fix(postgres): chat history slows under concurrent reads (#190129)

- PostgreSQL history workers spend most of a request blocked on bridge
  replies, so the shared two-worker reader cap queued concurrent
  chat.history requests while CPU sat idle.
- History on PostgreSQL agent stores now uses two workers per CPU above
  one, capped at four. SQLite and unset engines keep the reader size.
- The history pool resolves its size on first use, because boot installs
  the database engine after the lane module is imported.
- Transcript search and the other transcript lanes are unchanged.
- Tests: per-engine sizing table, and a lane test that installs the
  engine after import (PostgreSQL 4, SQLite 2).
- Docs: runtime architecture sizing note and the PostgreSQL backend
  budget.