Parser fixes: reject multi-paragraph clean-passthrough, add \b to predicate regex

- clean-passthrough fast path now requires single Prolog paragraph (no \n\n) so
multilingual outputs like 'este(T) :- \+ oeste(T).\n\nExplicación: ...' don't
pass trailing prose to swipl
- predicate counting regex pos_pred\( now uses \b word boundary so 'este\('
doesn't over-match 'oeste\(' (Spanish/French/Italian etc where the positive
predicate is a substring of the negative)

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

ipt_verifier.py

@@ -143,13 +143,19 @@ def extract_hypothesis_with_meta(text: str, enable_line_parsing: bool = True) ->
             "structured_parse": False,
         }
 
-    # Fast path: input is already clean Prolog (no markup/thinking/markers).
-    # Avoids letting the line-extractor heuristics misfire on legitimate
-    # multi-line rule bodies — e.g. lines starting with uppercase variables
-    # like `CarA \= CarB,` that aren't recognised as continuations.
+    # Fast path: input is a SINGLE clean Prolog paragraph (no markup, no
+    # trailing prose). Avoids letting the line-extractor heuristics misfire on
+    # legitimate multi-line rule bodies — e.g. lines starting with uppercase
+    # variables like `CarA \= CarB,` that aren't recognised as continuations.
+    # The `\n\n` check rejects outputs like
+    #     este(T) :- \+ oeste(T).
+    #
+    #     Explicación: ...
+    # so multilingual / explanatory prose doesn't get passed to swipl.
     stripped = text.strip()
     if (
         stripped.endswith(".")
+        and "\n\n" not in stripped
         and "<think>" not in text
         and "</think>" not in text
         and "```" not in text
@@ -358,8 +364,11 @@ def verify(
     if enable_parsing:
         hypothesis = extract_hypothesis(hypothesis)
 
-    pos_examples = re.findall(rf"{pos_pred}\(([^)]+)\)", validation_program)
-    neg_examples = re.findall(rf"{neg_pred}\(([^)]+)\)", validation_program)
+    # \b word boundary prevents over-counting when one predicate name is a
+    # substring of the other (e.g. Spanish este/oeste, French est/ouest):
+    # without \b, `este\(` would also match `oeste(` and inflate pos_negs.
+    pos_examples = re.findall(rf"\b{pos_pred}\(([^)]+)\)", validation_program)
+    neg_examples = re.findall(rf"\b{neg_pred}\(([^)]+)\)", validation_program)
     arity = 1
     if pos_examples:
         arity = pos_examples[0].count(",") + 1