S0-R1-C4 Unified Phase-1 Validation Barrier + Retry Budget Ownership

core/generation_runner.py

@@ -3087,46 +3087,18 @@ The previous attempt for this same manuscript unit was rejected by validation. R
                         # Model prose has no authority over quote presence/count.
                         en_raw = self._strip_model_controlled_qualitative_quotes(en_raw, data_metadata)
 
-                # S0-R1 phase 1: feedback-only critic + pure semantic validators.
-                pre_validation_rows = self._section_validation_bundle(
-                    mode, agent, quant_engine, prefix + en_raw, data_metadata,
-                    section_id=section_id, section_name=display_name,
-                    evidence_contract=evidence_packet,
-                    include_critic=not deterministic_appendix,
-                )
-                if pre_validation_rows:
-                    if section_attempt < max_section_attempts:
-                        attempt_feedback = self._section_retry_feedback(pre_validation_rows)
-                        logging.warning(
-                            "Section validation retry | unit=%s attempt=%s count=%s types=%s",
-                            display_name, section_attempt, len(pre_validation_rows),
-                            sorted({str(row.get("type") or "") for row in pre_validation_rows if isinstance(row, dict)}),
-                        )
-                        self._restore_section_citation_state(
-                            agent, citation_plan, display_name, frozen_citation_state
-                        )
-                        continue
-                    self._restore_section_citation_state(
-                        agent, citation_plan, display_name, frozen_citation_state
-                    )
-                    state.update({
-                        "citation_plan": citation_plan,
-                        "agent_state": agent.export_state(),
-                        "quality_diagnostics": quality_diagnostics,
-                    })
-                    self._checkpoint_section_retry_exhausted(
-                        state, section_id=section_id, section_name=display_name,
-                        violations=pre_validation_rows, pct=pct, attempts=max_section_attempts,
-                    )
-                    raise RuntimeError(
-                        f"section_content_invalid_after_retry:{display_name}:{pre_validation_rows[:3]}"
-                    )
-
-                # S0-R1-C2: the authoritative path never attaches a citation merely
-                # to satisfy section-level coverage.  The writer-owned claim/evidence
-                # plan is audited as written; any validated coverage shortfall flows
-                # to the existing claim-anchored recovery path below.  This rule is
-                # shared by ASQ and AMJ and keeps the sealed English citation binding
+                # S0-R1-C4 phase-1 barrier: all writer-owned semantic validators
+                # inspect the same attempt before the single same-Section retry budget
+                # can be consumed. Citation semantics therefore run before the one
+                # unified validation decision instead of being short-circuited by a
+                # validator that happens to execute earlier.
+
+                # S0-R1-C2/C4: citation semantics are part of the same phase-1
+                # decision.  The authoritative path never attaches a citation merely
+                # to satisfy section-level coverage; it audits the writer-owned
+                # claim/evidence binding exactly as written. Any validated coverage
+                # shortfall still flows to the existing claim-anchored recovery path.
+                # This rule is shared by ASQ and AMJ and keeps sealed English
                 # authoritative for ZH/RU/KK translation.
                 self._progress(pct, f"citation_audit:{current_index}:{total_steps}")
                 used_before = set(agent.used_citations)
@@ -3134,20 +3106,33 @@ The previous attempt for this same manuscript unit was rejected by validation. R
                     prefix + en_raw, section_name=display_name, section_id=section_id,
                     quarantine=is_quarantined
                 )
-                # Capture immediately: citation recovery may later replace the
-                # agent's last-violation buffer. These violations belong to the
-                # writer attempt and require the one bounded same-Section retry.
+                # Capture immediately: later recovery may replace the agent's last-
+                # violation buffer.  Only citation-owned rejected tags may enter the
+                # retry exclusion set.
                 citation_semantic_rows = copy.deepcopy(
                     list(getattr(agent, "_last_citation_semantic_violations", []) or [])
                 )
-                if citation_semantic_rows:
+                phase1_validation_rows = self._section_validation_bundle(
+                    mode, agent, quant_engine, prefix + en_raw, data_metadata,
+                    section_id=section_id, section_name=display_name,
+                    evidence_contract=evidence_packet,
+                    include_critic=not deterministic_appendix,
+                    citation_violations=citation_semantic_rows,
+                )
+                non_citation_count = max(
+                    0, len(phase1_validation_rows) - len(citation_semantic_rows)
+                )
+                if phase1_validation_rows:
+                    retry_excluded_tags.update(self._citation_violation_tags(citation_semantic_rows))
                     if section_attempt < max_section_attempts:
-                        retry_excluded_tags.update(self._citation_violation_tags(citation_semantic_rows))
-                        attempt_feedback = self._section_retry_feedback(citation_semantic_rows)
+                        attempt_feedback = self._section_retry_feedback(phase1_validation_rows)
                         logging.warning(
-                            "Citation semantic Section retry | unit=%s attempt=%s count=%s types=%s",
-                            display_name, section_attempt, len(citation_semantic_rows),
-                            sorted({str(row.get("type") or "") for row in citation_semantic_rows if isinstance(row, dict)}),
+                            "Unified Phase-1 Section retry | unit=%s attempt=%s count=%s types=%s",
+                            display_name, section_attempt, len(phase1_validation_rows),
+                            sorted({
+                                str(row.get("type") or "")
+                                for row in phase1_validation_rows if isinstance(row, dict)
+                            }),
                         )
                         self._restore_section_citation_state(
                             agent, citation_plan, display_name, frozen_citation_state
@@ -3163,10 +3148,15 @@ The previous attempt for this same manuscript unit was rejected by validation. R
                     })
                     self._checkpoint_section_retry_exhausted(
                         state, section_id=section_id, section_name=display_name,
-                        violations=citation_semantic_rows, pct=pct, attempts=max_section_attempts,
+                        violations=phase1_validation_rows, pct=pct, attempts=max_section_attempts,
+                    )
+                    terminal_kind = (
+                        "section_citation_invalid_after_retry"
+                        if citation_semantic_rows and non_citation_count == 0
+                        else "section_content_invalid_after_retry"
                     )
                     raise RuntimeError(
-                        f"section_citation_invalid_after_retry:{display_name}:{citation_semantic_rows[:3]}"
+                        f"{terminal_kind}:{display_name}:{phase1_validation_rows[:3]}"
                     )
 
                 used_after = set(agent.used_citations)