fix: replace Laplacian bokeh check with local-variance method in autocorrelation override

The old Laplacian-based bokeh detection in m05_autocorrelation() failed for macro
DSLR images because the 10%-of-peak-sharpness threshold didn't reliably capture
the bokeh region when the sharp-to-smooth transition was gradual.

New approach uses uniform_filter to compute local variance (E[X²]-E[X]²) over
32×32 windows, then thresholds at 5% of the P95 variance. This works because:
- Real bokeh: P95 is high (sharp subject), most pixels are far below → detected
- AI smooth: P95 is low (<50), guard clause blocks bokeh excuse → override fires
- Normal photo: variance spread evenly, few pixels below 5% of P95 → no false trigger

For the Unsplash macro DSLR test image, this should:
- Drop Autocorrelation Peak from +0.80 to +0.15 (bokeh explained)
- Remove override_suppression flag → modality ×0.1 can further reduce to +0.015
- Move Generative Model Agent from ~+0.054 to ~-0.02 (neutral/authentic)
- Push P(Fake) from ~46.9% to ~44-45%

agents/model_agent.py

@@ -78,37 +78,53 @@ def m05_autocorrelation(img):
     re=max(h,w)//20; Y,X=np.mgrid[0:h,0:w]; cm=((X-cx)**2+(Y-cy)**2)<re**2; acm[cm]=0
     ms=float(np.max(acm))
     
-    # Before firing hard override, check if high autocorrelation is explained by bokeh
-    # Macro/portrait photos have large uniform bokeh regions that create high autocorrelation
-    # from optical physics, not AI generation.
+    # Before firing hard override, check if high autocorrelation is explained by bokeh.
+    # Macro/portrait/shallow-DoF photos have large uniform bokeh regions that create
+    # high autocorrelation from optical physics, not AI generation.
+    #
+    # Method: Local variance via uniform_filter. Bokeh regions have very low local 
+    # variance (uniform pixel intensity). This is more robust than the Laplacian 
+    # approach because it directly measures what bokeh is — spatial uniformity —
+    # rather than inferring it from edge magnitude.
     bokeh_explained = False
+    bokeh_fraction = 0.0
     if ms > 0.95:
-        from scipy.signal import convolve2d
-        lap = np.array([[0,1,0],[1,-4,1],[0,1,0]], dtype=np.float64)
-        sharpness = gaussian_filter(np.abs(convolve2d(gray, lap, mode="same", boundary="symm")), sigma=10)
+        from scipy.ndimage import uniform_filter as uf
         
-        # Estimate bokeh region: pixels with sharpness below 10% of peak
-        peak_sharp = float(np.percentile(sharpness, 95))
-        if peak_sharp > 1:
-            bokeh_thresh = peak_sharp * 0.1  # Below 10% of peak = blurred
-            bokeh_region = sharpness < bokeh_thresh
-            bokeh_fraction = float(np.mean(bokeh_region))
-            
-            bokeh_vals = sharpness[bokeh_region] if np.sum(bokeh_region) > 100 else np.array([1])
-            bokeh_uniformity = 1.0 - min(float(np.std(bokeh_vals)) / (float(np.mean(bokeh_vals)) + 1e-9), 1.0)
-            
-            # The non-bokeh region must have genuine sharp detail (not smooth gradient)
-            has_genuine_sharp = peak_sharp > 10.0
-            
-            # Bokeh explains autocorrelation if:
-            # 1. Large blur region (>25% of image below 10% of peak sharpness)
-            # 2. AND genuine sharp foreground detail exists
-            if bokeh_fraction > 0.25 and has_genuine_sharp:
-                bokeh_explained = True
+        # Compute local variance: E[X²] - E[X]² over 32×32 windows
+        local_mean = uf(gray, size=32)
+        local_sq_mean = uf(gray**2, size=32)
+        local_var = np.clip(local_sq_mean - local_mean**2, 0, None)
+        
+        # Bimodal variance test: in a real bokeh/macro/shallow-DoF image, the 
+        # variance distribution is strongly bimodal — sharp foreground has high 
+        # local variance while the defocused background has near-zero variance.
+        #
+        # Strategy: Use the P95 of local variance (captures the sharp subject's
+        # typical variance) and threshold at 5% of that value. Pixels below this
+        # threshold are effectively uniform/bokeh. This works because:
+        # - Real bokeh: P95 is high (sharp subject), most pixels are far below it
+        # - AI smooth: P95 is low, so threshold is tiny, nothing qualifies as 
+        #   "bokeh relative to sharp detail" because there IS no sharp detail
+        # - Normal photo: variance is spread out, few pixels are <5% of P95
+        p95_var = float(np.percentile(local_var, 95))
+        
+        if p95_var > 50.0:  # Guard: sharp region must have real detail
+            var_thresh = p95_var * 0.05  # 5% of peak variance
+            low_var_mask = local_var < var_thresh
+            bokeh_fraction = float(np.mean(low_var_mask))
+            has_genuine_detail = True  # Already guaranteed by p95 > 50 check
         else:
-            # Entire image is smooth — no bokeh, just low detail
+            # No region with genuine high-variance detail — not a bokeh image
             bokeh_fraction = 0.0
-            bokeh_explained = False  # Smooth everywhere ≠ bokeh
+            low_var_mask = np.zeros_like(local_var, dtype=bool)
+            has_genuine_detail = False
+        
+        # Bokeh explains autocorrelation if:
+        # 1. Large uniform region (>35% of image has variance < 5% of P95)
+        # 2. AND genuine sharp detail exists (P95 variance > 50)
+        if bokeh_fraction > 0.35 and has_genuine_detail:
+            bokeh_explained = True
     
     if ms > 0.95 and not bokeh_explained:
         s, n = 0.8, f"CRITICAL: Autocorrelation {ms:.3f} exceeds physical camera limit — AI generated"
@@ -116,7 +132,7 @@ def m05_autocorrelation(img):
         result["override_suppression"] = True
         return result
     elif ms > 0.95 and bokeh_explained:
-        s, n = 0.15, f"High autocorrelation ({ms:.3f}) but large bokeh region ({bokeh_fraction:.0%}) — likely optical, not AI"
+        s, n = 0.15, f"High autocorrelation ({ms:.3f}) but large bokeh region ({bokeh_fraction:.0%} uniform background) — likely optical, not AI"
         return {"test":"Autocorrelation Peak","max_secondary":round(ms,4),"score":s,"note":n,"bokeh_explained":True}
     elif ms>0.3: s,n=0.6,f"Strong secondary peak ({ms:.3f}) — GAN checkerboard"
     elif ms>0.15: s,n=0.3,f"Moderate peak ({ms:.3f})"