feat: integrate Ontology pre-LLM filter into Pass 0b submit

Obligations classified before API call:
- evidence → skipped (saves API cost)
- composite → skipped (not atomic)
- framework_container → skipped (decompose separately)
- atomic → sent to LLM

Filter stats returned in submit response.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

control-pipeline/services/decomposition_pass.py

@@ -3640,6 +3640,33 @@ class DecompositionPass:
         if not prepared:
             return {"status": "empty", "total_candidates": 0}
 
+        # Pre-LLM filter: skip evidence, containers, framework references
+        from services.control_ontology import classify_obligation
+        atomic_prepared = []
+        filtered_stats = {"evidence": 0, "composite": 0, "framework_container": 0, "atomic": 0}
+        for obl in prepared:
+            classification = classify_obligation(
+                obl["obligation_text"], obl.get("action", "")
+            )
+            routing = classification["routing"]
+            filtered_stats[routing] = filtered_stats.get(routing, 0) + 1
+            if routing == "atomic":
+                atomic_prepared.append(obl)
+            else:
+                logger.info("Pre-LLM filter: %s skipped (%s): %s",
+                            obl["candidate_id"], routing, obl["obligation_text"][:80])
+
+        logger.info("Pre-LLM filter: %d → %d atomic (skipped: %d evidence, %d composite, %d framework)",
+                     len(prepared), len(atomic_prepared),
+                     filtered_stats.get("evidence", 0),
+                     filtered_stats.get("composite", 0),
+                     filtered_stats.get("framework_container", 0))
+
+        prepared = atomic_prepared
+        if not prepared:
+            return {"status": "empty", "total_candidates": 0,
+                    "filtered": filtered_stats}
+
         requests = []
         for i in range(0, len(prepared), batch_size):
             batch = prepared[i : i + batch_size]
@@ -3686,6 +3713,7 @@ class DecompositionPass:
             "total_candidates": len(prepared),
             "total_requests": len(requests),
             "batch_size": batch_size,
+            "pre_filter": filtered_stats,
         }
 
     async def process_batch_results(