fix(sre): bind backup DR evidence into MCP

apps/api/src/jobs/agent99_controlled_dispatch_reconciler_job.py

@@ -503,6 +503,8 @@ async def _ensure_mcp_evidence_writeback(
         "rag_writeback_ack_id",
         "playbook_trust_writeback_ack_id",
     }
+    if "backup_health" in identity.route_id:
+        required.add("dr_scorecard_writeback_ack_id")
     safe_refs = sanitize_agent99_public_receipt_refs(
         receipt_refs,
         allowed_keys=required,
@@ -794,13 +796,6 @@ async def _finalize_learning(
             "telegram_lifecycle_receipt_id",
             lambda: _ensure_telegram_receipt(identity, mode=mode),
         ),
-        (
-            "mcp_evidence_writeback_ack_id",
-            lambda: _ensure_mcp_evidence_writeback(
-                identity,
-                receipt_refs=receipt_refs,
-            ),
-        ),
     ]
     if "backup_health" in identity.route_id:
         assets.append((
@@ -810,6 +805,18 @@ async def _finalize_learning(
                 evidence_refs=evidence_refs,
             ),
         ))
+    # MCP is the final evidence binding: for backup routes it must include the
+    # already-persisted DR scorecard acknowledgement as well as Telegram,
+    # KM/RAG and PlayBook.  Writing MCP first would create a durable but
+    # incomplete evidence graph that could never prove the DR asset belonged
+    # to this same run.
+    assets.append((
+        "mcp_evidence_writeback_ack_id",
+        lambda: _ensure_mcp_evidence_writeback(
+            identity,
+            receipt_refs=receipt_refs,
+        ),
+    ))
     for key, writer in assets:
         if not await ensure_checkpointed(key, writer):
             return {