- What cost the most: the lowering cache. After my first edit,
prover_bodies and prover_passes were lowered as async and the build failed. A touch and a cache-salt run did not refresh the stale prover_passes.rs; only declaring both functions sync fixed it. That took about 8 turns and two extra lower-std runs. Second came step 2: three probes to learn that peek cannot be bound and that a published field may not hold both kinds of text, both after the step was already written.
- What the brief could have handed me:
- That the prover never walks grammar actions (
hold_unreached in prover_passes.nika says so outright), and where to add the walk: prover_bodies, plus prover_if as the pattern for a guard becoming a fact.
- ADR-282 D22 (no published field of both kinds), which contradicts the
Scalar.value it asked for.
- The Part II 10 element table, and that winnow-grammar refuses binding
peek, not and => (winnow-grammar-model/src/parser.rs).
- That
nikaia fmt does not cover tools/*.nika, and that commit.nika takes the attribution from the brief’s own report name.
- Checks:
- Useful: the scratch probes with
check --asserts, lower-std, tests/compiler --both-settings (it caught my .stderr column), the it guard set plus a grammar filter, clippy, pre-push-check, and the yaml and json package tests.
- Wasted:
cargo test --lib prove (0 tests), the NIKAIA_LOWER_CACHE_SALT run, and the step-2 yaml check (refused before I knew about D22).
- Weak: I rewrote the
.stderr from captured output without a second run, and the refusal test ran on the unfixed compiler only as a probe of the same form, not as the file itself.
- One change that would have saved the most: a pre-flight line in the brief, “probe every grammar form the step needs (
x:peek(text(..)), a mixed pub field) before writing it”, and an architect check of the brief’s design against ADR-282 D22. With both, step 2’s blockers would have shown up in the first five turns instead of after the code was written.
This site is open source.
Improve this page.
