Switch faithfulness to text_pair encoding, promote score logging to INFO

text_pair passes both sequences to T5Tokenizer separately — it inserts
the </s> separator between them, matching T5's pre-training format.
Concatenated string skipped that separator, likely causing the model to
receive malformed input and score faithful responses at ~0.14.

INFO log shows (label, score) per chunk — visible in HF Spaces logs
for threshold calibration without a debug flag.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

backend/grader.py

@@ -163,11 +163,10 @@ def grade_faithfulness(response: str, context: str) -> GradeResult:
     if not raw_chunks:
         return GradeResult(metric="faithfulness", passed=False, score=0.0, detail="No context")
     chunks = [_strip_chunk_title(c) for c in raw_chunks]
-    # Vectara HHEM v2: single concatenated string avoids T5 text_pair encoding issues.
-    # Label "Factually Consistent" = faithful; use startswith to avoid "inconsistent" false match.
-    inputs = [f"{chunk} {response}" for chunk in chunks]
-    results = model(inputs)
-    log.debug("Vectara raw results: %s", results)
+    # text_pair encodes sequences with T5 </s> separator — correct for T5-based models.
+    pairs = [{"text": chunk, "text_pair": response} for chunk in chunks]
+    results = model(pairs)
+    log.info("Vectara raw: %s", [(r["label"], round(r["score"], 3)) for r in results])
     scores = [
         r["score"] if r["label"].lower().startswith("factually consistent") else 1.0 - r["score"]
         for r in results