Both startup defects fixed in #988 were invisible in CI because e2e always boots a pristine profile. The failure class is call-frequency × per-call-cost × data-size — and CI has all three near zero.
Add one startup e2e seeded with a worn fixture: multi-MB config.json (dummy llm-history), a workspace with a few thousand files, hundreds of gitignored dirs (e.g. generated .wdio-*-style names), several tracked projects. Assert time-to-interactive under a threshold and that no Copse process exceeds a memory ceiling during the smoke window.
Both startup defects fixed in #988 were invisible in CI because e2e always boots a pristine profile. The failure class is
call-frequency × per-call-cost × data-size— and CI has all three near zero.Add one startup e2e seeded with a worn fixture: multi-MB config.json (dummy llm-history), a workspace with a few thousand files, hundreds of gitignored dirs (e.g. generated
.wdio-*-style names), several tracked projects. Assert time-to-interactive under a threshold and that no Copse process exceeds a memory ceiling during the smoke window.