Fix sentinel edge cases: hallucination combo guard + UI formatting

Replace ad-hoc refusal regexes with NOT IN DOCUMENTS sentinel

Fix refusal detector missing 'not able to' and 'outside scope' phrases

Replace HHEM with sentence-level NLI, add claim decomposition and drift detection

Fix compat, bugs, and types; expand retail KB

Switch faithfulness to text_pair encoding, promote score logging to INFO

Fix Vectara label check and input format

Load T5-small tokenizer for Vectara HHEM v2

Use T5Tokenizer directly for Vectara HHEM v2

Fix Vectara pipeline: explicitly load tokenizer before pipeline init

Load Vectara model via transformers pipeline, not CrossEncoder

Add trust_remote_code=True for Vectara hallucination model

Switch faithfulness grader to Vectara hallucination evaluation model

Faithfulness: mean sentence scoring, strip chunk title prefix, lower threshold to 0.35

Address Gate 5 audit gaps

Fix faithfulness: score per chunk, take max entailment

Replace Anthropic with free-tier stack

Add full RAG evaluation pipeline with L1 metrics and UI