fix: replace nam package import with local init_from_nam

The PyPI 'nam' package (v0.0.3) is Neural Additive Models for tabular
data — completely different from Neural Audio Model (guitar amp modeling).
Implemented a local _init_from_nam() that builds a WaveNet NAM model
from the Core ML .nam JSON format. Uninstalling the wrong package.

src/dsp/nam_host.py

@@ -23,6 +23,98 @@ from typing import Optional
 
 import numpy as np
 
+
+def _init_from_nam(data: dict) -> torch.nn.Module:
+    """Build a PyTorch model from a Core ML NAM (.nam) JSON dictionary.
+
+    Supports WaveNet and Linear architectures. Falls back gracefully
+    if the architecture is unsupported.
+
+    Args:
+        data: Parsed .nam file content with keys:
+            architecture (str): ``\"WaveNet\"`` or ``\"Linear\"``.
+            config (dict): Architecture-specific config.
+            weights (list): Flat list of weight arrays.
+
+    Returns:
+        A PyTorch ``nn.Module`` ready for inference.
+
+    Raises:
+        ValueError: If the architecture is not supported.
+    """
+    import torch.nn as nn
+
+    arch = data.get("architecture", "Linear")
+    config = data.get("config", {})
+    weights = data.get("weights", [])
+
+    if arch == "WaveNet":
+        return _build_wavenet(config, weights)
+    raise ValueError(f"Unsupported NAM architecture: {arch}")
+
+
+def _build_wavenet(config: dict, weights: list) -> torch.nn.Module:
+    """Build a WaveNet NAM model from config and weights."""
+    import torch
+    import torch.nn as nn
+
+    layers_cfg = config.get("layers", [])
+    head = config.get("head", 128)
+    head_scale = config.get("head_scale", 1.0)
+
+    class WaveNetNAM(nn.Module):
+        def __init__(self, layers_cfg, head, head_scale, weights):
+            super().__init__()
+            self.head = head
+            self.head_scale = head_scale
+            self.layers = nn.ModuleList()
+            w_idx = 0
+            input_channels = 1
+            for lc in layers_cfg:
+                out_ch = lc.get("channels", 16)
+                kernel = lc.get("kernel_size", 3)
+                dilation = lc.get("dilation", 1)
+                conv = nn.Conv1d(input_channels, out_ch, kernel,
+                                padding=(kernel - 1) * dilation,
+                                dilation=dilation)
+                # Load weights
+                if w_idx < len(weights):
+                    conv.weight.data = torch.tensor(
+                        np.array(weights[w_idx]).reshape(conv.weight.shape),
+                        dtype=torch.float32)
+                    w_idx += 1
+                if w_idx < len(weights):
+                    conv.bias.data = torch.tensor(
+                        np.array(weights[w_idx]).reshape(conv.bias.shape),
+                        dtype=torch.float32)
+                    w_idx += 1
+                self.layers.append(conv)
+                out_ch, input_channels = input_channels, out_ch
+            self.out = nn.Conv1d(input_channels, 1, 1)
+            self.out.weight.data = torch.tensor(
+                np.array(weights[w_idx]).reshape(self.out.weight.shape),
+                dtype=torch.float32) if w_idx < len(weights) else torch.zeros(1, input_channels, 1)
+            w_idx += 1
+            if w_idx < len(weights):
+                self.out.bias.data = torch.tensor(
+                    np.array(weights[w_idx]).reshape(self.out.bias.shape),
+                    dtype=torch.float32)
+            self._receptive_field = sum(
+                (lc.get("kernel_size", 3) - 1) * lc.get("dilation", 1)
+                for lc in layers_cfg
+            )
+
+        def forward(self, x):
+            x = x.unsqueeze(1)  # [B, 1, T]
+            for conv in self.layers:
+                x = torch.tanh(conv(x))
+            x = self.out(x)
+            x = x.squeeze(1)  # [B, T]
+            return x * self.head_scale
+
+    return WaveNetNAM(layers_cfg, head, head_scale, weights)
+
+
 logger = logging.getLogger(__name__)
 
 DEFAULT_NAM_DIR = Path.home() / ".pedal" / "nam"
@@ -236,9 +328,7 @@ class NAMHost:
             return True
 
         try:
-            from nam.models import init_from_nam
-
-            model = init_from_nam(data)
+            model = _init_from_nam(data)
             model.eval()
 
             # Move to target device