = _DelayLine(max_tap + self._block_size + 1)
            state["er_taps"] = [
                int(t * size_factor * self._sample_rate / 1000.0)
                for t in base_taps_ms
            ]
            state["tail_combs"] = [
                _CombFilter(int(d * size_factor * self._sample_rate / 1000.0 + 1), block_size=self._block_size)
                for d in tail_comb_ms
            ]
            state["tail_allpasses"] = [
                _AllpassFilter(
                    int(d * size_factor * self._sample_rate / 1000.0 + 1),
                    block_size=self._block_size)
                for d in tail_ap_ms
            ]
            state["predelay"] = _DelayLine(
                int(predelay_ms * self._sample_rate / 1000.0) + 1)
            state["predelay"].write_block(np.zeros(self._block_size))

        predelay_line: _DelayLine = state["predelay"]
        delayed = predelay_line.read_block(
            float(predelay_ms * self._sample_rate / 1000.0), len(buf))
        predelay_line.write_block(buf)

        er_delay: _DelayLine = state["er_delay"]
        er_delay.write_block(delayed)

        # ── Early reflections ──
        taps = state["er_taps"]
        num_taps = len(taps)
        er = np.zeros(len(buf), dtype=np.float64)
        amp = 0.5 + decay * 0.4  # 0.5-0.9

        for i, tap in enumerate(taps):
            tap_gain = amp * (1.0 - i * 0.6 / max(num_taps, 1))
            tap_gain = max(tap_gain, 0.0)
            reflected = er_delay.read_block(float(tap), len(buf))
            er += reflected.astype(np.float64) * tap_gain

        er_sum = sum(1.0 - i * 0.6 / num_taps for i in range(num_taps))
        if er_sum > 0:
            er /= er_sum

        # ── Late reverb tail ──
        tail_combs: list[_CombFilter] = state["tail_combs"]
        tail_ap: list[_AllpassFilter] = state["tail_allpasses"]

        param_tag = (decay, damping, size)
        if state.get("_tag_tail") != param_tag:
            fb = 0.2 + decay * 0.6      # 0.2-0.8
            damp = 0.1 + damping * 0.6  # 0.1-0.7
            for c in tail_combs:
                c.feedback = min(fb, 0.92)
                c.damping = min(damp, 0.85)
            for ap in tail_ap:
                ap.gain = 0.3 + damping * 0.2
            state["_tag_tail"] = param_tag

        tail = np.zeros(len(buf), dtype=np.float64)
        for c in tail_combs:
            tail += c.process(delayed)
        tail /= len(tail_combs)
        for ap in tail_ap:
            tail = ap.process(tail)

        # Blend: early reflections dominate early, tail fills in
        fade_len = int(len(buf) * 0.3)
        if fade_len > 0:
            blend = np.ones(len(buf), dtype=np.float64)
            blend[:fade_len] = np.linspace(0.0, 1.0, fade_len)
            wet = er * (1.0 - blend * 0.3) + tail * blend
        else:
            wet = er + tail * 0.5

        return np.clip(wet, -1.0, 1.0).astype(np.float32)

    # ── 12. Volume ──────────────────────────────────────────────────

    def _apply_volume(self, buf: np.ndarray, params: dict,
                      state: dict) -> np.ndarray:
        """Simple volume/level control."""
        level = params.get("level", 1.0)
        return buf * level

    # ── 13. IR Cabinet Simulator ─────────────────────────────────────

    def _apply_ir_cab(self, buf: np.ndarray, params: dict,
                      state: dict) -> np.ndarray:
        """Apply IR convolution via IRLoader.process().

        Delegates to the IRLoader's FFT overlap-add engine.
        Supports wet/dry mix control per-preset.

        Params:
            - ir_file: str (path to .wav IR) — already set via load_ir()
            - enabled: bool
            - wet: float 0.0-1.0
            - dry: float 0.0-1.0
        """
        # Update mix from preset params
        wet = params.get("wet", 1.0)
        dry = params.get("dry", 0.0)
        self.ir.set_mix(wet=wet, dry=dry)
        self.ir.enabled = params.get("enabled", True) and not params.get("bypass", False)

        return self.ir.process(buf)

    # ── 14. Octaver ──────────────────────────────────────────────────

    def _apply_octaver(self, buf: np.ndarray, params: dict,
                       state: dict) -> np.ndarray:
        """Sub-octave via full-wave rectification + divide-by-2.
        Mix control blends between dry and octave-down.
        """
        mix = params.get("mix", 0.5)

        # Full-wave rectification produces even harmonics including sub-octave
        rectified = np.abs(buf)
        # Low-pass to isolate fundamental (below ~300Hz corner)
        # Simple one-pole: alpha ~ 0.95 at 48kHz
        alpha = state.get("lp_alpha", 0.95)
        lp = np.zeros_like(buf)
        prev = state.get("lp_prev", 0.0)
        for i in range(len(buf)):
            prev = alpha * prev + (1.0 - alpha) * rectified[i]
            lp[i] = prev
        state["lp_prev"] = prev

        # Scale to match input level
        sub = lp * 0.5
        out = buf * (1.0 - mix) + sub * mix
        return np.clip(out, -1.0, 1.0)

    # ── 15. Pitch Shifter ────────────────────────────────────────────
    # BETA — test on RPi 4B for xruns

    def _apply_pitch_shifter(self, buf: np.ndarray, params: dict,
                             state: dict) -> np.ndarray:
        """Granular/overlap-add pitch shifter for subtle shifts (±2 semitones).
        CPU-heavy, tagged as beta.
        """
        # BETA — test on RPi 4B for xruns
        shift_semitones = params.get("shift", 0.0)
        mix = params.get("mix", 0.5)

        if shift_semitones == 0.0:
            return buf

        # Pitch factor: 2^(shift/12)
        factor = 2.0 ** (shift_semitones / 12.0)

        # Windowed overlap-add with grain size ~40ms
        grain_size = int(0.040 * self._sample_rate)  # 1920 samples at 48kHz
        hop_in = grain_size
        hop_out = int(grain_size / factor) if factor > 0 else grain_size

        if "ring" not in state:
            state["ring"] = np.zeros(grain_size * 2, dtype=np.float32)
            state["write_pos"] = 0
            state["read_pos"] = 0

        ring: np.ndarray = state["ring"]
        wpos = state["write_pos"]
        rpos = state["read_pos"]

        # Write input into ring buffer
        for i, s in enumerate(buf):
            ring[wpos % len(ring)] = s
            wpos += 1

        # Read with resampling via linear interpolation
        out = np.zeros_like(buf)
        for i in range(len(buf)):
            idx_floor = int(np.floor(rpos))
            frac = rpos - idx_floor
            a = ring[idx_floor % len(ring)]
            b = ring[(idx_floor + 1) % len(ring)]
            out[i] = a + frac * (b - a)
            rpos += factor
            if rpos >= wpos:
                rpos = wpos - 1

        state["write_pos"] = wpos
        state["read_pos"] = rpos

        # Apply raised-cosine window
        window = 0.5 * (1.0 - np.cos(2.0 * np.pi * np.arange(len(buf)) / len(buf)))
        out = out * window

        return buf * (1.0 - mix) + out.astype(np.float32) * mix

    # ── 16. Harmonizer ──────────────────────────────────────────────

    def _apply_harmonizer(self, buf: np.ndarray, params: dict,
                          state: dict) -> np.ndarray:
        """Pitch shift + dry mix, with MIDI-note-aware shift amount.
        shift parameter is in semitones (e.g. 3 = minor third above).
        """
        shift_semitones = params.get("shift", 3.0)
        mix = params.get("mix", 0.5)

        if shift_semitones == 0.0:
            return buf

        factor = 2.0 ** (shift_semitones / 12.0)
        grain_size = int(0.040 * self._sample_rate)
        hop_out = int(grain_size / factor)

        if "ring" not in state:
            state["ring"] = np.zeros(grain_size * 2, dtype=np.float32)
            state["wpos"] = 0
            state["rpos"] = 0

        ring = state["ring"]
        wpos = state["wpos"]
        rpos = state["rpos"]

        for s in buf:
            ring[wpos % len(ring)] = s
            wpos += 1

        out = np.zeros_like(buf)
        for i in range(len(buf)):
            idx = int(np.floor(rpos))
            frac = rpos - idx
            a = ring[idx % len(ring)]
            b = ring[(idx + 1) % len(ring)]
            out[i] = a + frac * (b - a)
            rpos += factor
            if rpos >= wpos:
                rpos = wpos - 1

        state["wpos"] = wpos
        state["rpos"] = rpos

        window = 0.5 * (1.0 - np.cos(2.0 * np.pi * np.arange(len(buf)) / len(buf)))
        out = out * window
        return buf * (1.0 - mix) + out.astype(np.float32) * mix

    # ── 17. Whammy ──────────────────────────────────────────────────

    def _apply_whammy(self, buf: np.ndarray, params: dict,
                      state: dict) -> np.ndarray:
        """Continuous pitch bend via parameter (expression pedal target).
        bend parameter: 0.0 = no shift, 1.0 = +12 semitones up.
        """
        bend = params.get("bend", 0.0)  # 0.0-1.0 maps to 0-12 semitones
        mix = params.get("mix", 0.7)

        shift_semitones = bend * 12.0

        if "ring" not in state:
            grain_size = int(0.040 * self._sample_rate)
            state["ring"] = np.zeros(grain_size * 2, dtype=np.float32)
            state["wpos"] = 0
            state["rpos"] = 0

        factor = 2.0 ** (shift_semitones / 12.0)
        ring = state["ring"]
        wpos = state["wpos"]
        rpos = state["rpos"]

        for s in buf:
            ring[wpos % len(ring)] = s
            wpos += 1

        out = np.zeros_like(buf)
        for i in range(len(buf)):
            idx = int(np.floor(rpos))
            frac = rpos - idx
            a = ring[idx % len(ring)]
            b = ring[(idx + 1) % len(ring)]
            out[i] = a + frac * (b - a)
            rpos += factor
            if rpos >= wpos:
                rpos = wpos - 1

        state["wpos"] = wpos
        state["rpos"] = rpos

        window = 0.5 * (1.0 - np.cos(2.0 * np.pi * np.arange(len(buf)) / len(buf)))
        out = out * window
        return buf * (1.0 - mix) + out.astype(np.float32) * mix

    # ── 18. Detune ──────────────────────────────────────────────────

    def _apply_detune(self, buf: np.ndarray, params: dict,
                      state: dict) -> np.ndarray:
        """Very short modulated delay (0.5-5ms), chorus-like thickening.
        Detunes by shifting pitch via modulated read of a tiny delay line.
        """
        depth = params.get("depth", 0.5)
        mix = params.get("mix", 0.5)

        base_samples = 1.0 * self._sample_rate / 1000.0  # 1ms base
        mod_range = depth * 3.0 * self._sample_rate / 1000.0

        if "delay" not in state:
            max_d = int(base_samples + mod_range + 5.0 * self._sample_rate / 1000.0) + 1
            state["delay"] = _DelayLine(max_d)
            state["delay"].write_block(np.zeros(max_d))

        delay_line: _DelayLine = state["delay"]
        phase = self._lfo_phase(4.0, state, len(buf))  # 4Hz LFO
        lfo = self._lfo_wave(phase, "sine")
        mod_delay = base_samples + lfo * mod_range

        wet = delay_line.read_block_varying(mod_delay)
        delay_line.write_block(buf)
        return np.clip(buf * (1.0 - mix) + wet * mix, -1.0, 1.0)

    # ── 19. Ring Modulator ───────────────────────────────────────────

    def _apply_ring_modulator(self, buf: np.ndarray, params: dict,
                              state: dict) -> np.ndarray:
        """Multiply signal with sine LFO. Rate, Depth, Mix."""
        rate = params.get("rate", 100.0)     # Hz
        depth = params.get("depth", 0.5)
        mix = params.get("mix", 0.5)

        phase = self._lfo_phase(rate, state, len(buf))
        lfo = self._lfo_wave(phase, "sine")
        carrier = lfo * 2.0 - 1.0  # map to [-1, 1]
        ring = buf * carrier
        rm_signal = buf * (1.0 - depth) + ring * depth
        return buf * (1.0 - mix) + rm_signal * mix

    # ── 20. Auto-Wah ────────────────────────────────────────────────

    def _apply_auto_wah(self, buf: np.ndarray, params: dict,
                        state: dict) -> np.ndarray:
        """Envelope follower drives bandpass filter cutoff.
        Sensitivity, Q, Mix.
        """
        sensitivity = params.get("sensitivity", 0.5)
        q = params.get("q", 2.0)
        mix = params.get("mix", 0.5)

        # Envelope follower
        rms = np.sqrt(np.mean(buf ** 2) + _EPS)
        env = state.get("envelope", 0.0)
        env = env * 0.9 + rms * 0.1  # smoothed
        state["envelope"] = env

        # Map envelope to cutoff frequency: 400-4000 Hz
        cutoff = 400.0 + env * sensitivity * 3600.0
        cutoff = np.clip(cutoff, 200.0, 5000.0)

        # BPF biquad
        coeffs = _compute_bpf_coeffs(cutoff, q, self._sample_rate)
        b0, b1, b2, a1, a2 = coeffs
        b = np.array([b0, b1, b2], dtype=np.float64)
        a = np.array([1.0, a1, a2], dtype=np.float64)
        zi = state.get("bpf_zi", np.zeros(2, dtype=np.float64))
        sig, zf = lfilter(b, a, buf.astype(np.float64), zi=zi)
        state["bpf_zi"] = zf

        wet = sig.astype(np.float32)
        return np.clip(buf * (1.0 - mix) + wet * mix, -1.0, 1.0)

    # ── 21. Envelope Filter ──────────────────────────────────────────

    def _apply_envelope_filter(self, buf: np.ndarray, params: dict,
                               state: dict) -> np.ndarray:
        """Filter cutoff follows pick attack envelope.
        Decay, Range, Mix.
        """
        decay = params.get("decay", 0.5)
        freq_range = params.get("range", 0.5)
        mix = params.get("mix", 0.5)

        # Envelope follower with decay
        rms = np.sqrt(np.mean(buf ** 2) + _EPS)
        env = state.get("envelope", 0.0)
        if rms > env:
            env = rms  # instant attack
        else:
            env = env * np.exp(-self._block_size / (decay * 0.1 * self._sample_rate))
        state["envelope"] = env

        cutoff = 200.0 + env * freq_range * 3800.0
        cutoff = np.clip(cutoff, 100.0, 4000.0)

        coeffs = _compute_lpf_coeffs(cutoff, 0.707, self._sample_rate)
        b0, b1, b2, a1, a2 = coeffs
        b = np.array([b0, b1, b2], dtype=np.float64)
        a = np.array([1.0, a1, a2], dtype=np.float64)
        zi = state.get("lpf_zi", np.zeros(2, dtype=np.float64))
        sig, zf = lfilter(b, a, buf.astype(np.float64), zi=zi)
        state["lpf_zi"] = zf

        wet = sig.astype(np.float32)
        return np.clip(buf * (1.0 - mix) + wet * mix, -1.0, 1.0)

    # ── 22. Rotary Speaker ──────────────────────────────────────────

    def _apply_rotary_speaker(self, buf: np.ndarray, params: dict,
                              state: dict) -> np.ndarray:
        """Dual modulated allpass filters (Leslie sim).
        Speed, Drive, Mix.
        """
        speed = params.get("speed", 0.5)   # 0.0-1.0, slow-fast
        drive = params.get("drive", 0.3)
        mix = params.get("mix", 0.5)

        # Map speed: slow rotor ~0.5Hz, fast ~6.5Hz
        rotor_rate = 0.5 + speed * 6.0
        # Horn: ~2x rotor speed
        horn_rate = rotor_rate * 2.0

        # Dual modulated delays (rotor + horn)
        if "rotor_delay" not in state:
            state["rotor_delay"] = _DelayLine(int(0.020 * self._sample_rate + 1))
            state["horn_delay"] = _DelayLine(int(0.015 * self._sample_rate + 1))
            state["rotor_delay"].write_block(np.zeros(int(0.020 * self._sample_rate)))
            state["horn_delay"].write_block(np.zeros(int(0.015 * self._sample_rate)))

        rotor_phase = self._lfo_phase(rotor_rate, state, len(buf))
        horn_phase = self._lfo_phase(horn_rate, state, len(buf))

        # Rotor: deeper modulation, slower
        rotor_lfo = self._lfo_wave(rotor_phase, "sine")
        rotor_delay_s = 0.003 + rotor_lfo * 0.005  # 3-8ms

        # Horn: shallower, faster
        horn_lfo = self._lfo_wave(horn_phase, "sine")
        horn_delay_s = 0.001 + horn_lfo * 0.002  # 1-3ms

        rotor_wet = state["rotor_delay"].read_block_varying(
            rotor_delay_s * self._sample_rate)
        state["rotor_delay"].write_block(buf)

        horn_wet = state["horn_delay"].read_block_varying(
            horn_delay_s * self._sample_rate)
        state["horn_delay"].write_block(buf * 0.7)

        # Mix with subtle drive (soft clip)
        combined = rotor_wet + horn_wet
        if drive > 0:
            combined = np.tanh(combined * (1.0 + drive * 2.0))

        return buf * (1.0 - mix) + combined.astype(np.float32) * mix

    # ── 23. Uni-Vibe ────────────────────────────────────────────────

    def _apply_uni_vibe(self, buf: np.ndarray, params: dict,
                        state: dict) -> np.ndarray:
        """Photoresistor-style phaser.
        Rate, Depth, Feedback, Mix.
        """
        rate = params.get("rate", 0.8)
        depth = params.get("depth", 0.5)
        feedback = params.get("feedback", 0.3)
        mix = params.get("mix", 0.5)

        phase = self._lfo_phase(rate, state, len(buf))
        lfo = self._lfo_wave(phase, "sine")

        # Uni-Vibe uses 4 allpass stages with LFO-driven modulation
        # Centre frequency sweeps between 200-2000 Hz
        if "fb_buf" not in state:
            state["fb_buf"] = np.zeros(len(buf), dtype=np.float64)

        fb_buf = state["fb_buf"]
        fb_input = buf.astype(np.float64, copy=False) + fb_buf * feedback

        # Frequency per sample varies with LFO
        sig = fb_input.copy()
        for stage in range(4):
            freq = 200.0 + lfo * depth * 1800.0
            freq_mean = float(np.mean(freq))
            w = 2.0 * np.pi * freq_mean / self._sample_rate
            tan_half_w = np.tan(w / 2.0)
            coeff = (1.0 - tan_half_w) / (1.0 + tan_half_w)

            b = np.array([coeff, 1.0], dtype=np.float64)
            a = np.array([1.0, coeff], dtype=np.float64)
            zi = state.get(f"uv_zi_{stage}", np.zeros(1, dtype=np.float64))
            sig, zf = lfilter(b, a, sig, zi=zi)
            state[f"uv_zi_{stage}"] = zf

        state["fb_buf"] = sig * 0.3
        sig = np.clip(sig, -1.0, 1.0)
        return (buf * (1.0 - mix) + sig.astype(np.float32) * mix).astype(np.float32)

    # ── 24. Auto-Pan ────────────────────────────────────────────────

    def _apply_auto_pan(self, buf: np.ndarray, params: dict,
                        state: dict) -> np.ndarray:
        """LFO-driven stereo pan.
        Rate, Depth, Waveform.
        Currently returns mono with pan applied; full stereo needs
        2-channel pipeline. Simulated here as amplitude modulation.
        """
        rate = params.get("rate", 0.3)
        depth = params.get("depth", 0.7)
        waveform = params.get("waveform", "sine")

        phase = self._lfo_phase(rate, state, len(buf))
        lfo = self._lfo_wave(phase, waveform)
        # Map LFO (0-1) to pan [-1, 1]
        pan = (lfo * 2.0 - 1.0) * depth
        # Apply pan as amplitude modulation
        left_gain = np.clip(1.0 - pan, 0.0, 1.0)
        right_gain = np.clip(1.0 + pan, 0.0, 1.0)
        # Mono output: average both channels
        mod = (left_gain + right_gain) * 0.5
        return (buf * mod).astype(np.float32)

    # ── 25. Stereo Widener ──────────────────────────────────────────

    def _apply_stereo_widener(self, buf: np.ndarray, params: dict,
                              state: dict) -> np.ndarray:
        """Mid/side processing, widen by boosting side channel.
        Width, Mix.
        Mono input: creates synthetic side from filtered difference.
        """
        width = params.get("width", 0.5)
        mix = params.get("mix", 0.5)

        # From mono: create synthetic stereo by delaying a copy
        if "delay" not in state:
            state["delay"] = _DelayLine(int(0.030 * self._sample_rate + 1))
            state["delay"].write_block(np.zeros(int(0.030 * self._sample_rate)))

        delay_line: _DelayLine = state["delay"]
        # Read delayed signal for "side" channel
        side = delay_line.read_block(0.010 * self._sample_rate, len(buf))
        delay_line.write_block(buf)

        # Mid = original, Side = delayed difference
        mid = buf * 0.5 + side * 0.5
        side_boosted = (buf - side) * (0.5 + width * 0.5)

        # Mix: blend widened signal with original
        widened = mid + side_boosted
        return (buf * (1.0 - mix) + widened * mix).astype(np.float32)

    # ── 26. Bitcrusher ──────────────────────────────────────────────

    def _apply_bitcrusher(self, buf: np.ndarray, params: dict,
                          state: dict) -> np.ndarray:
        """Reduce bit depth + sample rate. Bits, Rate, Mix."""
        bits = params.get("bits", 8)       # 1-16
        rate_reduction = params.get("rate", 1.0)  # 1-10, higher = more crush
        mix = params.get("mix", 0.5)

        # Bit depth reduction
        n_levels = max(2, 2 ** min(bits, 16))
        crushed = np.round(buf * n_levels) / n_levels

        # Sample rate reduction (hold each sample for N samples)
        hold = max(1, int(rate_reduction))
        if hold > 1:
            out = np.zeros_like(crushed)
            held_val = state.get("held_val", 0.0)
            hold_count = state.get("hold_count", 0)
            for i in range(len(crushed)):
                if hold_count >= hold:
                    held_val = crushed[i]
                    hold_count = 0
                out[i] = held_val
                hold_count += 1
            state["held_val"] = float(held_val)
            state["hold_count"] = hold_count
            crushed = out

        return buf * (1.0 - mix) + crushed * mix

    # ── 27. Wavefolder ──────────────────────────────────────────────

    def _apply_wavefolder(self, buf: np.ndarray, params: dict,
                          state: dict) -> np.ndarray:
        """Fold signal back at thresholds (Klanggebiet-style).
        Gain, Fold, Mix.
        """
        gain = params.get("gain", 2.0)
        fold = params.get("fold", 3.0)
        mix = params.get("mix", 0.5)

        shaped = buf * gain

        # Multi-stage wavefolding
        for _ in range(int(fold)):
            # Fold: if abs > 1, fold back
            above = np.abs(shaped) > 1.0
            shaped = np.where(above, np.sign(shaped) * (2.0 - np.abs(shaped)), shaped)

        return (buf * (1.0 - mix) + shaped * mix).astype(np.float32)

    # ── 28. Rectifier ───────────────────────────────────────────────

    def _apply_rectifier(self, buf: np.ndarray, params: dict,
                         state: dict) -> np.ndarray:
        """Full/half-wave rectification for octave-up fuzz.
        Mode (full/half), Mix.
        """
        mode = params.get("mode", "full")   # "full" or "half"
        mix = params.get("mix", 0.5)

        if mode == "full":
            rectified = np.abs(buf)  # full-wave: all positive, octave up
        else:
            rectified = np.where(buf > 0, buf, 0.0)  # half-wave

        return buf * (1.0 - mix) + rectified * mix

    # ── 29. Expander ────────────────────────────────────────────────

    def _apply_expander(self, buf: np.ndarray, params: dict,
                        state: dict) -> np.ndarray:
        """Inverse compressor, widens dynamic range.
        Threshold, Ratio, Attack, Release.
        """
        threshold_db = params.get("threshold", -30.0)
        ratio = params.get("ratio", 3.0)
        attack_ms = params.get("attack", 5.0)
        release_ms = params.get("release", 100.0)

        rms = np.sqrt(np.mean(buf ** 2) + _EPS)
        envelope = state.get("envelope", 0.0)

        if rms > envelope:
            alpha = np.exp(-self._block_size / (attack_ms * self._sample_rate / 1000.0))
        else:
            alpha = np.exp(-self._block_size / (release_ms * self._sample_rate / 1000.0))

        envelope = envelope * alpha + rms * (1.0 - alpha)
        state["envelope"] = envelope

        if envelope > 1e-10:
            env_db = 20.0 * np.log10(envelope)
        else:
            env_db = -120.0

        if env_db < threshold_db:
            # Expand: gain_db = (env - threshold) * (ratio - 1)
            gain_db = (env_db - threshold_db) * (ratio - 1.0)
        else:
            gain_db = 0.0

        gain_lin = 10 ** (gain_db / 20.0)
        return np.clip(buf * gain_lin, -1.0, 1.0)

    # ── 30. De-esser ────────────────────────────────────────────────

    def _apply_de_esser(self, buf: np.ndarray, params: dict,
                        state: dict) -> np.ndarray:
        """Compress sibilance band (EQ split + compressor).
        Frequency, Threshold, Mix.
        """
        freq = params.get("frequency", 6000.0)   # sibilance centre
        threshold = params.get("threshold", -30.0)
        mix = params.get("mix", 0.5)

        # Bandpass the sibilance band
        q = 3.0  # narrow Q for sibilance band
        coeffs = _compute_bpf_coeffs(freq, q, self._sample_rate)
        b0, b1, b2, a1, a2 = coeffs
        b = np.array([b0, b1, b2], dtype=np.float64)
        a = np.array([1.0, a1, a2], dtype=np.float64)
        zi = state.get("bpf_zi", np.zeros(2, dtype=np.float64))
        sibilance, zf = lfilter(b, a, buf.astype(np.float64), zi=zi)
        state["bpf_zi"] = zf

        # RMS of sibilance band
        sib_rms = np.sqrt(np.mean(sibilance ** 2) + _EPS)

        # Compress sibilance
        if sib_rms > (10.0 ** (threshold / 20.0)):
            gain = 0.3  # heavy reduction
        else:
            gain = 1.0

        # Split: clean signal minus reduced sibilance
        clean = buf.astype(np.float64) - sibilance * (1.0 - gain)
        return (buf * (1.0 - mix) + clean.astype(np.float32) * mix).astype(np.float32)

    # ── 31. Transient Shaper ────────────────────────────────────────

    def _apply_transient_shaper(self, buf: np.ndarray, params: dict,
                                state: dict) -> np.ndarray:
        """Attack boost/sustain per note. Attack, Sustain, Mix."""
        attack = params.get("attack", 0.5)     # boost amount
        sustain = params.get("sustain", 0.5)   # sustain amount
        mix = params.get("mix", 0.5)

        # Envelope follower with fast attack/slow release
        rms = np.sqrt(np.mean(buf ** 2) + _EPS)
        env = state.get("envelope", 0.0)
        if rms > env:
            env = env * 0.5 + rms * 0.5  # fast attack
        else:
            env = env * 0.995 + rms * 0.005  # slow release
        state["envelope"] = env

        # Differentiate to find transients
        if "prev" not in state:
            state["prev"] = np.float32(0.0)
        diff = np.abs(buf.astype(np.float64) - state["prev"])
        state["prev"] = np.float32(buf[-1])

        # Transient detection: diff > envelope * threshold
        trans_thresh = max(env * 0.1, 0.001)
        is_transient = diff > trans_thresh

        # Attack boost on transients
        boost = np.ones_like(buf)
        boost[is_transient] = 1.0 + attack * 2.0

        # Sustain: amplify quiet tail
        tail = env > 0.0
        sustain_gain = 1.0 + (sustain - 0.5) * 2.0

        shaped = buf * boost
        shaped[~is_transient] = buf[~is_transient] * sustain_gain

        return (buf * (1.0 - mix) + shaped.astype(np.float32) * mix).astype(np.float32)

    # ── 32. Sidechain Compressor ────────────────────────────────────

    def _apply_sidechain_compressor(self, buf: np.ndarray, params: dict,
                                    state: dict) -> np.ndarray:
        """Compression triggered by external input (4CM return channel).
        Threshold, Ratio, Attack, Release, Mix.
        NOTE: In mono mode, simulates sidechain from the input signal itself.
        """
        threshold_db = params.get("threshold", -20.0)
        ratio = params.get("ratio", 4.0)
        attack_ms = params.get("attack", 2.0)
        release_ms = params.get("release", 50.0)
        mix = params.get("mix", 0.5)

        # Sidechain signal = the input itself (in mono mode)
        # In 4CM mode, the return channel would be the sidechain
        sidechain = buf

        rms = np.sqrt(np.mean(sidechain ** 2) + _EPS)
        envelope = state.get("envelope", 0.0)

        if rms > envelope:
            alpha = np.exp(-self._block_size / (attack_ms * self._sample_rate / 1000.0))
        else:
            alpha = np.exp(-self._block_size / (release_ms * self._sample_rate / 1000.0))

        envelope = envelope * alpha + rms * (1.0 - alpha)
        state["envelope"] = envelope

        if envelope > 1e-10:
            env_db = 20.0 * np.log10(envelope)
        else:
            env_db = -120.0

        if env_db > threshold_db:
            gain_db = threshold_db + (env_db - threshold_db) / ratio - env_db
        else:
            gain_db = 0.0

        gain_lin = 10 ** (gain_db / 20.0)
        compressed = np.clip(buf * gain_lin, -1.0, 1.0)
        return buf * (1.0 - mix) + compressed * mix

    # ── 33. Parametric EQ ──────────────────────────────────────────

    def _apply_parametric_eq(self, buf: np.ndarray, params: dict,
                             state: dict) -> np.ndarray:
        """N-band biquad peaking filters.
        Takes array-style params: freq_0, gain_0, q_0, freq_1, ...
        Up to 4 bands.
        """
        # Parse bands from flattened params
        sig = buf.astype(np.float64, copy=False)
        for band in range(4):
            freq = params.get(f"freq_{band}", 0.0)
            gain_db = params.get(f"gain_{band}", 0.0)
            q = params.get(f"q_{band}", 0.707)
            if freq == 0.0 or gain_db == 0.0:
                continue

            coeffs = _compute_peaking_coeffs(freq, gain_db, q, self._sample_rate)
            b0, b1, b2, a1, a2 = coeffs
            b = np.array([b0, b1, b2], dtype=np.float64)
            a = np.array([1.0, a1, a2], dtype=np.float64)
            zi = state.get(f"peq_zi_{band}", np.zeros(2, dtype=np.float64))
            sig, zf = lfilter(b, a, sig, zi=zi)
            state[f"peq_zi_{band}"] = zf

        return np.clip(sig, -1.0, 1.0).astype(np.float32)

    # ── 34-37. HPF / LPF / BPF / Notch ──────────────────────────────

    def _apply_hpf(self, buf: np.ndarray, params: dict,
                   state: dict) -> np.ndarray:
        """High-pass filter. Frequency, Slope (6/12dB oct)."""
        freq = params.get("frequency", 200.0)
        slope = params.get("slope", 12.0)

        q = 0.707 if slope >= 12 else 0.5  # 12dB = Q=0.707 (Butter), 6dB = Q=0.5
        coeffs = _compute_hpf_coeffs(freq, q, self._sample_rate)
        b0, b1, b2, a1, a2 = coeffs
        b = np.array([b0, b1, b2], dtype=np.float64)
        a = np.array([1.0, a1, a2], dtype=np.float64)
        zi = state.get("hpf_zi", np.zeros(2, dtype=np.float64))
        sig, zf = lfilter(b, a, buf.astype(np.float64), zi=zi)
        state["hpf_zi"] = zf

        # If 6dB slope, mix in a first-order version (less aggressive)
        if slope < 12:
            # First-order HPF: simpler - just one pole
            alpha = 1.0 / (1.0 + self._sample_rate / (2.0 * np.pi * freq))
            zi1 = state.get("hpf_zi_1", np.float64(0.0))
            out = np.zeros(len(buf), dtype=np.float64)
            for i in range(len(buf)):
                zi1 = alpha * (zi1 + buf[i] - state.get("hpf_last", np.float64(0.0)))
                out[i] = zi1
                state["hpf_last"] = np.float64(buf[i])
            state["hpf_zi_1"] = zi1
            return np.clip(out, -1.0, 1.0).astype(np.float32)

        return np.clip(sig, -1.0, 1.0).astype(np.float32)

    def _apply_lpf(self, buf: np.ndarray, params: dict,
                   state: dict) -> np.ndarray:
        """Low-pass filter. Frequency, Slope (6/12dB)."""
        freq = params.get("frequency", 5000.0)
        slope = params.get("slope", 12.0)

        q = 0.707 if slope >= 12 else 0.5
        coeffs = _compute_lpf_coeffs(freq, q, self._sample_rate)
        b0, b1, b2, a1, a2 = coeffs
        b = np.array([b0, b1, b2], dtype=np.float64)
        a = np.array([1.0, a1, a2], dtype=np.float64)
        zi = state.get("lpf_zi", np.zeros(2, dtype=np.float64))
        sig, zf = lfilter(b, a, buf.astype(np.float64), zi=zi)
        state["lpf_zi"] = zf

        if slope < 12:
            alpha = 1.0 / (1.0 + self._sample_rate / (2.0 * np.pi * freq))
            zi1 = state.get("lpf_zi_1", np.float64(0.0))
            out = np.zeros(len(buf), dtype=np.float64)
            for i in range(len(buf)):
                zi1 = zi1 + alpha * (buf[i] - zi1)
                out[i] = zi1
            state["lpf_zi_1"] = zi1
            return np.clip(out, -1.0, 1.0).astype(np.float32)

        return np.clip(sig, -1.0, 1.0).astype(np.float32)

    def _apply_bpf(self, buf: np.ndarray, params: dict,
                   state: dict) -> np.ndarray:
        """Band-pass filter. Frequency, Q."""
        freq = params.get("frequency", 1000.0)
        q = params.get("q", 0.707)
        coeffs = _compute_bpf_coeffs(freq, q, self._sample_rate)
        b0, b1, b2, a1, a2 = coeffs
        b = np.array([b0, b1, b2], dtype=np.float64)
        a = np.array([1.0, a1, a2], dtype=np.float64)
        zi = state.get("bpf_zi", np.zeros(2, dtype=np.float64))
        sig, zf = lfilter(b, a, buf.astype(np.float64), zi=zi)
        state["bpf_zi"] = zf
        return np.clip(sig, -1.0, 1.0).astype(np.float32)

    def _apply_notch(self, buf: np.ndarray, params: dict,
                     state: dict) -> np.ndarray:
        """Notch filter. Frequency, Q."""
        freq = params.get("frequency", 60.0)
        q = params.get("q", 10.0)  # High Q for narrow notch
        coeffs = _compute_notch_coeffs(freq, q, self._sample_rate)
        b0, b1, b2, a1, a2 = coeffs
        b = np.array([b0, b1, b2], dtype=np.float64)
        a = np.array([1.0, a1, a2], dtype=np.float64)
        zi = state.get("notch_zi", np.zeros(2, dtype=np.float64))
        sig, zf = lfilter(b, a, buf.astype(np.float64), zi=zi)
        state["notch_zi"] = zf
        return np.clip(sig, -1.0, 1.0).astype(np.float32)

    # ── 38. Formant Filter ──────────────────────────────────────────

    def _apply_formant_filter(self, buf: np.ndarray, params: dict,
                              state: dict) -> np.ndarray:
        """Fixed resonant peaks (vowel shapes). Vowel (a/e/i/o/u), Mix."""
        vowel = params.get("vowel", "a")
        mix = params.get("mix", 0.5)

        # Formant frequencies and bandwidths for vowels (F1, F2, F3)
        formants = {
            "a": [(700, 80), (1100, 90), (2500, 120)],
            "e": [(400, 60), (1800, 80), (2600, 100)],
            "i": [(300, 50), (2300, 80), (2900, 100)],
            "o": [(450, 70), (800, 80), (2700, 110)],
            "u": [(300, 50), (700, 70), (2700, 110)],
        }
        bands = formants.get(vowel, formants["a"])

        sig = buf.astype(np.float64, copy=False)
        for stage, (freq, bw) in enumerate(bands):
            q_val = freq / bw  # Q from bandwidth
            coeffs = _compute_peaking_coeffs(freq, 12.0, q_val, self._sample_rate)
            b0, b1, b2, a1, a2 = coeffs
            b = np.array([b0, b1, b2], dtype=np.float64)
            a = np.array([1.0, a1, a2], dtype=np.float64)
            zi = state.get(f"form_zi_{stage}", np.zeros(2, dtype=np.float64))
            sig, zf = lfilter(b, a, sig, zi=zi)
            state[f"form_zi_{stage}"] = zf

        wet = np.clip(sig, -1.0, 1.0).astype(np.float32)
        return np.clip(buf * (1.0 - mix) + wet * mix, -1.0, 1.0)

    # ── 39. Ping-pong Delay ─────────────────────────────────────────

    def _apply_ping_pong_delay(self, buf: np.ndarray, params: dict,
                               state: dict) -> np.ndarray:
        """Alternate left/right taps. Time, Feedback, Mix.
        Mono input: creates stereo effect by alternating.
        """
        time_ms = params.get("time", 400.0)
        feedback = params.get("feedback", 0.3)
        mix = params.get("mix", 0.4)

        delay_samples = int(time_ms * self._sample_rate / 1000.0)
        if "delay" not in state:
            max_d = max(delay_samples * 2, self._sample_rate)
            state["delay"] = _DelayLine(max_d + 1)
            state["delay"].write_block(np.zeros(max_d // 2))
            state["ping"] = 1  # 1 = left, -1 = right

        delay_line: _DelayLine = state["delay"]
        wet = delay_line.read_block(float(delay_samples), len(buf))
        fb_gain = min(feedback, 0.98)

        # Alternate pan per block
        pan = state.get("ping", 1)
        panned_wet = wet * (0.5 + pan * 0.5)

        write_sig = buf + panned_wet * fb_gain
        delay_line.write_block(write_sig)
        state["ping"] = -pan

        return buf * (1.0 - mix) + panned_wet * mix

    # ── 40. Multi-tap Delay ─────────────────────────────────────────

    def _apply_multi_tap_delay(self, buf: np.ndarray, params: dict,
                               state: dict) -> np.ndarray:
        """Rhythmic repeats (dotted eighth, triplet). Pattern, Feedback, Mix."""
        pattern = params.get("pattern", "quarter")  # quarter/dotted/triplet
        feedback = params.get("feedback", 0.3)
        mix = params.get("mix", 0.4)

        # Tap times in ms, relative to quarter note = 500ms (120 BPM default)
        quarter = 500.0
        tap_patterns = {
            "quarter": [quarter],
            "dotted": [quarter * 1.5],
            "triplet": [quarter / 3.0, quarter * 2.0 / 3.0, quarter],
        }
        taps = tap_patterns.get(pattern, tap_patterns["quarter"])

        max_delay = int(max(taps) * self._sample_rate / 1000.0)
        if "delay" not in state:
            max_d = max(max_delay * 2, self._sample_rate)
            state["delay"] = _DelayLine(max_d + 1)
            state["delay"].write_block(np.zeros(max_d // 2))

        delay_line: _DelayLine = state["delay"]

        # Sum all tap reads
        wet = np.zeros(len(buf), dtype=np.float32)
        for tap_ms in taps:
            tap_samples = int(tap_ms * self._sample_rate / 1000.0)
            tap_sig = delay_line.read_block(float(tap_samples), len(buf))
            wet += tap_sig * (1.0 / len(taps))  # Normalise

        fb_gain = min(feedback, 0.98)
        delay_line.write_block(np.clip(buf + wet * fb_gain, -1.0, 1.0))
        return np.clip(buf * (1.0 - mix) + wet * mix, -1.0, 1.0)

    # ── 41. Reverse Delay ───────────────────────────────────────────

    def _apply_reverse_delay(self, buf: np.ndarray, params: dict,
                             state: dict) -> np.ndarray:
        """Buffer + reverse before playback. Time, Feedback, Mix."""
        time_ms = params.get("time", 400.0)
        feedback = params.get("feedback", 0.3)
        mix = params.get("mix", 0.4)

        delay_samples = int(time_ms * self._sample_rate / 1000.0)
        if "delay" not in state:
            max_d = max(delay_samples * 2, self._sample_rate)
            state["delay"] = _DelayLine(max_d + 1)
            state["delay"].write_block(np.zeros(max_d // 2))
            state["read_pos"] = 0

        delay_line: _DelayLine = state["delay"]
        rpos = state.get("read_pos", 0)

        # Read reversed: read from buffer start, write at current pos
        # Reverse: read from (write_pos - read_offset) going backwards
        wpos = delay_line.write_idx
        buf_len = len(buf)

        out = np.zeros(buf_len, dtype=np.float32)
        for i in range(buf_len):
            # Read sample at (wpos - rpos) mod max_len
            rev_idx = (wpos - rpos) % delay_line.max_len
            out[i] = delay_line.buf[rev_idx]
            rpos = (rpos + 1) % delay_samples

        state["read_pos"] = rpos % delay_samples

        fb_gain = min(feedback, 0.98)
        delay_line.write_block(np.clip(buf + out * fb_gain, -1.0, 1.0))
        return np.clip(buf * (1.0 - mix) + out * mix, -1.0, 1.0)

    # ── 42. Tape Echo ───────────────────────────────────────────────

    def _apply_tape_echo(self, buf: np.ndarray, params: dict,
                         state: dict) -> np.ndarray:
        """Delay with wow/flutter + high-cut feedback. Time, Wow, High-Cut, Mix."""
        time_ms = params.get("time", 300.0)
        wow = params.get("wow", 0.3)        # wow/flutter intensity
        high_cut = params.get("high_cut", 0.5)  # 0.0-1.0 high cut
        feedback = params.get("feedback", 0.3)
        mix = params.get("mix", 0.4)

        delay_samples = int(time_ms * self._sample_rate / 1000.0)
        if "delay" not in state:
            max_d = max(delay_samples * 2, self._sample_rate)
            state["delay"] = _DelayLine(max_d + 1)
            state["delay"].write_block(np.zeros(max_d // 2))

        delay_line: _DelayLine = state["delay"]

        # Wow/flutter: LFO on read position (wobble)
        phase = self._lfo_phase(4.0, state, len(buf))  # 4Hz wow
        lfo = self._lfo_wave(phase, "sine")
        wow_offset = (lfo * 2.0 - 1.0) * wow * 3.0  # up to ±3 samples wobble
        read_delay = float(delay_samples) + wow_offset

        wet = delay_line.read_block_varying(read_delay)

        # High-cut: one-pole low-pass on feedback
        if high_cut > 0:
            cutoff_factor = 1.0 - high_cut * 0.9  # 0.1-1.0
            lp_prev = state.get("lp_prev", 0.0)
            lp_out = np.zeros_like(wet)
            for i in range(len(wet)):
                lp_prev = lp_prev * cutoff_factor + wet[i] * (1.0 - cutoff_factor)
                lp_out[i] = lp_prev
            state["lp_prev"] = float(lp_prev)
            wet = lp_out.astype(np.float32)

        fb_gain = min(feedback, 0.98)
        delay_line.write_block(np.clip(buf + wet * fb_gain, -1.0, 1.0))
        return np.clip(buf * (1.0 - mix) + wet * mix, -1.0, 1.0)

    # ── 43. Shimmer Reverb ──────────────────────────────────────────
    # BETA — test on RPi 4B for xruns

    def _apply_shimmer_reverb(self, buf: np.ndarray, params: dict,
                              state: dict) -> np.ndarray:
        """Reverb + pitch shift + re-verb. Shift, Decay, Mix.
        CPU-heavy, tagged as beta.
        """
        # BETA — test on RPi 4B for xruns
        shift = params.get("shift", 0.0)   # semitones
        decay = params.get("decay", 0.5)
        mix = params.get("mix", 0.4)

        # Use regular reverb as base
        if "combs" not in state:
            comb_delays = [29, 37, 44, 50, 31, 39, 47, 53]
            ap_delays = [5, 7, 11, 13]
            state["combs"] = [_CombFilter(int(d * self._sample_rate / 1000.0), block_size=self._block_size) for d in comb_delays]
            state["allpasses"] = [_AllpassFilter(int(d * self._sample_rate / 1000.0), block_size=self._block_size) for d in ap_delays]
            state["predelay"] = _DelayLine(int(30.0 * self._sample_rate / 1000.0 + 1))
            state["predelay"].write_block(np.zeros(int(30.0 * self._sample_rate / 1000.0)))

        combs: list[_CombFilter] = state["combs"]
        allpasses: list[_AllpassFilter] = state["allpasses"]
        predelay_line: _DelayLine = state["predelay"]

        scaled_fb = 0.3 + decay * 0.6
        scaled_damp = 0.1 + decay * 0.5
        for comb in combs:
            comb.feedback = min(scaled_fb, 0.95)
            comb.damping = min(scaled_damp, 0.85)
        for ap in allpasses:
            ap.gain = 0.3 + decay * 0.3

        # Predelay
        delayed = predelay_line.read_block(30.0 * self._sample_rate / 1000.0, len(buf))
        predelay_line.write_block(buf)

        # Comb + allpass reverb
        wet = np.zeros_like(buf, dtype=np.float64)
        for comb in combs:
            wet += comb.process(delayed)
        wet /= len(combs)
        for ap in allpasses:
            wet = ap.process(wet)

        # Pitch shift the reverb tail (shimmer!)
        if shift != 0.0:
            factor = 2.0 ** (shift / 12.0)
            if "shift_ring" not in state:
                grain_size = int(0.040 * self._sample_rate)
                state["shift_ring"] = np.zeros(grain_size * 2, dtype=np.float32)
                state["swpos"] = 0
                state["srpos"] = 0
            ring = state["shift_ring"]
            wpos = state["swpos"]
            rpos = state["srpos"]
            wet_float = wet.astype(np.float32)
            for i, s in enumerate(wet_float):
                ring[wpos % len(ring)] = s
                wpos += 1
            shifted_out = np.zeros_like(buf, dtype=np.float32)
            for i in range(len(buf)):
                idx = int(np.floor(rpos))
                frac = rpos - idx
                a = ring[idx % len(ring)]
                b = ring[(idx + 1) % len(ring)]
                shifted_out[i] = a + frac * (b - a)
                rpos += factor
                if rpos >= wpos:
                    rpos = wpos - 1
            state["swpos"] = wpos
            state["srpos"] = rpos
            window = 0.5 * (1.0 - np.cos(2.0 * np.pi * np.arange(len(buf)) / len(buf)))
            shifted_out = shifted_out * window
            wet = wet + shifted_out.astype(np.float64) * 0.5  # feedback shimmer

        wet = np.clip(wet, -1.0, 1.0).astype(np.float32)
        return np.clip(buf * (1.0 - mix) + wet * mix, -1.0, 1.0)

    # ── 44. Looper ─────────────────────────────────────────────────

    def _apply_looper(self, buf: np.ndarray, params: dict,
                      state: dict) -> np.ndarray:
        """Record/overdub/playback (~30s buffer = 6MB).
        Controls: record, overdub, play, stop.
        State machine: idle -> recording -> playing -> overdub -> idle
        """
        max_buffer = int(30.0 * self._sample_rate)  # 30 seconds

        if "looper_buf" not in state:
            state["looper_buf"] = np.zeros(max_buffer, dtype=np.float32)
            state["looper_pos"] = 0
            state["looper_len"] = 0
            state["looper_mode"] = "idle"

        looper_buf: np.ndarray = state["looper_buf"]
        pos = state.get("looper_pos", 0)
        buf_len = state.get("looper_len", 0)
        mode = state.get("looper_mode", "idle")

        # Control actions
        if params.get("record", 0):
            mode = "recording"
            state["looper_len"] = 0
            state["looper_pos"] = 0
            looper_buf[:] = 0.0
        elif params.get("overdub", 0):
            mode = "overdub"
            state["looper_pos"] = 0
        elif params.get("play", 0):
            mode = "playing"
            state["looper_pos"] = 0
        elif params.get("stop", 0):
            mode = "idle"

        state["looper_mode"] = mode

        if mode == "idle":
            return buf  # passthrough

        out = np.zeros_like(buf)
        n = len(buf)

        if mode == "recording":
            # Write input to buffer
            end = min(pos + n, max_buffer)
            looper_buf[pos:end] = buf[:end - pos]
            state["looper_pos"] = pos + n
            state["looper_len"] = max(state["looper_len"], pos + n)
            out = buf * 0.5  # Monitor at half volume during recording

        elif mode == "overdub":
            # Play existing + record new
            for i in range(n):
                looper_buf[pos % max_buffer] = (
                    looper_buf[pos % max_buffer] + buf[i]
                ) * 0.7  # Mix down to avoid clipping
                out[i] = looper_buf[pos % max_buffer]
                pos += 1
            state["looper_pos"] = pos
            state["looper_len"] = max(state["looper_len"], pos)

        elif mode == "playing":
            # Play from buffer
            for i in range(n):
                out[i] = looper_buf[pos % buf_len] if buf_len > 0 else 0.0
                pos += 1
            state["looper_pos"] = pos

        return out.astype(np.float32)

    # ── 45. Early Reflections ───────────────────────────────────────

    def _apply_early_reflections(self, buf: np.ndarray, params: dict,
                                 state: dict) -> np.ndarray:
        """Sparse delay taps + allpass for room simulation.
        Size, Decay, Mix.
        """
        room_size = params.get("size", 0.5)
        decay = params.get("decay", 0.4)
        mix = params.get("mix", 0.4)

        # Early reflection tap times (ms) scaled by room size
        base_taps = [5, 12, 20, 30, 45, 65]
        size_factor = 0.5 + room_size * 2.0  # 0.5-2.5x
        taps = [int(t * size_factor * self._sample_rate / 1000.0) for t in base_taps]

        if "delay" not in state:
            max_d = max(taps) + self._block_size + 1
            state["delay"] = _DelayLine(max_d)

        delay_line: _DelayLine = state["delay"]
        delay_line.write_block(buf)

        # Read reflections with decreasing amplitude
        wet = np.zeros(len(buf), dtype=np.float32)
        amplitude = 0.6 + decay * 0.5  # 0.6-1.1
        for i, tap in enumerate(taps):
            gain = amplitude * (1.0 - i / len(taps))
            reflected = delay_line.read_block(float(tap), len(buf))
            wet += reflected * gain

        # Normalize
        wet = wet / max(len(taps), 1) * 0.5

        return (buf * (1.0 - mix) + wet * mix).astype(np.float32)

    def set_audio_profile(self, block_size: int, sample_rate: int) -> None:
        """Update block size and sample rate at runtime.

        Recomputes profile-dependent constants and clears DSP state so
        effects reinitialise with the new buffer sizes on the next
        process() call.

        Called by the web server when the JACK latency profile changes
        (``POST /api/audio/profile``). Follows the same pattern as
        :meth:`NAMEngineRouter.set_block_size`.

        Args:
            block_size: JACK period size in frames (e.g. 64, 128, 256, 512).
            sample_rate: Sample rate in Hz (e.g. 44100, 48000, 96000).
        """
        changed = (self._block_size != block_size
                   or self._sample_rate != sample_rate)
        if not changed:
            return
        self._block_size = block_size
        self._sample_rate = sample_rate
        self._vu_alpha = np.exp(-block_size / (0.05 * sample_rate))
        # Clear DSP state — effects will reinit with new block/sample rate
        self._state.clear()
        self._coeffs.clear()
        logger.info("Audio profile updated: block=%d, sr=%d",
                     block_size, sample_rate)

    @property
    def block_size(self) -> int:
        """Current JACK period / block size in frames."""
        return self._block_size

    @property
    def sample_rate(self) -> int:
        """Current audio sample rate in Hz."""
        return self._sample_rate

    # ── Properties ─────────────────────────────────────────────────

    @property
    def master_volume(self) -> float:
        return self._master_volume

    @master_volume.setter
    def master_volume(self, value: float) -> None:
        self._master_volume = max(0.0, min(1.0, value))

    @property
    def bypassed(self) -> bool:
        return self._bypassed

    @bypassed.setter
    def bypassed(self, value: bool) -> None:
        self._bypassed = value
        logger.info("Global bypass: %s", "ON" if value else "OFF")

    @property
    def tuner_enabled(self) -> bool:
        return self._tuner_enabled

    @tuner_enabled.setter
    def tuner_enabled(self, value: bool) -> None:
        self._tuner_enabled = value

    # ── 4CM routing properties ────────────────────────────────────

    @property
    def routing_mode(self) -> str:
        return self._routing_mode

    @routing_mode.setter
    def routing_mode(self, value: str) -> None:
        if value not in ("mono", "4cm"):
            raise ValueError(f"routing_mode must be 'mono' or '4cm', got {value!r}")
        self._routing_mode = value
        logger.info("Routing mode: %s", value)

    @property
    def routing_breakpoint(self) -> int:
        return self._routing_breakpoint

    @routing_breakpoint.setter
    def routing_breakpoint(self, value: int) -> None:
        self._routing_breakpoint = max(0, value)
        logger.info("Routing breakpoint: %d", self._routing_breakpoint)

    # ── Chain access (read-only for external inspection) ────────────

    @property
    def chain(self) -> list[dict]:
        """Read-only access to the DSP chain."""
        return list(self._chain)

    def set_routing(self, mode: str, breakpoint: int = 7) -> None:
        """Set 4CM routing configuration at runtime.

        Args:
            mode: ``\"mono\"`` or ``\"4cm\"``.
            breakpoint: Chain index where pre/post split occurs.
        """
        self.routing_mode = mode
        self.routing_breakpoint = breakpoint