tests/compiler package, which the gate runs in full whenever a check_*.nika changes. I ran it three times:
target/fast/nikaia under it.Before that, the costliest was the fallout of putting element reads in view_of: a_changing_view also reads view_of, so lower-std refused emit_sums.nika with NK2108. Fixing it took a bootstrap round (lower-std with the old binary, then a rebuild). Row e cost a probe cycle too: the brief’s read_a_map is set only by get, not by m[k].
view_of has a second reader, a_changing_view in check_takes.nika.facts.method_callees keys (Vec::last, HashMap::get, …) and facts.map_read_views for map brackets.lowered.rs already held another agent’s hunk, so my new helper had to be a method, which stays out of the index.commit.nika now leaves the commit local (Pushed: local), while builder.md and the brief still say it pushes.fmt --check on both new tests, and the lower-std fixpoint.it filters (309 passed) and nikaia-std --lib check_.Wasted: the second and third whole-package runs. Shared-machine noise was already the likely cause; running only the failing tests settled it.
commit.nika), the gate’s test step runs only the touched tests even when check/ changed. A whole-package red that a one-test copy can’t reproduce should be reported as flaky, not as Needs: fix. The brief should also list every reader of a field it tells you to extend (here view_of → a_changing_view).