Most time: the tests/compiler run, which was the whole 379-test suite. I ran it twice, once to get the expected .stderr and once to confirm it passed. I also lost a few turns when nikaia check in tests/compiler checked the project instead of my file. That wrote another test’s error into my .stderr.
.stderr. --bless only writes .stdout, so the brief should say to copy the refusal block from the test run’s failure output, with the path form tests/NAME.nika:6:9.nikaia test has no filter and always runs the whole project.lower-std / build / lower-std cycle gives the lowered diff directly.nikaia check, the project test run, and the guard set..stderr attempt, and the project-wide test run used just to read one test’s output. I did not run the new test against the unfixed compiler. I judged it failing by construction, which is a shortcut the rule did not allow, and the report says so.pre-push-check.sh printed nothing, and I treated that as passing without seeing a positive result.builder.md for a new refusal test: write an empty .stderr, run the project once, copy the refusal block from the failure into it, then re-run. It should also say that the check against the unfixed compiler is done by running the test with the old binary copy. Separately, a push rejected for a non-fast-forward should tell the builder what to do, for example whether to fetch, since commit.nika only reports it.