It appears that git created files in .git/objects for the new commit, but didn’t successfully write to them. I solved it by deleting them one at a time and re-running git fsck --full to find the next one. I started with the one originally reported by git fsck:
% rm -f .git/objects/43/46883490a0990e68db0187241abc1642765a73
% git fsck --full
Checking object directories: 100% (256/256), done.
error: inflate: data stream error (incorrect header check)
error: unable to unpack 86e7247af5865e857a3b61eed99986e2d9538df1 header
error: inflate: data stream error (incorrect header check)
fatal: loose object 86e7247af5865e857a3b61eed99986e2d9538df1 (stored in .git/objects/86/e7247af5865e857a3b61eed99986e2d9538df1) is corrupt
% rm -f .git/objects/86/e7247af5865e857a3b61eed99986e2d9538df1
% git fsck --full
Checking object directories: 100% (256/256), done.
error: inflate: data stream error (incorrect header check)
error: unable to unpack a94406345ac44982b00cf57b4b9660a35436637f header
error: inflate: data stream error (incorrect header check)
fatal: loose object a94406345ac44982b00cf57b4b9660a35436637f (stored in .git/objects/a9/4406345ac44982b00cf57b4b9660a35436637f) is corrupt
And so on. I deleted five objects before git fsck came up clean, corresponding (as I suppose) to the five files in the commit I was trying to make. I guess that the file history was not corrupted at all.
Incidentally, I thought of another method that seems to work as well. git clone copies the bad objects, but git push does not. After backing up, I created a new empty repository (–bare, because otherwise you can’t push to master), then unstaged my changes and pushed both branches into the new repository. Then it was just a matter of checking it out again and restoring the latest changes from my backups.
Still interested if anyone cares to shed light on the failure mechanism here.