v1.15.3
decision-records 0.2.3. A sixth review, of the fifth round's corrections, found the oldest
defect in the file and a regression introduced by the previous release.
The convention report was nondeterministic: producer | grep -q under pipefail is a race that
only bites once the output passes the pipe buffer, so every fixture was too small to see it.
Measured 5 flips in 8 runs on identical input; now 12 of 12 agree. And 1.15.2's index fix had
made the verdict depend on the ambient locale, which produced the wrong one.
The coverage sweep covered one of two families and reported none uncovered, truthfully about
the half it looked at. Both families are swept now: 18 say sites, 10 skipped sites, none
uncovered. 118 cases; against an always-zero stub, 54 passed 64 failed; no verdict changed
across 27 real collections.