test_concurrent_inserts_settle_at_cap created 160 real AIAgent instances
(8 threads × 20 inserts). Each real init does tool-registry scan, MCP
discovery, config loading and file I/O — ~200 ms — which exhausts the
30 s pytest-timeout budget under parallel xdist load on a saturated
runner (load avg 16-37 on 8 CPUs).
Replace _real_agent() with a _StubAgent that exposes only the two
attributes the cache enforcement logic actually touches: .client
(non-None, checked by active-safety tests) and .close() (cleanup).
The stub has no init work and runs in microseconds, so the concurrent
insert test finishes in <1 s even under xdist pressure.
Sister tests test_fill_to_cap_then_spillover and
test_spillover_all_active_keeps_cache_over_cap use the same helper
and benefit equally from the lighter fixture.
Closes#5.
Molecule AI Dev Engineer A (Kimi)
7aea9c32b7