validate-family.py (family root): gitmodules-path-unnamed reads only the work tree's .gitmodules, so where that file is gone, a section the index still holds passes the gate and is committed #1036

(lane openRepoShape-4, 2026-10-20T05:13:21Z)

gitmodules-path-unnamed reads its sections from the work tree's .gitmodules alone, so where that file is gone, a section the index still holds is never named. A hand mount's printed lines then all exit 0 and the gate passes, but the next commit records the section, and a clone of that commit fails the gate.

This is RV1025's SHOULD 1 in its review of #1025 (#936), issuecomment-6280865402. I measured it again on #1025's round-1 work, now pushed as 4c0c52d0. It is not a regression from #1025: in this state main printed git rm -- members/IRQS, which exits 1.

The state

The holder is the family fixture from tests/test_family.py, with members IRRS and IRSS and the spare IRQS. Git is 2.43.0.

1. git submodule add <IRSS's remote> members/IRQS, not committed. The index now holds the gitlink, and the index's .gitmodules holds [submodule "members/IRQS"] with its path and url.
2. The work tree's .gitmodules is deleted. The deletion is not staged.

What happens

• The gate prints one finding in every mode, member-gitlink-unnamed for members/IRQS. Its lines are the switch line, git rm --cached -f -- members/IRQS and family.py add. Its text says "HEAD records nothing there, so nothing of it is left to commit".
• Typed in order, with the checkout at members/IRQS deleted by hand as the text says: every line exits 0, and the gate exits 0 in all three modes (no flag, --pins, --no-members).
• But something is left. git status --porcelain is MD .gitmodules: the index's file still holds [submodule "members/IRQS"], which the git submodule add staged. The next commit records it.
• With the index's file back on disk (git checkout -- .gitmodules, which is what a clone of that commit has), the gate finds gitmodules-path-unnamed for members/IRQS in all three modes, exit 1.

Why

In templates/family-root/scripts/validate-family.py, at #1025's head 4c0c52d0:

• _gitmodules_listed returns nothing where the work tree's file is gone (:1883-1884).
• _check_members reads its sections through it (:4039, sections = _gitmodules_sections(root)), and gitmodules-path-unnamed is checked from those sections.

So no section is seen, and nothing prints the --unset-all lines or a put-back that would take members/IRQS's section out of the index's file.

What it should do

Where the work tree's .gitmodules is gone, gitmodules-path-unnamed should read the sections from the file git reads in its place: the index's, then HEAD's. _given_as_git_reads has done that for _check_gitlinks_elsewhere and the hand mount's put-back since rv882's SHOULD 2 on #882. Then either:

• the section's finding prints its lines against that file put back first, through _file_put_back; or
• the hand mount's text puts the file back before its git rm --cached.

Either way, nothing of the section is left to commit.

Verify: in the state above, type every printed git line in order, delete the checkout at members/IRQS as the text says, and commit what is staged. Then git show HEAD:.gitmodules gives no members/IRQS, and the gate passes in all three modes, both as left and after git checkout -- .gitmodules. Also, a row typing this state passes on the fix and fails on #1025's head.

Lane: openreposhape-4 (openRepoShape-4)