feat(fingerprint): improve CLIP attribution accuracy and add 4th detector

- Rewrite GENERATOR_PROMPTS to describe visual forensic artifacts instead of
content style — CLIP picks up texture/frequency cues better than generator names
- Add umm-maybe/AI-image-detector (EfficientNet fine-tuned on AI vs real) as 4th
ensemble detector with weight 0.1; rebalance existing weights to [0.4,0.3,0.2,0.1]
- Raise CLIP confidence threshold 0.25 → 0.32 (2.9× above chance for 9 classes)
to reduce incorrect attributions when CLIP is uncertain

src/engines/fingerprint/engine.py

@@ -26,18 +26,19 @@ DETECTOR_CANDIDATES = [
     "Organika/sdxl-detector",
     "haywoodsloan/ai-image-detector-deploy",
     "dima806/deepfake_vs_real_image_detection",
+    "umm-maybe/AI-image-detector",
 ]
 
 GENERATOR_PROMPTS: dict[str, str] = {
-    "real":               "a real photograph taken by a camera with natural lighting and film grain",
-    "sora":               "a Sora text-to-video frame with temporal coherence and photorealistic lighting",
-    "runway":             "a Runway Gen-2 frame with painterly dreamlike motion blur and color grading",
-    "wav2lip":            "a Wav2Lip face-swap with sharp lip boundary artifacts and texture inconsistency at mouth edges",
-    "stable_diffusion":   "an image generated by Stable Diffusion with painterly soft textures and dreamlike quality",
-    "sdxl":               "an image generated by SDXL with high resolution detail, sharp edges and crisp textures",
-    "midjourney":         "an image generated by Midjourney with cinematic dramatic lighting and extreme hyperdetail",
-    "dall_e":             "an image generated by DALL-E with clean flat illustration style and smooth gradients",
-    "unknown_generative": "an AI-generated image with unidentifiable generator-specific artifacts and synthetic patterns",
+    "real":               "photograph with natural film grain, uneven organic noise, authentic lens distortion, and real-world lighting imperfections",
+    "sora":               "AI video frame with unnaturally smooth temporal transitions, photorealistic but physically implausible motion, and over-consistent lighting",
+    "runway":             "AI video frame with painterly color grading artifacts, dreamlike motion blur inconsistencies, and synthetic depth-of-field",
+    "wav2lip":            "face with sharp unnatural lip boundary artifacts, texture discontinuity around the mouth region, and mismatched skin tone at lip edges",
+    "stable_diffusion":   "image with soft overly-smooth skin, color bleeding at object edges, dreamlike over-saturation, and repeating background texture patterns",
+    "sdxl":               "image with hyper-sharp commercial detail, perfect noise-free skin, unnaturally crisp edges, and over-rendered textures lacking real-world imperfection",
+    "midjourney":         "image with dramatic cinematic vignette, fantasy color palette, exaggerated contrast, hyper-detailed surreal aesthetic, and painterly over-rendering",
+    "dall_e":             "image with clean flat graphic style, smooth AI-blended gradients, slightly plastic surface quality, and uniformly lit commercial illustration look",
+    "unknown_generative": "image with subtle AI artifacts including unnatural smoothness, inconsistent frequency patterns, and synthetic pixel-level regularities absent in real photos",
 }
 
 FAKE_LABEL_KEYWORDS = (
@@ -185,7 +186,7 @@ class FingerprintEngine:
         if not _detectors:
             logger.warning("No fingerprint detectors loaded; using neutral fallback score.")
 
-        detector_weights = [0.5, 0.3, 0.2]
+        detector_weights = [0.4, 0.3, 0.2, 0.1]
         total_w = 0.0
         weighted_fake = 0.0
 
@@ -245,8 +246,8 @@ class FingerprintEngine:
             probs = logits.softmax(dim=0).cpu().numpy()
             max_prob = float(np.max(probs))
 
-            # Low confidence attribution → unknown generator
-            if max_prob < 0.25:
+            # Low confidence attribution → unknown generator (9 classes: chance=0.11, threshold=2.9×)
+            if max_prob < 0.32:
                 generator = "unknown_generative"
             else:
                 generator = list(GENERATOR_PROMPTS.keys())[int(np.argmax(probs))]