research(phase3b): benchmark 1D CNN on RAW pulses — data-starved, benched

Trains a PyTorch 1D CNN on encoded RAW pulse arrays as the Phase 3B leg
of the designed ensemble, scored on the SAME group-disjoint held-out test
set as the heuristic and statistical models (GroupShuffleSplit by device
sub-folder, no near-duplicate leakage).

RESULT — CNN is data-starved and loses decisively:
  CNN        balanced 0.338, top-1 0.795 (garage-inflated)
  statistical balanced 0.625, top-1 0.868
  heuristic  top-1 0.300
Only 489 train samples with severe class imbalance (Garage 565, Security 2,
Weather 12). A blend sweep confirmed every non-zero CNN weight degrades the
statistical model (0.625 -> 0.613 at 15% CNN, worse beyond), so the CNN is
NOT wired into decode(). Production ensemble stays heuristic + statistical,
both already live.

Committing the trainer + shared pulse_encoder (trainer/inference parity) +
metrics.json to document the reproducible negative result. The benched
.pt/.onnx binaries are intentionally NOT committed.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

models/category_cnn_metrics.json

@@ -0,0 +1,36 @@
+{
+  "n_samples": 803,
+  "n_train": 489,
+  "n_val": 94,
+  "n_test": 220,
+  "classes": [
+    "Doorbell",
+    "Fan Controller",
+    "Garage Door Opener",
+    "Remote Control",
+    "Security Sensor",
+    "Weather Sensor"
+  ],
+  "pulse_len": 512,
+  "epochs": 60,
+  "class_balance": {
+    "Doorbell": 39,
+    "Garage Door Opener": 565,
+    "Weather Sensor": 12,
+    "Fan Controller": 91,
+    "Security Sensor": 2,
+    "Remote Control": 94
+  },
+  "cnn": {
+    "accuracy": 0.7954545454545454,
+    "balanced_accuracy": 0.33796618290289177,
+    "macro_f1": 0.3552945963506287
+  },
+  "heuristic_same_test": {
+    "top1": 0.3
+  },
+  "statistical_same_test": {
+    "top1": 0.8681818181818182,
+    "balanced_accuracy": 0.6253728690437551
+  }
+}