R14-TQ1-R4-R3: align generator and validator lexical contract

core/pipeline_agent.py

@@ -3751,6 +3751,23 @@ Output only the repaired section."""
             'contamination_policy': 'Treat sustained untranslated English prose as a repair signal; normal acronyms, model names and protected terms are allowed.',
         }
 
+    @staticmethod
+    def _translation_lexical_retention_contract(target_lang):
+        """Single prompt-side authority for legal Latin retention in ZH/RU/KK.
+
+        The deterministic lexical-purity validator remains the enforcement
+        authority. This contract mirrors that policy so generation, QA, repair,
+        source regeneration, and adjudication are instructed by the same rule.
+        """
+        return (
+            'LEXICAL RETENTION CONTRACT: In Chinese, Russian, or Kazakh target-language prose, '
+            'every ordinary English lexical item must be translated. Latin script may remain only '
+            'for protected immutable research objects, approved acronyms, model/tool/variable identifiers, '
+            'citations, DOI/URL strings, source-matched proper names or bounded titles/institutions, and '
+            'legitimate Latin glossary targets recognized by the deterministic lexical-purity policy. '
+            'Do not retain ordinary English words or phrases merely because they appear technical or familiar.'
+        )
+
     @classmethod
     def _translation_language_rule(cls, target_lang):
         profile = cls._translation_language_profile(target_lang)
@@ -3760,7 +3777,11 @@ Output only the repaired section."""
             core = 'Use idiomatic academic Kazakh with correct case and possessive suffixes, person/plural endings, vowel harmony, consonant assimilation, natural word order, and correct Kazakh-specific letters. Translate directly from English and avoid Russian syntactic calques.'
         else:
             core = 'Use formal academic Simplified Chinese with explicit concepts, natural Chinese syntax, clear hierarchy, and logically connected claims. Avoid carrying English nominalization and clause order into Chinese mechanically.'
-        return core + ' ' + profile['term_policy'] + ' ' + profile['contamination_policy']
+        return (
+            core + ' ' + profile['term_policy'] + ' '
+            + cls._translation_lexical_retention_contract(target_lang) + ' '
+            + profile['contamination_policy']
+        )
 
     @classmethod
     def _russian_contamination_evidence(cls, source, target, glossary=None):
@@ -5062,10 +5083,11 @@ SECTION: {section_name}
                 'items': [{'id': row['id'], 'english': row['english'], 'target': row['target']} for row in rows],
             }
             profile = self._translation_language_profile(target_lang)
+            lexical_contract = self._translation_lexical_retention_contract(target_lang)
             system_prompt = f"""You are a bilingual academic translation QA classifier for {target_lang}.
 Review EACH claim independently. Do not rewrite and do not provide free-form explanations. Check semantic equivalence and language quality: omissions/additions, negation, uncertainty, association-versus-causation strength, terminology, grammar, language contamination, citation attachment, and acronym attachment.
-LANGUAGE ADAPTER: {profile['adapter_id']}; morphology={profile['morphology']}. {profile['term_policy']} {profile['contamination_policy']}
-For Russian, do not flag a terminology error merely because a translated term is inflected for grammatical case/number/gender. Russian LANGUAGE_CONTAMINATION includes sustained unprotected English prose and ordinary untranslated English lexemes; never use it for protected acronyms, AI/LLM/SEM, model or tool names, variable names, citations, canonical glossary terms, or source-matched proper names/titles. For Chinese, ordinary untranslated English lexical words are contamination while protected technical objects remain legal. For Kazakh, do not flag a terminology error merely because productive suffixes attach to the correct term stem. Kazakh LANGUAGE_CONTAMINATION includes ordinary untranslated English lexemes and clearly Russian lexical/syntactic intrusions not justified by shared international terminology; Cyrillic script or absence of Kazakh-specific letters alone is never sufficient evidence.
+LANGUAGE ADAPTER: {profile['adapter_id']}; morphology={profile['morphology']}. {profile['term_policy']} {lexical_contract} {profile['contamination_policy']}
+For Russian, do not flag a terminology error merely because a translated term is inflected for grammatical case/number/gender; sustained non-Russian prose remains an additional contamination signal. For Kazakh, do not flag a terminology error merely because productive suffixes attach to the correct term stem; clearly Russian lexical/syntactic intrusion remains an additional language-specific signal when the pair supports that conclusion. Cyrillic script or absence of Kazakh-specific letters alone is never sufficient evidence.
 Return valid JSON only in this exact form: {{"items":[{{"id":"SP-P000001-C000001","ok":true,"codes":[],"evidence_codes":[]}}]}}.
 If ok=false, codes MUST contain only values from allowed_issue_codes. For SEMANTIC_DRIFT, NEGATION_DRIFT, CAUSALITY_STRENGTH_DRIFT, UNCERTAINTY_DRIFT, OMITTED_CLAIM, or ADDED_CLAIM, evidence_codes should identify the affected semantic dimension using only: {sorted(allowed_semantic_evidence)}. For GRAMMAR, evidence_codes should identify the narrow grammatical defect using only: {sorted(allowed_grammar_evidence)}. Use the narrowest supported evidence; use MEANING_CHANGED or GRAMMAR_UNSPECIFIED only when the narrower dimension cannot be identified from the pair. Do not invent an error dimension. Never return numbered prose, examples, quotes, or corrective wording. Never return replacement text. Return every supplied id exactly once."""
             response = self.call_deepseek(
@@ -5604,12 +5626,13 @@ Each code must be one of: {sorted(allowed)}. Return one or more specific dimensi
         ]
         if grammar_evidence_codes:
             evidence_payload['grammar_evidence_codes'] = grammar_evidence_codes[:4]
+        lexical_contract = self._translation_lexical_retention_contract(target_lang)
         retry_mode = (
             '\nCLAIM SOURCE REGENERATION. The previous target wording has failed bounded repair and MUST NOT be used as drafting context. Generate a fresh translation of only this one claim directly from the frozen marked English source. '
             if strategy == 'SOURCE_REGENERATE' else
             '\nCLAIM QUALITY RETRY. Repair only this one claim. '
         )
-        effective_prompt = system_prompt + retry_mode + (
+        effective_prompt = system_prompt + '\n' + lexical_contract + retry_mode + (
             'The only permitted repair signals are the following bounded issue codes: '
             + json.dumps(codes[:5], ensure_ascii=False)
             + '. Use the bounded detector evidence below only to repair the supplied claim; do not invent additional defects. '
@@ -5859,9 +5882,11 @@ Each code must be one of: {sorted(allowed)}. Return one or more specific dimensi
                 for row in model_rows
             ],
         }
+        lexical_contract = self._translation_lexical_retention_contract(target_lang)
         system_prompt = f"""You are an independent bilingual translation QA adjudicator for {target_lang}.
 This is verification-only: do not rewrite, suggest wording, or emit free-form explanations. Each supplied issue already survived two bounded claim-local repair rounds. Confirm a BLOCKING defect only when the English/target pair itself provides clear evidence. Do not treat stylistic preference or a legitimate grammatical/morphological variant as blocking.
-For Russian, allow required case/number/gender inflection. For Kazakh, allow productive case, possessive, person/plural and vowel-harmony suffixes on the correct term stem, but treat clearly Russian lexical/syntactic substitution as blocking only when the pair itself supports that conclusion. For Chinese, allow natural reordering and sentence reshaping inside the same claim while preserving meaning. For all three languages, an ordinary untranslated English lexical word is blocking; protected acronyms, model/tool names, variables, citations, canonical glossary terms, and source-matched proper names are not.
+{lexical_contract}
+For Russian, allow required case/number/gender inflection and retain sustained non-Russian prose as an additional contamination signal. For Kazakh, allow productive case, possessive, person/plural and vowel-harmony suffixes on the correct term stem, but treat clearly Russian lexical/syntactic substitution as blocking only when the pair itself supports that conclusion. For Chinese, allow natural reordering and sentence reshaping inside the same claim while preserving meaning.
 Return valid JSON only: {{"items":[{{"id":"SP-P000001-C000001","issue_code":"GRAMMAR","decision":"CONFIRMED_BLOCKING","evidence_code":"GRAMMATICAL_VIOLATION"}}]}}.
 Decision must be CONFIRMED_BLOCKING or NOT_CONFIRMED. evidence_code must be one of: {sorted(allowed_evidence)}. Return every supplied id+issue_code pair exactly once."""
         try: