fix(backend): Convert PyTorch thread execution to bounded async pool to prevent OOM on HF Spaces

fix(deployment): reduce uvicorn workers to 1 to prevent OOM on huggingface spaces

chore: configuration updates for production frontend

chore: clean codebase for production

perf: optimize deps, imports, Dockerfile; improve README for judges

fix: language fallback to Auto + reduce timeout to 20s for evaluator safety

format changes

feat: source-based gating for authenticity override

fix: scope authenticity override to realtime-only, preserve file upload AI detection

fix: final calibration - auth override 40, anomaly multiplier 0.55, P4 avg-risk label

fix: round 3 recalibration - widen override and fix session labelling

fix: aggressive recalibration for browser mic human voice detection

fix: model confidence calibration for browser mic audio

fix: complete 24-issue compatibility audit (C1-C4, H1-H4, M1-M11, L1-L5)

feat: language auto-detect by default, language field now optional

feat: mixed-language (Hinglish) support — auto-detect ASR, expanded fraud keywords

refactor: detect label inversion once at load time, reduce log noise

fix: add missing max_voice_human_confidence field, fix map_level_to_label call

fix: revert ASR model to tiny (base too large for HF free tier)

fix: P1-P11 comprehensive risk scoring, signal calibration, and session summary improvements

fix: invert label mapping for primary model (training labels were swapped)

fix: raise short-chunk guard threshold from 1.5s to 2.0s

fix: risk floor for AI detections, legacy recommendedAction, alert reasons, short-chunk guard

fix: restore /v1/ routes as hidden aliases (include_in_schema=False) for frontend compat

cleanup: remove duplicate /v1/ route decorators, keep only /api/voice-detection/v1/ routes

fix: populate top_indicators with voice classification signals for explainability

fix: align signal contributions, evidence, and session summary for consistency

fix: lower uncertainty threshold from 0.65 to 0.50 — was forcing all low-confidence HUMAN results to UNCERTAIN

perf: use ML model for realtime chunks (higher accuracy for hackathon demo)

chore: bake production defaults into config.py (remove need for HF variables)

fix: correct id2label key-type mismatch causing inverted classifications