Optimize eviction candidate selection at scale

[DarlingtonDeveloper]

Problem

RetentionEngine::select_eviction_candidates() loads ALL nodes into memory, deserializes them, sorts by (importance ASC, created_at ASC), then takes the top N.

At 100k+ nodes this becomes expensive — full table scan + deserialize + sort.

Current code

crates/cortex-core/src/policies/retention.rs — select_eviction_candidates()

Fix options

1. Secondary index — maintain a NODES_BY_IMPORTANCE multimap table (bucketed f32 → node IDs), query lowest bucket first
2. Materialized priority queue — maintain a sorted eviction queue in a dedicated redb table, updated on put_node
3. Streaming sort with early exit — iterate by created_at (already ordered in redb), keep a bounded heap of size N by importance

Option 3 is simplest and avoids new indexes. Options 1/2 are O(N) on write but O(1) on eviction.

Priority

Low — only matters at scale (100k+ nodes). Current usage is well below this.