Applying the full schema baseline on every fresh test database cost ~530ms
per file and dominated PG gate wall-time under fork contention. Apply the
schema once per worker process into a pid-keyed template database, then
create each test DB via CREATE DATABASE ... TEMPLATE (fast server-side copy).
Dead-pid templates from crashed/prior runs are swept before creating a new
one; template copies are serialized since CREATE DATABASE ... TEMPLATE forbids
concurrent access to the source. Removes the throwaway probe2 timing test.
Fusion-Task-Id: FN-8146
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
c64c1014e2