Released22 Aug 2026Commit0b7c5cacLine0.1.xLateston its line

v0.1.4

Six gates and sabotage drivers that could not fail, now able to

Notes

This release changes no product behaviour. Every change is in test/, and the reason it is worth a release is that the suite was reporting greens it had not earned. One gate was outright red on master against passing code, and four others carried assertions that could never have failed.

Changes

  • gate-sd-p1.sh was RED on master while the product was fine. Line 49 asserted a hardcoded Tests: 11 passed (11 total) while the suite declares and runs 12, so a twelfth test added at some point turned the gate red and left it accusing healthy code. The count is now derived from the suite file, and it is split into two assertions rather than one: every test the file declares must be collected, and every test collected must have passed. The first catches a test that silently stopped running; the second catches a failure the summary would bury.
  • Every setsid invocation now waits. Sixteen bare invocations against six that waited, spanning gate-d2, gate-d3, gate-sd-p1, gate-sd-p4 and the sabotage-sd-p1 / sabotage-sd-p4 drivers. The sabotage scripts were the worse case: an arm whose command does not block can read a truncated log, find no match, and report that the gate can fail when it has demonstrated nothing about the content it was checking.
  • New gate-setsid-waits.sh keeps the property from regressing. Its first version was disarmed by a single apostrophe, because a quote-parity heuristic let echo "the parser's output"; setsid ... through along with eval and sh -c strings. It was rebuilt with a positive discriminator and a control, and it now goes red on the exact states it used to pass.
  • sabotage-f1a.sh's attribution check matched the passing line. The folio runner echoes every test name, so all eight arm signatures appeared on passing lines, and "red, but not at the expected arm" was unreachable for every suite leg since that driver's first commit. The anchor is now strict, and it was exercised against all three real log shapes rather than only against synthetic controls.
  • sabotage-sd-p1.sh violated set -u on every leg, printing label: unbound variable into the committed transcript seven times and collapsing all eight per-leg logs into a single file. Fixed and verified by effect: zero unbound-variable occurrences and nine distinct per-leg logs.
  • gate-d3.sh's rg -q '7|8|' had an empty alternative, so it matched any non-empty line.

Known gaps, stated rather than left to be discovered

  • gate-sd-p4.sh:45 still hardcodes 17 the way gate-sd-p1.sh hardcoded 11. It matches its suite today, so it is green, and it is the next one to go stale.
  • sabotage-sd-p4.sh has neither a python3 guard nor a restore trap, alone among its three siblings. The missing trap is the live hazard: a driver that can leave the tree modified after a failure is a hazard to every run after it.
  • gate-d3 arm 10 caps a nested sigil-stdlib build at 30 seconds inside a verify that takes 299 seconds on an idle machine, so it goes red under load with a message indistinguishable from a real build break.
Commit

0b7c5cac

The commit this tag names.

Releases