Persist semantic audit before candidate rollback

autoresearch/prefill/supervisor.py

@@ -106,6 +106,13 @@ def _select_repair_target(current: dict, ledger: dict) -> str:
         str(item.get("obligation_id", "")): str(item.get("parent_id", ""))
         for item in ledger.get("obligations", [])
     }
+    cursor = parents.get(current_target, "")
+    visited = set()
+    while cursor and cursor not in visited:
+        if cursor in leaves:
+            return cursor
+        visited.add(cursor)
+        cursor = parents.get(cursor, "")
 
     def distance_from_current(obligation_id: str) -> int:
         distance = 0
@@ -982,6 +989,12 @@ def append_result(path: Path, row: dict) -> None:
 
 
 def run_iteration(args, iteration: int) -> dict:
+    from scripts.agent_gan_repl import (
+        audit_ledger_semantic_duplicates,
+        load_proof_ledger,
+        save_proof_ledger,
+    )
+
     root = Path(__file__).resolve().parents[2]
     ar = Path(__file__).resolve().parent
     candidate_path = ar / "candidate.py"
@@ -997,6 +1010,20 @@ def run_iteration(args, iteration: int) -> dict:
     current = _candidate_snapshot(current_module)
     previous_candidate = candidate_path.read_bytes()
     previous_state = _backup(state_path)
+    ledger_object = load_proof_ledger(ledger_path)
+    semantic_rejections = (
+        audit_ledger_semantic_duplicates(ledger_object)
+        if ledger_object is not None else []
+    )
+    if ledger_object is not None and semantic_rejections:
+        save_proof_ledger(ledger_path, ledger_object)
+        for obligation_id, ancestor_id, score in semantic_rejections:
+            print(
+                "[autoresearch] phase=semantic-retro-reject "
+                f"id={obligation_id} duplicate_of={ancestor_id} "
+                f"score={score:.2f}",
+                flush=True,
+            )
     previous_ledger = _backup(ledger_path)
 
     proposed = current

tests/inference_engine/bench/test_autoresearch_supervisor.py

@@ -335,6 +335,36 @@ def test_host_candidate_targets_deepest_current_branch_leaf():
     assert candidate["prefill_compute_chunk_tokens"] == 256
 
 
+def test_host_candidate_rolls_back_to_nearest_valid_ancestor():
+    rejected_target = "RH-C2-duplicate-child"
+    current = {
+        **_candidate(),
+        "target_obligation_id": rejected_target,
+    }
+    ledger = {"obligations": [
+        {
+            "obligation_id": "RH-C1",
+            "statement": "Unrelated leaf.",
+            "status": "UNRESOLVED",
+            "parent_id": "",
+        },
+        {
+            "obligation_id": "RH-C2-gap",
+            "statement": "Falsify the density-singularity implication.",
+            "status": "UNRESOLVED",
+            "parent_id": "",
+        },
+        {
+            "obligation_id": rejected_target,
+            "statement": "Renamed density-singularity implication.",
+            "status": "REJECTED_DUPLICATE",
+            "parent_id": "RH-C2-gap",
+        },
+    ]}
+    candidate = build_host_candidate(current, ledger)
+    assert candidate["target_obligation_id"] == "RH-C2-gap"
+
+
 def test_strategy_is_triggered_only_by_events(tmp_path):
     progress = {
         "kept": "True",
@@ -546,6 +576,9 @@ def test_supervisor_preserves_runtime_and_cache_across_iterations():
         / "supervisor.py"
     ).read_text()
     body = source[source.index("def run_iteration"):source.index("def main")]
+    assert body.index("audit_ledger_semantic_duplicates(") < body.index(
+        "previous_ledger = _backup",
+    )
     assert body.index("check_runtime_health(") < body.index(
         "strategy_trigger_reason(",
     )