1. What cost the most. Three build cycles that each found one more layer of the same problem, so the checker fix needed three passes:
   • the #997 look-ahead refused callees that keep their argument;
   • rustc would not unsize Kept<closure> without a cast;
   • the parameter types were unknown (i32 against i64).

   On top of that, another session ran git pull --rebase --autostash during my build. That build silently lacked my edits, so one probe round was wasted, and it left UU and foreign files staged.

2. What I looked up that the brief could have handed me:
   • the path from #997’s look-ahead to run_code and lambda_params_by_use, and that check_core_stmt_code_parameter skips keeping callees;
   • that a kept lambda’s emitted form is Kept(Arc::new(..)) and needs a dyn Fn cast when the let has no type;
   • that row d (a run-only callee) already fails on develop.

   One probe line per row, with the unfixed output, would have saved two cycles.

3. Checks:
   • Useful: the per-row probes, the lower-std fixpoint, the guard set plus selfhosting_defects:: (43 passed), clippy, pre-push-check, and tests/compiler --both-settings (434 passed; the one failure is environmental).
   • Wasted: the probe round on the build that lost my edits to the autostash, and one ^error grep that --message-format short can hide.
   • Not possible: a stash-based unfixed run, since stash is forbidden here. Early probes on the old binary stood in for it.

4. One change. Give builder.md, or the brief, a “run the rows on the current binary and paste their outputs” step before any edit. It should also say that a rejected push only needs the supervisor’s rebase.

   Separately, the rules forbid --autostash for every session, but another session still ran it during my build. That hazard cost more than anything in the brief.

This site is open source. Improve this page.