lower-std, built, probed, then put it back, ran lower-std and built again. After that came the long tests/compiler --both-settings run, and sorting out its 11 failures, which all came from other people’s uncommitted work (#999 text_tiers.nika, cargo file-lock noise in one test’s output). Two small slips cost a turn each:
S=... && (clippy) & chain, so $S was empty afterwards. builder-facts.md already warns about exactly this.sed command broke because # was both the delimiter and in the text.What I had to look up: whether to adopt the fix and test an earlier run of this brief had left uncommitted. The brief’s “every file modified at your start is not yours” rule, read literally, forbade it. A line like “an earlier run may have left your own change in the tree: check the diff and adopt it” would have settled it. I also had to check whether a binary without the fix existed (there was no target/release). Locations, the cause and the test text were all in the brief.
nikaia lower check that the clone comes out, the unfixed rebuild, the lower-std fixpoint, the guard set, clippy and pre-push-check.tests/compiler --both-settings. Its red came from a shared dirty tree, so it could not judge my change. A run over the NK1131 tests alone (where.nika -- NK1131) would have answered the question in a fraction of the time.target/fast/nikaia.prev from the last build), so the unfixed run costs no rebuild. Second: in builder.md, say that when tests/compiler failures sit in files of another agent’s uncommitted work, the builder escalates them under Needs: and does not treat them as its own gate red.