fix(ensemble): apply feedback_manager weight multipliers in combine_signals()

fix(mcmc): replace biased 1e-10 lower bound in MH acceptance draw

fix(own_embedding): wrap centroid pickle.load in try/except

fix(ensemble): log XGBoost feature-name mismatches before predict_proba

fix(ensemble): apply feedback_manager weight multipliers in combine_signals()

fix(ensemble): filter top_reasons by confidence > 0

fix(ensemble): move Platt calibration to weighted-sum fallback path only

fix: CLIP placeholder init logs at ERROR not INFO; add placeholder_mode to return dict (batch3 #4)

refactor: extract combine_signals() from detect() so SSE can reuse scoring without re-running detectors (audit #12)

feat: add GPT-4o, Flux, Imagen3, Ideogram, Recraft generator profiles and heuristics (audit #14)

fix: SSE now runs all 30 signals exactly once; DIRE and OwnEmbedding now appear as individual streamed events (audit #13)

refactor: extract combine_signals() from detect() so SSE can reuse scoring without re-running detectors (audit #12)

fix: DIRE now encodes to VAE latent before UNet; decode after denoise; align noise timestep (audit #11)

perf: vectorize DCT radial spectrum loop with np.bincount (audit #8)

fix: add raw_value and expected_range to noise_map success dict (audit #5)

fix: heatmap_generator path now uses __file__-relative, not CWD-relative (audit #4)

fix: cache Platt params at module level, add reload_params() (audit #2)

fix: add extension_valid key to validate_file() — was always True (audit #1)

fix: enable LFS in CI checkout + safe pickle load

fix: CLIP confidence scaling, attribution path, natural model cache cap

fix: own_result missing from DIRE branch and inconclusive classification zone

fix: DCT kurtosis threshold in wrong mathematical domain

fix: GPS rejects western hemisphere and JPEG Ghost PNG false positives

fix: methods_used fallback, HEIC extensions, CORS, Content-Length pre-check

fix: ensemble includes zero-confidence signals biasing PNG/WebP toward 0.5

fix: DIRE pixel/latent mismatch and timestep scheduling bug

fix: CLIP always returning 0.5 — move _load_model() before db_available check

fix: invert guilt condition in feedback_manager Nash weight update

feat: Nash Equilibrium adaptive detection feedback system

feat: segment-level AI detection endpoint

fix: NameError in image_forensics — use hashes[sha256] not bare sha256

fix: NameError in image_forensics — file_hash -> sha256

fix: Platt calibration sign convention and test isolation

feat: stable UUID5 evidence IDs derived from file SHA-256

feat: wire MCMC and Platt calibration into ensemble detector

feat: Platt scaling confidence calibration

feat: MCMC probabilistic authenticity engine

fix: add missing threading import in statistical_detector

fix: sanitize numpy types, forensics isolation, admin warning, docstring, remove dotenv

fix: DIRE confidence fragile check and missing available flag

perf: replace O(n^2) radial FFT loops with vectorized np.bincount

fix: KL divergence missing fftshift and ELA neutral score for lossless

fix: three thread-safety bugs — global random seed, race condition, unseeded sampling

fix: inter-block blockiness and missing method keys in advanced_ai_detector

fix: double-stacking boost and wrong blockiness metric in ai_detector

fix: resolve model/database paths relative to __file__ not CWD

fix: CLIP detection always disabled — db_available never set True

feat: Noiseprint camera fingerprint and CFA Bayer artifact detectors

fix(phase-20): wire jpeg_ghost and noise_map into ensemble scoring

fix(phase-20): add missing jpeg_ghost_detector.py and noise_map_detector.py