Fix: CsI(Tl) non-linear response correction + detector calibration overhaul

Root cause of Am-241 misidentification: the Radiacode 103's CsI(Tl) crystal
shifts low-energy peaks upward (59.5 keV → 71.6 keV for Am-241) due to
non-proportional scintillation response. The model was trained on theoretical
peak positions and couldn't match the shifted real peaks.

Changes:
- Add inverse CsI(Tl) non-linear correction to inference pipeline
  (radiacode_monitor.py, web/config.py, test_detection.py)
  E_apparent = E_true * (1 + 0.37 * exp(-E_true/100))
  Corrects channel mapping so peaks appear at theoretical energies
- Fix energy calibration: DetectorConfig now uses E = 0.33 + 2.97*ch
  with 1023 channels, matching the real detector (was energy_min=20,
  skip_first_channel=True, different channel width)
- Add K-escape peaks for CsI(Tl) iodine X-ray escape (E - 28.5 keV)
- Add asymmetric peak shapes for low-energy tails (< 200 keV)
- Add log1p normalization in dataset and inference (replaces max-norm)
- Add background-subtracted training mode (subtract_background flag)
- Add low-signal augmentation (0.01-5 Bq activities, 30-300s durations)
- Update docker-compose.yml: batch_size=32, duration=30-300s,
  CSI_NONLINEAR_ALPHA/BETA env vars for detect and web
- Web dashboard: apply CsI correction to displayed spectra
- Various UI fixes (Chart.js width, zoom/pan, isotope lines)

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

train/vega_ml/training/vega/__init__.py

@@ -14,7 +14,12 @@ Features:
 
 from .model import VegaModel, VegaConfig
 from .dataset import SpectrumDataset, create_data_loaders
-from .train import train_vega, VegaTrainer
+
+def __getattr__(name):
+    if name in ('train_vega', 'VegaTrainer'):
+        from .train import train_vega, VegaTrainer
+        return locals()[name]
+    raise AttributeError(f"module {__name__!r} has no attribute {name!r}")
 
 __all__ = [
     'VegaModel',