1. What cost the most: waiting on the tests/compiler --both-settings run (about 10 minutes, 398 programs), not turns. The one real design step was the Rust ComputeHost. Its units are numbered by position in a list, and the sub-program’s unit is not next to the others in that list. So driven needed lowered_units to take &[&Parsed], and without_readers needed a helper that works on units picked out by index. One cargo check found nothing more than two now-unused re-exports.
2. What I looked up that the brief could have given me:
   • The Wall variants (InsideAnother { grammar } maps to name).
   • The exact Rust signatures of the lowered ComptimeComputeHost (&impl Fn on the trait, &dyn Fn on the functions).
   • That lowered_units and NOT_COMPUTED_YET have no other callers, while without_readers is also called from grammar_run.rs:336.
   • That the K3a row of docs/selfhosting-plan.md already held another session’s uncommitted note.
   • The path of the rs2nika binary for --score.
3. Checks:
   • Useful: cargo check -p nikaia-std --features compiler right after the first lower-std (green at once), cargo check -p nikaia, one build cycle with a stable fixpoint, the it filters next to clippy, the gate, tests/compiler, and pre-push-check.
   • Wasted: none really. I could have skipped badges.py --check, since pre-push-check covers it.
4. One change: give the brief a line on what the Rust host needs from code outside the cut. The arena’s units are not next to each other in file order, so lowered_units/without_readers must take &[&Parsed] or &mut [&mut Parsed]. Also say that the consts’ Rust re-exports become dead, and clippy refuses dead ones. A builder could then write the host in one pass without rereading the call sites.
   • Separately: when the brief orders the code taken from a patch, say whether --score is still expected to pass 60%. Here it measured how far the host-call shape is from the Rust, not what the converter is worth.

This site is open source. Improve this page.