mirror of
https://github.com/stablyai/orca.git
synced 2026-10-03 16:02:11 +00:00
Merge Sim
af2e140e70
An eleven-PR chain merged main during a window in which main itself was broken; six nodes went red and the reds were diagnosed as merge damage. They were not. The tell was already in the data: an identical failure set on every node, including the root. Damage from a merge varies with what each node changed, so identical failures across a chain are upstream. main's own CI cannot answer this. pr.yml is `on: pull_request` only and unit-tests.yml is `workflow_call` only, so no commit on main carries a test result at all. The probe reconstructs main's health from the PRs whose CI ran against it in a window, classifies each check by whether its reds form one unbroken stretch in time (upstream) or interleave with greens (branch-specific), and reports broken / clean / unknown. It never reports clean without corroborating independent witnesses, and it refuses evidence whose checks ran more than a day apart, since those observed different mains. The roll-up `verify` job, the known-false reds and third-party app checks are excluded and printed, never dropped silently.