Most cost: the one tests/compiler --both-settings run, which was long, plus one build cycle. It ended with 7 failures that I could not cheaply isolate. Another agent’s uncommitted text_tiers.nika/.rs was in the shared tree and so in my binary. I put the failures down to that file but never proved it. That left the gate verdict shaky, and that was the real cost. The code itself took few turns, because the brief’s Locations were accurate.
for builds its frame (lent, CheckLocal.id, bound = view_of_type(..), so r’s type is a view and needs .unseen());check_core_stmt_for and the let;number_list_name and open_items_of, which only accept a Variable;check_flow.nika:~405, and no Rust one.The brief said 5 CheckOpenNumber { literals; there are 3. I also guessed that the bound type of [[1,2],[3]] is Vec[Vec[?]]; I never looked it up, but the probes confirmed it.
lower-std, the build, then lower-std again;fmt on the test;it modules plus the guard set, with clippy beside them;pre-push-check, run twice: the first run piped through tail showed nothing, so a second was needed to see the exit code (wasted);tests/compiler at both settings.Also wasted: one probe written with i16, which Nikaia does not have.
git status shows another agent’s uncommitted tools/*.nika or lowered/*.rs at the start. That file gets compiled into your binary and makes the tests/compiler verdict impossible to attribute. Either the brief lists the failures already known on that tree, or builder.md gives a cheap way to check one: rerun the failing test with a binary that does not contain the foreign file.