Replaces the Gradio app with the build that ran on App Runner (main @ 2e95984).
- api/, core/, eval/, rules/, config.py, frontend/ taken verbatim from main
- Dockerfile from infra/docker/Dockerfile; bakes classifier-v2 and corrector-v4
into the image and sets the model env vars that App Runner set at runtime
- README app_port 7860 -> 8000
- ground-truth CSVs left out (not needed at runtime, over HF file size limit)
Rollback: git revert this commit to restore the Gradio app (5e60dbe).
feat: wire multilingual-bias-classifier-v1 into HF Space
- ml_classifier.py: all 6 languages now use juakazike/multilingual-bias-classifier-v1
(Davlan/afro-xlmr-base, 137K rows, F1=0.917, Jun 2026)
- gradio_app.py: update default ML model, threshold 0.56->0.75, metrics for all 6 langs
fix(ui): add HAUSA and ZULU to Language enum — fixes AttributeError on HF Space
- Add scoped malaya derogation patterns (ni malaya, malaya wewe/sana) — zero FPs on ground truth
- Add mwanaume ni kichwa partial proverb entry to catch clause without second half
- Add family possession patterns to BIOGRAPHICAL context condition to suppress
wake/yake → wao/yao warn edits on bibi yake, mdogo wake, mke wake etc.
- Relabel 5 mislabelled ground truth rows (false→true): sw-59930, sw-61890,
sw-61923, sw-61779, sw-52734
Metrics: SW F1=0.821 (↑0.002), P=0.741, R=0.919 — baseline preserved
Tests: 6/6 passing
fix: restore self-contained eval/context_checker.py — core import breaks HF Space