1. What cost the most. The 9-minute tests/compiler --both-settings run cost the most time. The most turns went to working out how to run one test program on its own. nikaia run tests/X.nika refuses because it is not the entry point, and nikaia test has no name filter, so I wrote a scratch-project script. Finding where a type’s parts are stored took about 5 turns: state.structs, enum_payloads and world.parts_beside.

2. Lookups the brief could have handed me:
   • The text of ADR-388 D2. It is short, so it could have been quoted.
   • Where the catalogue entry goes. The brief says “page”, but it is one long table row in 30-nikaia-tooling.md, inserted before NK1331.
   • The type-parts maps above, and any_takes_space in check_decls.nika as a walk to copy.
   • The check_refusal signature, and Ty::owned().
   • That a .stderr ends with the N type errors line.
3. Checks I ran:
   • Needed: the probes on the old binary (they caught the owned-table NK1102 defect and the Symbol? defect before I built anything), two build cycles, the it tests on errors:: and grammar, clippy, fmt and pre-push.
   • Mostly wasted: the full tests/compiler run. It was 9 minutes, and the brief asked for it while builder.md says to run only touched tests. The three changed programs had already passed by probe.
   • A real slip: I piped pre-push into tail, so its exit code did not stop the commit. dfb6502e went in while pre-push was red on the out-of-date self-hosted badge. That badge is a known trap and yours to fix, but the check was not gating.
4. One change: add a documented way to run a single tests/compiler program, for example a nikaia test --only NAME filter or a scripts/one_test.nika NAME. Also settle in the brief whether to run the full tests/compiler or only the touched programs, because the brief and builder.md contradict each other.

This site is open source. Improve this page.