THE SUITE IS THE REGRESSION NET FOR THIS FRAMEWORK and it lives in a container prefix rather than in buck-out, since the cases are guest binaries. The prefix tmp had been cleared since the last run, so re-establishing the net today meant redoing the staging by hand, and it went wrong in the way it always will. RESOURCES KEEP THEIR PATHS. A case asks for its resource by the path it has in the source tree, so DARLING_TESTSUITE_RESOURCE_PATH must point at a directory that still has testsuite/... underneath. Copying every resource directory into one flat directory makes test_read_file fail its fd >= 0 assertion, which reads exactly like a defect in our read and is nothing of the kind. That is the second time a rig has made that case look broken. THE RUNNER MUST NOT EAT THE CASE OUTPUT. An AppKit case prints two lines of backend greeting before anything of its own, so a runner that keeps the first two lines keeps nothing that matters, and a case that prints on its happy path looks as silent as one that never ran. The staged runner drops the backend lines and keeps three of the case own. MEASURED WITH THE SCRIPT, and it reproduces the state exactly: 69 case targets, 36 built, 36 run, 28 exit zero, AppKit 11 of 12. That is the same as before today's AppKit changes, so the table background fix, the alert button metrics and the view tree instrument cost nothing. The libxpc group still does not build and its one case that does still fails, which is unchanged and known.