fix: ship persistent bgp ai briefs and optimize bgp queries

rules.md

@@ -224,3 +224,18 @@ class BaseCollector:
 - Dependency injection for testability
 - Feature flags for incomplete features
 - Use config files for environment-specific settings
+
+---
+
+## Query Performance - MANDATORY
+
+- **NEVER** load whole tables into Python just to do filtering, pagination, counting, dedupe, or summary aggregation
+- Filters, sorting, pagination, `count`, `distinct`, and grouped statistics **MUST** be pushed down to the database whenever the ORM/query builder can express them
+- Summary/dashboard endpoints should prefer dedicated aggregate queries or aggregate endpoints, not multiple full-table scans
+- For hot paths, avoid selecting large JSON/text payload columns unless the response really needs them
+- If an endpoint returns a list, default to database-side pagination instead of `scalars().all()` followed by Python slicing
+- When you suspect a query is slow, first check for:
+  - full-table ORM loads
+  - Python-side post-filtering
+  - repeated summary queries that can be merged
+  - repeated per-request recomputation that should be cached or aggregated once