feat(P03 W2): outcome backfill (REQ-317) + escalation_reason (REQ-318)

REQ-317: core/metrics/outcome_backfill.py backfills fact_decision.outcome
pending -> succeeded/failed after run.completed/run.failed; idempotent +
terminal (does not overwrite a non-pending outcome); wired into the
collector. The Post-Pilot AI Decision Accuracy denominator is now grounded
(fact_decision.outcome is not stuck pending).

REQ-318: ai.decision.made on a block band carries escalation_reason:
'confidence' (the only value in v1.26 — a block is always confidence-
driven; future milestones may add 'policy'). Persisted into fact_run by
the collector. The Post-Pilot Human Escalation Frequency denominator is
now grounded.

---ci---
project: acdl
phase: 3
milestone: v1.26
status: execute
wave: W2
---

core/metrics/collector.py

@@ -54,7 +54,8 @@ def _init_store(db_path=None):
             confidence_band TEXT,
             hitl_block INTEGER,
             cost_estimate_usd REAL,
-            decision_id TEXT
+            decision_id TEXT,
+            escalation_reason TEXT
         );
 
         CREATE TABLE IF NOT EXISTS fact_capability (
@@ -110,7 +111,9 @@ def _init_store(db_path=None):
             confidence REAL,
             alternatives TEXT,
             human_override INTEGER,
+            escalation_reason TEXT,
             outcome TEXT,
+            backfilled_at TEXT,
             event_time TEXT,
             PRIMARY KEY (decision_id)
         );
@@ -217,14 +220,15 @@ def collect_run_manifests(db_path=None, runs_dir=None):
             INSERT OR REPLACE INTO fact_run
             (run_id, contract_id, environment, started_at, completed_at,
              exit_code, outcome, confidence_score, confidence_band,
-             hitl_block, cost_estimate_usd, decision_id)
-            VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
+             hitl_block, cost_estimate_usd, decision_id, escalation_reason)
+            VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
         """, (run_id, manifest.get("contract_id", ""), manifest.get("environment", ""),
               manifest.get("started_at", ""), manifest.get("completed_at", ""),
               manifest.get("exit_code", 0), manifest.get("outcome", ""),
               conf.get("score", 0), conf.get("band", ""),
               1 if hitl.get("block") else 0,
-              manifest.get("cost_estimate_usd", 0), manifest.get("decision_id", "")))
+              manifest.get("cost_estimate_usd", 0), manifest.get("decision_id", ""),
+              manifest.get("escalation_reason")))
         count += 1
     conn.commit()
     conn.close()
@@ -232,7 +236,16 @@ def collect_run_manifests(db_path=None, runs_dir=None):
 
 
 def collect_decision_ledger(db_path=None, ledger_db=None):
-    """Read the Decision Ledger SQLite → fact_decision."""
+    """Read the Decision Ledger SQLite → fact_decision.
+
+    REQ-317: preserves a backfilled outcome. The ledger is append-only
+    and the `nova.ai.decision.made` event always carries outcome=pending
+    (it is emitted before apply). Once `outcome_backfill.backfill()` has
+    transitioned the `fact_decision` row to succeeded/failed, a re-run of
+    the collector must NOT clobber it back to pending. We therefore
+    coalesce: if the existing row has a non-pending outcome, keep it +
+    its backfilled_at; otherwise write pending (the event default).
+    """
     if db_path is None:
         db_path = _STORE_PATH
     if ledger_db is None:
@@ -251,15 +264,27 @@ def collect_decision_ledger(db_path=None, ledger_db=None):
         payload = json.loads(payload_json)
         data = payload.get("data", {})
         decision_id = data.get("decision_id", run_id)
+        # Preserve a backfilled outcome across collector re-runs (REQ-317).
+        existing = conn.execute(
+            "SELECT outcome, backfilled_at FROM fact_decision WHERE decision_id = ?",
+            (decision_id,),
+        ).fetchone()
+        if existing and existing[0] and existing[0] != "pending":
+            outcome = existing[0]
+            backfilled_at = existing[1]
+        else:
+            outcome = data.get("outcome", "pending")
+            backfilled_at = data.get("backfilled_at")
         conn.execute("""
             INSERT OR REPLACE INTO fact_decision
             (decision_id, run_id, chosen_action, confidence, alternatives,
-             human_override, outcome, event_time)
-            VALUES (?, ?, ?, ?, ?, ?, ?, ?)
+             human_override, escalation_reason, outcome, backfilled_at, event_time)
+            VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
         """, (decision_id, run_id, data.get("chosen_action", ""),
               data.get("confidence", 0), json.dumps(data.get("alternatives", {})),
               1 if data.get("human_override") else 0,
-              data.get("outcome", "pending"), event_time))
+              data.get("escalation_reason"),
+              outcome, backfilled_at, event_time))
         count += 1
     conn.commit()
     conn.close()