The suite wrote its own verdict to disk for a week. The CI runner never opened the file.

The interesting defect is not the abort. It is the default the classifier picked when the counts were missing. Crash with no counts was read as a process that died partway through its assertions, so it blocked the commit chain; crash with green counts was reported and waved through ^[8]. That is a reasonable default when stdout is a faithful record of how far a process got. It stops being reasonable when the same abort that kills the process also decides whether the last line ever left the buffer ^[9]. At that point both branches are conclusions about flush ordering, and the refusal to let changes land ^[1] is no longer a statement about the code.

The visibility is lopsided in the wrong direction. The blocking branch is loud, printing "do NOT commit" ^[9]; the continuing branch is quiet. So the week of "it keeps failing" ^[10] was spent on the arm of the coin flip that shouts, while the arm that lets the chain proceed on a dead process makes no noise at all.

The measurement that preceded all this deserves more attention than the crash it failed to fix. Four userland variants, interleaved A/B, six runs each inside one window so machine drift could not flatter the result ^[4]. The targeted fix crashed in 3 of 6 runs; leaving the code alone crashed in 0 of 6 in that same window ^[5], which is a 50 point crash rate for the fix against zero for doing nothing ^[1]. The identical untouched code had crashed 8 of 8 twenty minutes earlier ^[5], a swing across the entire range with no code change at all ^[2]. The author's read, that the environment dominated anything testable from userland ^[6], is what the numbers say.

Then the mechanics of the channel that actually worked. The battery had been writing its pass and fail counts synchronously to a JSON file in the temp directory, along with a note that the exit code may come from a libuv teardown race after all assertions complete, and to trust the file over the exit status ^[11]. A synchronous write returns before the abort; a buffered console line does not ^[11][9]. That file was created during the original crash investigation a week before the verdict mystery began, and the runner never read it ^[12].

The repair is a contract rather than a heuristic: the runner hands each battery a summary path through the environment, the battery writes its counts there before exiting, the file is authoritative, and stdout is demoted to the fallback for batteries that do not have one ^[13]. The general form is worth keeping. An exit status is a summary produced by the process being summarised, at the moment it is least capable of producing one. Anything that has to survive an abort belongs on disk, written by a call that has already returned.

The harness's stated premise is that a verdict you cannot trust is worse than no verdict ^[14]. For a week it issued exactly that, and the correction was one file read away.