feat: overhaul screenshot detection + suppression + explanation text (v7)

Detection improvements:
- Standard screen resolution matching (19 common resolutions including Retina)
- Low unique color count (<32 per channel — real UI renders use exact colors)
- H/V edge dominance with minimum edge strength gate
- Hard Bayer CFA exclusion (cameras can't be screenshots)
- 2-trait gate requiring at least 1 content-based trait (prevents false triggers)
- Safety guard exception: screenshots allowed through low-detail guard at score>=0.6

Suppression expansion (15 → 37 tests):
- All optical lens tests (no lens on screenshots)
- All sensor tests (no camera sensor)
- Diffusion Notches ×0.15 (LCD pixel grid creates false spectral harmonics)
- Benford's Law ×0.2, Color Histogram ×0.2 (UI color distributions violate natural priors)
- DCT/Wavelet Kurtosis ×0.3, Gradient Sparsity ×0.3
- EXIF/Maker Note/Thumbnail/GPS ×0.1 (screenshots have no camera metadata)

Explanation text:
- Screenshot-specific disclosure in forensic report explaining methodology limitations
- Screenshot-aware conclusion/recommendation in court brief
- Explicit note that photographic forensic tests don't apply to UI screenshots

agents/modality_detector.py

@@ -1,15 +1,18 @@
 """
-FORENSIQ — Capture Modality Detector v6
+FORENSIQ — Capture Modality Detector v7
+
+v7: Overhauled SCREENSHOT detection — now detects OS screenshots (pixel-perfect,
+standard resolution, few unique colors, no Bayer CFA, uniform sharpness) in
+addition to photographed screens (H/V edge dominance). Expanded SCREENSHOT 
+suppression from 15 to 30+ tests covering lens physics, sensor characteristics,
+LCD pixel grid frequency artifacts (Diffusion Notches), and UI color distribution
+violations (Benford's Law, Color Histogram).
 
 v6: Added Vignetting cos⁴θ, Bokeh Shape, and Fixed Pattern Noise ×0.3 suppression
-for MACRO_DSLR modality. Macro lenses have non-standard field illumination (not cos⁴θ),
-extremely smooth circular bokeh from long focal length, and low row/column variance 
-in the smooth background region.
+for MACRO_DSLR modality.
 
 v5: Added minimum bayer_margin threshold (0.005) to prevent JPEG re-encoding
-from creating false Bayer CFA patterns on AI images. Real cameras have 
-bayer_margin > 0.01; AI images through JPEG have margin < 0.001.
-Also: sharpness_ratio > 3.0 hard gate for macro detection.
+from creating false Bayer CFA patterns on AI images.
 """
 
 import sys
@@ -44,8 +47,6 @@ def detect_modality(img: Image.Image) -> ModalityResult:
     rgb = np.array(img.convert("RGB")).astype(np.float64)
     
     # ═══ CRITICAL PRE-CHECK: Bayer CFA pattern ═══════════════════════
-    # Real camera: σ_green < σ_red ≈ σ_blue (green has 2x sampling density)
-    # AI image through JPEG: all channels have ~same noise (margin ≈ 0)
     noise_std = {}
     for c, nm in enumerate(["red", "green", "blue"]):
         ch = rgb[:, :, c]
@@ -53,7 +54,6 @@ def detect_modality(img: Image.Image) -> ModalityResult:
         noise_std[nm] = float(np.std(ch - dn))
     
     bayer_margin = min(noise_std["red"], noise_std["blue"]) - noise_std["green"]
-    # Require BOTH: green < min(red, blue) AND margin above noise floor
     has_bayer = (noise_std["green"] < min(noise_std["red"], noise_std["blue"]) 
                  and bayer_margin > MIN_BAYER_MARGIN)
     
@@ -83,7 +83,6 @@ def detect_modality(img: Image.Image) -> ModalityResult:
     sharp_region = sharpness > sharp_thresh
     sharp_frac = float(np.mean(sharp_region))
     
-    # Blur region: content-based threshold (20% of p95)
     blur_content_thresh = p95 * 0.20
     blur_region = sharpness < blur_content_thresh
     blur_frac = float(np.mean(blur_region))
@@ -177,6 +176,11 @@ def detect_modality(img: Image.Image) -> ModalityResult:
         indicators["macro_detected"] = True
     
     # ── Screenshot detection ──────────────────────────────────────────
+    # Two types: (1) OS screenshot (PrintScreen/cmd+Shift) — pixel-perfect,
+    #   no lens physics, exact resolution match, few unique colors
+    # (2) Photographed screen — has lens physics but also display grid artifacts
+    
+    # H/V edge dominance (UI elements are rectilinear)
     edge_mag = np.hypot(sobel(gray, 0), sobel(gray, 1))
     strong = edge_mag > np.percentile(edge_mag, 95)
     gx = sobel(gray, axis=1); gy = sobel(gray, axis=0)
@@ -184,11 +188,70 @@ def detect_modality(img: Image.Image) -> ModalityResult:
     v_edges = np.abs(gy) > np.abs(gx) * 3
     hv_ratio = float(np.sum(h_edges | v_edges)) / (float(np.sum(strong)) + 1e-9)
     
-    ratio = max(w, h) / (min(w, h) + 1e-9)
-    if hv_ratio > 0.6 and ratio > 1.8:
-        scores["SCREENSHOT"] = 0.6
+    # H/V ratio is only meaningful if edges are genuinely strong (not noise)
+    median_edge = float(np.median(edge_mag))
+    strong_edges_present = float(np.percentile(edge_mag, 95)) > max(5.0, median_edge * 3)
+    
+    aspect = max(w, h) / (min(w, h) + 1e-9)
+    
+    # Standard screen resolutions (and 2× Retina variants)
+    standard_res = {
+        (1920, 1080), (2560, 1440), (3840, 2160), (1366, 768), (1280, 720),
+        (1440, 900), (1680, 1050), (2560, 1600), (2880, 1800), (3024, 1964),
+        (1536, 864), (1600, 900), (3456, 2234), (2560, 1080), (3440, 1440),
+        (1280, 800), (1024, 768), (2048, 1536), (2732, 2048),
+    }
+    is_standard_res = (w, h) in standard_res or (h, w) in standard_res
+    
+    # Count unique colors per channel — UI screenshots use few exact colors
+    unique_colors = [len(np.unique(rgb[:, :, c].astype(np.uint8))) for c in range(3)]
+    max_unique = max(unique_colors)
+    low_color_count = max_unique < 32  # Pure UI screenshots use <30 distinct values per channel
+    
+    # Uniform sharpness — no blur gradient (unlike camera photos)
+    sharpness_cv = float(np.std(sharpness)) / (float(np.mean(sharpness)) + 1e-9)
+    uniform_sharpness = sharpness_cv < 0.5
+    
+    # No Bayer CFA = not a camera sensor
+    no_bayer = not has_bayer
+    
+    screenshot_score = 0.0
+    screenshot_traits = 0  # Count distinct UI traits for gating
+    
+    # Hard gate: images with real Bayer CFA pattern are from cameras, not screenshots
+    if not has_bayer:
+        # OS screenshot path — pixel-perfect, no lens physics
+        if is_standard_res:
+            screenshot_score += 0.2
+            screenshot_traits += 1
+            indicators["standard_resolution"] = True
+        if low_color_count:
+            screenshot_score += 0.25
+            screenshot_traits += 1
+            indicators["low_color_count"] = max_unique
+        if hv_ratio > 0.6 and strong_edges_present:
+            screenshot_score += 0.25
+            screenshot_traits += 1
+        if uniform_sharpness:
+            screenshot_score += 0.15
+        if low_color_count:
+            screenshot_score += 0.1  # No sensor + few colors = UI render
+        
+        # Gate: require at least 2 distinct UI traits AND at least one must be
+        # content-based (colors or edges), not just resolution match.
+        # This prevents smooth AI images at 1920×1080 from false-triggering.
+        has_content_trait = low_color_count or (hv_ratio > 0.6 and strong_edges_present)
+        if screenshot_traits < 2 or not has_content_trait:
+            screenshot_score = 0.0
+    
+    if screenshot_score >= 0.4:
+        scores["SCREENSHOT"] = min(1.0, screenshot_score)
         indicators["screenshot_detected"] = True
     
+    indicators["hv_ratio"] = round(hv_ratio, 3)
+    indicators["max_unique_colors"] = max_unique
+    indicators["sharpness_cv"] = round(sharpness_cv, 3)
+    
     # ── Double JPEG detection (messaging) ─────────────────────────────
     hc, wc = (gray.shape[0] // 8) * 8, (gray.shape[1] // 8) * 8
     blockiness = 1.0
@@ -260,10 +323,17 @@ def detect_modality(img: Image.Image) -> ModalityResult:
         conf = min(1.0, scores["PORTRAIT_MODE"])
     
     # SAFETY GUARD 1: No detail = possible AI
-    if not has_detail:
+    # Exception: SCREENSHOT modality is legitimately low-detail (UI renders exact colors,
+    # not photographic texture). If screenshot signals are strong, let it through.
+    if not has_detail and modality != "SCREENSHOT":
         modality = "UNKNOWN"
         conf = 0.2
         indicators["safety_override"] = "Low-detail image — suppression disabled"
+    elif not has_detail and modality == "SCREENSHOT" and scores.get("SCREENSHOT", 0) < 0.6:
+        # Weak screenshot signal + no detail → still override
+        modality = "UNKNOWN"
+        conf = 0.2
+        indicators["safety_override"] = "Low-detail, weak screenshot signal — suppression disabled"
     
     # SAFETY GUARD 2: No real Bayer = not a real camera
     if not has_bayer and modality in ("PORTRAIT_MODE", "SMARTPHONE", "MESSAGING"):
@@ -338,12 +408,27 @@ def _get_modality_adjustments(modality: str) -> dict:
         }
     elif modality == "SCREENSHOT":
         return {
+            # Optical physics — screenshots have no lens, suppress all lens tests
             "Vignetting cos⁴θ": 0.1, "Vignetting Symmetry": 0.1,
             "Lens Distortion": 0.1, "Field Curvature": 0.1,
             "CA Magnitude": 0.1, "CA Radial Gradient": 0.1, "Lateral CA": 0.1,
             "Purple Fringing": 0.1, "Bokeh Shape": 0.1,
+            "Sharpness Falloff": 0.1, "DoF Consistency": 0.1, "DoF Gradient Direction": 0.1,
+            "Diffraction Limit": 0.1, "Optical Center": 0.1,
+            # Sensor — screenshots have no camera sensor
             "PRNU Uniformity": 0.1, "Bayer CFA Pattern": 0.1, "CFA Nyquist": 0.1,
             "Hot/Dead Pixels": 0.1, "Noise Autocorrelation": 0.1, "Demosaic Interpolation": 0.1,
+            "PRNU Cross-Channel": 0.1, "Poisson-Gaussian Model": 0.1,
+            "Fixed Pattern Noise": 0.1, "Green Pixel Imbalance": 0.1,
+            # Generative model — LCD pixel grid creates false frequency artifacts
+            "Diffusion Notches": 0.15, "FFT Grid 8×8": 0.3, "FFT Grid 16×16": 0.3,
+            # Statistical — UI color distributions violate natural-image priors
+            "Benford's Law": 0.2, "Color Histogram": 0.2,
+            "DCT Kurtosis": 0.3, "Wavelet Kurtosis": 0.3,
+            "Gradient Sparsity": 0.3,
+            # Metadata — screenshots lack camera EXIF by definition
+            "EXIF Completeness": 0.1, "Maker Note": 0.1, "Thumbnail Check": 0.1,
+            "GPS Plausibility": 0.1, "ICC Color Profile": 0.2,
         }
     elif modality == "SMARTPHONE":
         return {