1. What cost the most. Before the summary, it was reading enough to design the step. That meant grammar_types.nika, j1’s parser hook, the AST shapes, the kind names, how findings are built and the Rust walker. It filled one context. After that, the pre-existing defect cost about 8 turns: an action’s T value was not wrapped into its rule’s T?. Finding it and fixing it in check_function_action took that long.
2. What the brief could have handed me:
   • The kind names of Part II 10.3. The sketch’s Int does not exist; it is I64.
   • That the catch-all arm is else, not _.
   • That check_reflect_arm already reads an arm.
   • Where NK1164’s wording lives.
   • j1’s Finding/refuse_finding pattern in expand_grammar_types.
   • text_tiers::visit_block_mut as the walker for substituting types in actions.
   • That f[i64?](x) does not parse, so a test cannot pass a nullable type at the call.
   • How to run one tests/compiler program: nikaia run refuses a file inside the project and nikaia test has no filter, so I copied files to a scratch project.
3. Checks run: the probes on the old binary, lower-std twice, cargo check, the build, each new test against the unfixed binary, clippy, fmt, pre-push-check.sh, tests/compiler --both-settings, the it binary, and the parser unit tests.
   • Wasted: the parser unit tests. parser:: matches 0 tests.
   • Partly wasted: running it beside tests/compiler. Its rebuild of target/fast/nikaia made 5 tests fail for no real reason, and I had to re-run them.
4. One change: builder.md should say that tests/compiler programs spawn target/fast/nikaia, so nothing may rebuild it while they run, and it should give the command for running a single test program. A close second: the brief’s sketch should use the real kind names, which would have spared me a lookup.

This site is open source. Improve this page.