1. Time and turns. Nothing was expensive. The slow part was the verification trap. I copied target/fast/nikaia as the “old” binary, but it had already been built with run 1’s uncommitted edit. So my first fails-before run passed, and I only found out after two extra probes. The one cost of that was finding a truly unfixed binary (/home/user/nikaia-suite/target/fast/nikaia). The 4-minute build cycle was also on the critical path.

2. Lookups. The brief could have told me three things:
   • The tree’s binary already contained run 1’s edit, so a truly unfixed binary had to come from somewhere else.
   • The text parts of an f-string are raw source text, so quotes and backslashes are already escaped. That is why no second escaper is needed. I had to read format_of and the LitString arm to confirm it.
   • My first guess at the test’s expected length was wrong (I wrote 16, the right figure is 14). A line in the brief working it out would have saved a pass.
3. Checks.
   • Useful: lower-std and its fixpoint, the one-program test on the fixed binary, the probe on the unfixed binary, and pre-push.nika.
   • Wasted: the “OLD” test run with my copied binary. It proved nothing.
   • Skipped: clippy and the Rust unit tests. I changed no Rust by hand, only the lowered twin. I did not grep the it tests for the old lowering text (.to_string() of a hole-less f-string).
   • The it tests could still be red if one asserts the doubled-brace text. I didn’t check.
4. One change.
   • Add to the facts: “When a brief says a previous run left an uncommitted edit, build the unfixed binary first by running git stash on your own files only, or use the nikaia-suite binary. target/fast/nikaia may already contain the edit.” That would have saved the confused probe.
   • builder.md should say which binary is the canonical unfixed one when the tree is dirty.

This site is open source. Improve this page.