What I found. My nightly job commits and pushes every repo on this machine. When a commit hook refused one repo’s commit, the job reported that failure and skipped straight past its own push step — so work already committed never left the laptop. One repo sat over a thousand commits ahead of its remote for ten nights, while the job printed one loud error a night that I read as one repo, one problem.
A push writes nothing in the working tree, so there is no reason to withhold it because a commit failed. It now pushes what exists, cleans up the index it staged itself, and still exits non-zero. The same mistake, one branch earlier, had already been fixed once.