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/synthetic_spectra/generator.py

@@ -49,14 +49,14 @@ class IsotopeSource:
 @dataclass
 class SpectrumConfig:
     """Configuration for a single spectrum generation."""
-    
+
     # Time parameters
     duration_seconds: float = 60.0
     time_interval_seconds: float = 1.0  # Each row in the spectrogram
-    
+
     # Sources to include
     sources: List[IsotopeSource] = field(default_factory=list)
-    
+
     # Background options
     include_background: bool = True
     background_cps: float = 5.0
@@ -64,18 +64,25 @@ class SpectrumConfig:
     include_radon: bool = True
     include_thorium: bool = True
     measured_background_path: Optional[str] = None
-    
+
+    # Background subtraction simulation
+    # When True, generates a second independent background realization
+    # and subtracts it from the spectrum, then clips negatives to 0.
+    # This simulates what happens at inference time (measured bg subtraction).
+    subtract_background: bool = False
+
     # Detector configuration
     detector_name: str = "radiacode_103"
-    
+
     # Noise options
     apply_poisson: bool = True
     apply_electronic: bool = False
     electronic_noise_sigma: float = 0.5
-    
-    # Normalization
+
+    # Normalization — "log1p" preserves relative signal levels,
+    # works well after background subtraction where many channels are ~0.
     normalize: bool = True
-    normalization_method: str = "max"  # max, sum, log, sqrt
+    normalization_method: str = "log1p"  # max, sum, log, sqrt, log1p
 
 
 @dataclass
@@ -272,7 +279,7 @@ class SpectrumGenerator:
         all_source_isotopes.extend(src_iso)
         all_background_isotopes.extend(bg_iso)
 
-        # Apply noise
+        # Apply noise before any subtraction (Poisson noise on raw counts)
         if config.apply_poisson:
             spectrum = apply_poisson_noise(spectrum)
 
@@ -282,6 +289,24 @@ class SpectrumGenerator:
                 config.electronic_noise_sigma
             )
 
+        # Simulate background subtraction (matches inference pipeline)
+        if config.subtract_background and config.include_background:
+            # Generate an independent background realization
+            bg_spectrum2, _ = generate_environmental_background(
+                self.energy_bins,
+                config.duration_seconds,
+                background_cps=config.background_cps,
+                include_k40=config.include_k40,
+                include_radon=config.include_radon,
+                include_thorium=config.include_thorium,
+                detector_config=self.detector_config,
+                measured_background_path=config.measured_background_path,
+            )
+            if config.apply_poisson:
+                bg_spectrum2 = apply_poisson_noise(bg_spectrum2)
+            # Subtract and clip — same as inference: net = clip(rate - bg_rate, 0, inf)
+            spectrum = np.maximum(spectrum - bg_spectrum2, 0)
+
         # Normalize if requested
         if config.normalize:
             spectrum = normalize_spectrum(spectrum, config.normalization_method)