family.py add/bump/remove: a members file deleted unstaged, a git bisect, an empty sparse-checkout file, an empty rebase-merge/ and a merge conflicted in family.yaml meet a refusal or remedy that does not fit (#626's gaps) #684

Goal: where the holder is mid-operation or sparse in a state #626's checks do not model, the verbs refuse before their first write with a remedy that works, or say nothing wrong.

Filed by lane openRepoShape-5 on Brett Heap's word of 2026-10-16T12:17:02Z ("#459's family lines (~38)"), from lane 4's R612 lines on #459 (comment 6213530456). Lane 4 judged each OPEN on 95780dd (cluster T7); re-read on d4a7fab.

Lines (line numbers at d4a7fab; probes were run on a scratch export of main)

• R612 E1, a tracked members FILE gone from disk, deletion unstaged, or moved by family.py: add, bump and remove refuse before their first write a merge, cherry-pick, revert, rebase or git am in progress, a .gitmodules, members/ or family.yaml the sparse checkout leaves out, and a file named members, which each failed after writing and was rolled back, or committed in the wrong place (#612) #626's own git mv remedy and not committed, passes every add check. _checked_mount_path asks _members_is_a_file only where the disk lists members as a file (scripts/family.py:3196, the function at :3081), so add's commit carries D members beside A members/<P> and leaves the rename's other half staged (R612 measured; not re-run, since it needs a holder that tracks members as a file). Direction: refuse an index entry members of mode 100644 as well (_index_entries(root, MEMBERS_DIR) is already asked at :3092), in the words of the on-disk case, with its git mv then commit remedy.
• R612 E2, git bisect in progress: family-holder-detached's git switch -c would branch from the commit under test, and the exit is git bisect reset. OPERATION_MARKS (scripts/family.py:827-852) has no BISECT_LOG, and _checked_branch (:744-770) refuses the detached HEAD with the switch -c remedy. Probed: git bisect start; bad; good <root>, then family.py add prints REFUSED family-holder-detached ... switch -c family/add-irqs. The dry run prints a plan. Direction: ask BISECT_LOG with the other marks (where git keeps it, _git_paths), name git bisect reset as the exit and say the dry run refuses the same.
• R612 E3, an EMPTY .git/info/sparse-checkout with sparse checkout on refuses every path, and the first command the remedy names takes every other file out of the work tree. Probed: git sparse-checkout init --no-cone, the file truncated, family.py add --dry-run gives REFUSED family-sparse-excluded ... leaves out .gitmodules, members/IRQS, family.yaml and git sparse-checkout add .gitmodules members/IRQS family.yaml, or git sparse-checkout disable (scripts/family.py:1173-1177). R612 measured the top level going from 13 entries to 4 after the add. Direction: where the rules file is empty, name git sparse-checkout disable alone.
• R612 E4, reftable (git 2.45+): CHERRY_PICK_HEAD and REVERT_HEAD may live in the ref store, not as files. _operation_in_progress probes the marks with os.path.lexists (scripts/family.py:942), so a cherry-pick would be a late failure and a revert the harm family.py: add, bump and remove refuse before their first write a merge, cherry-pick, revert, rebase or git am in progress, a .gitmodules, members/ or family.yaml the sparse checkout leaves out, and a file named members, which each failed after writing and was rolled back, or committed in the wrong place (#612) #626 refuses. NOT MEASURED: this host has git 2.43. Direction: verify first, on git 2.45 or later (git init --ref-format=reftable, stop a cherry-pick and a revert, list .git); if they are files, close this as not a defect, and if not, ask the ref store (git rev-parse -q --verify CHERRY_PICK_HEAD, which also resolves a branch of that name, so qualify it).
• R612 E5, a stale, EMPTY rebase-merge/ is refused as a rebase, and the remedy names only --abort, which fails there. Probed: mkdir .git/rebase-merge; family.py add --dry-run gives REFUSED family-operation-in-progress ... finish it with git rebase --continue ... or give it up with git rebase --abort (scripts/family.py:828-829, the text at :1025-1029); git rebase --abort there prints could not read '.git/rebase-merge/head-name' and exits 1; git rebase --quit exits 0 and removes the directory. Direction: where rebase-merge/ holds no head-name, name git rebase --quit.
• R612 N4, a merge conflicted in family.yaml itself is never named as a merge. Family.open parses the manifest first (scripts/family.py:279), so the commonest holder merge conflict says only REFUSED yaml-unparsable: .../family.yaml: line 108: merge keys are not supported (probed; scripts/repo_shape.py:789), with no remediation, before _checked_no_operation_in_progress (:953) is asked. Asking the check inside Family.open broke tests/test_family_member_path.py::test_the_refusal_is_made_before_any_git_call[remove|bump], which forbids any git call before the member-path refusal. Direction: ask the operation check only when the parse fails (W612's fallback), and let the yaml-unparsable refusal carry a remediation naming the merge.

Overlaps: none with an open PR: #639 edits family.py's _checked_gitmodules_as_staged, a new _checked_gitmodules_held and cmd_remove (other functions), and lane 5's #646 edits _spellings_refusal (:2917). Neither file this issue edits is a pinned copy.

Left out, by the lines' own words: R612 E7 (git merge --squash and git cherry-pick -n leave staged changes with no mark) is benign and stays on #459; R612's cosmetic raw path in the family-operation-in-progress line is in the path-spelling issue.

Mark the six lines → this issue on #459, and tick them there when this closes.

🤖 Generated with Claude Code