adopt-project.py: plan writes its follow-ups once, so after a hand edit of the plan check and the split commit name a GNUmakefile the root no longer keeps, or say nothing of one it now keeps #1206

(lane openRepoShape-5, split out of #1132 as its item 5; RV1105's NIT 8 on #1105)

Problem

plan works out the follow-ups once and stores them in the plan's follow_ups: list: the CONTRACTS_DIR line (_makefile_follow_up, #977), the collision lines (_predict_collisions, whose one caller is plan), and the rest of follow_ups_for. check prints the stored items, and execute carries them into the split commit's message, which is also the pull request's body.

The plan is meant to be edited by hand:

• contracts/path-classification.yaml (root-project-makefile) says that moving a Makefile to the code leg "is a one-line edit of the plan";
• every review_required question is answered by editing leg:.

After either edit, the stored follow-ups still describe the plan as plan wrote it, not as it stands. execute's NEXT does not have this problem, because it reads the plan's current root paths (paths_for["root"]).

Measured on main 961640c68

1. A hand move of a kept GNUmakefile to the code leg. This used an untracked probe in the lane's scratch folder: test_zz_probe_n1132_stale.py, built on tests/test_adopt_contracts_dir.py's TREE and EDITS and run through the lane's runner. There were two sources: one keeps GNUmakefile, and one keeps Makefile + GNUmakefile. For each:

• run plan;
• replace - path: GNUmakefile\n leg: root\n with leg: code and a resolution:;
• run check, then execute --local-remote-dir.

2. A question answered root. This is row answered-root of tests/test_adopt_make_bootstrap_followups.py, on #1132's branch. The source holds GNUmakefile/rules.mk.

• plan asks about GNUmakefile/ (rule default, leg: null). It writes the root Makefile wires CONTRACTS_DIRto thecontracts/ of the spec leg, ..., which says nothing of GNUmakefile/, because the plan sent nothing GNU make reads to the root.
• The question is answered leg: root.
• check prints that same line, and the split commit's message carries it.
• NEXT, which reads the plan as it stands, says that GNU make stops at GNUmakefile/.
• In the clone, make -n bootstrap exits 2 with make: *** GNUmakefile: Is a directory. Stop..

So the plan's CONTRACTS_DIR line says the root's Makefile wires the variable, at a root where make reads no Makefile at all.

The collision lines come from the same stored list and would go stale the same way, for example after a hand move of a kept Makefile. That case was not measured.

Recommended fix

Have check and execute work out the follow-ups from the plan as it stands, and stop storing them, as #600 did for the read-across lines. Since #600, cross_leg_references takes no mount path, and contracts_from_code_leg prints the path from the code leg to the spec leg beside the CURRENT mounts. Work out the CONTRACTS_DIR line and the collision lines the same way, from the current entries:

• follow_ups_for over the plan's entries;
• _predict_collisions over its root paths, with attributes_link read as execute reads it.

A plan written before the change still holds a follow_ups: list. Read it as nothing, or refuse it by name; never print it beside the recomputed list.

The second-best fix is for check to compare the stored list with a recomputed one, and to report each item that differs. That still leaves two lists to keep in agreement.

Related cases found while porting #1132 (main, or #1132's branch, as marked)

1. A kept makefile/ DIRECTORY gets "until the two are merged". _makefile_follow_up asks _root_keeps(surviving, "Makefile", fold=True), which takes makefile/ for Makefile on a disk that ignores case. So, rendered on adopt-project.py and the assembly-root template: follow-ups from RV1105's and RV1121's reviews of #1105 and #1121 (make bootstrap in the template, a kept GNUmakefile/ directory, python3 on Windows, GNU_MAKE_READS twice, stale follow-ups) #1132's branch for a root that keeps only makefile/, the line says the shape's Makefile is written "as shape/Makefile once case is ignored, as a macOS or Windows disk does, where the source's makefile is that path, until the two are merged". A directory cannot be merged with a Makefile. (adopt-project.py and the assembly-root template: follow-ups from RV1105's and RV1121's reviews of #1105 and #1121 (make bootstrap in the template, a kept GNUmakefile/ directory, python3 on Windows, GNU_MAKE_READS twice, stale follow-ups) #1132 adds the clause that says GNU make stops at makefile/, and leaves this part as main has it.)

2. A directory kept only through deeper entries is not seen. _root_keeps(surviving, name) asks only for name and for name/. Measured on adopt-project.py and the assembly-root template: follow-ups from RV1105's and RV1121's reviews of #1105 and #1121 (make bootstrap in the template, a kept GNUmakefile/ directory, python3 on Windows, GNU_MAKE_READS twice, stale follow-ups) #1132's branch:

   • _root_keeps({"GNUmakefile/rules.mk"}, "GNUmakefile") is False;
   • _make_reads_first of that root is None;
   • the CONTRACTS_DIR line is then "the root Makefile wires ...".

   But GNU make finds the directory GNUmakefile/ all the same and stops there. plan folds such a folder into one GNUmakefile/ entry, so this needs a hand-split entry or a policy that writes file entries.

3. A kept Makefile/ DIRECTORY is treated as a Makefile of the source's own. _root_keeps({"Makefile/"}, "Makefile") is True. So the shape's Makefile goes beside it as shape/Makefile, and the line says "the source keeps a Makefile of its own at the root, so that block waits in shape/Makefile until the two are merged". GNU make 4.3 at such a root exits 2 with make: *** Makefile: Is a directory. Stop.. Measured in a scratch folder that held only a Makefile/ directory.

Verify

• A row per edit: move a kept GNUmakefile to code, and answer a GNUmakefile/ question root. After each edit, check's follow-ups and the split commit's message must match a fresh plan of a source whose rules give the edited answer directly. That source is tests/test_adopt_make_bootstrap_followups.py's a-rule-keeps-it row for the directory.
• A row for a stored follow_ups: list in an old plan.
• Rows for the three related cases, or a named ruling that leaves them as they are.

Lane: openreposhape-5 (openRepoShape-5)