fix(pm): git-history's touch leg reads the graft, not just the parent's presence - #18398

Fixes #18355

touchIsProvable proved leg 1 by asking whether every parent the commit OBJECT names is present locally. A later fetch — of another ref, or of one sha — brings that parent into the object store without moving the shallow boundary, so the graft stays registered and git's traversal goes on treating the commit as parentless. Leg 1 then passed, diff-tree diffed the commit against its EMPTY grafted parent list and printed nothing, and the verdict came back:

{ provable: false, boundary: false, reason: "SHA has its parent(s) locally but its diff does not touch PATH" }

The refusal is correct (no caller is handed the sha) and both the boundary flag and the sentence beside it were wrong: an operator was told the file was simply untouched, when a grafted walk is what answered. Which of the two sentences came out depended on when the last fetch ran.

The discriminator, and why it is not the shallow file

Leg 1 now reads the graft where git applies it, per sha. git rev-parse SHA^@ lists the parents the TRAVERSAL uses — the --max-parents=0 reading asked of ONE commit — and an object naming a parent that list does not carry is grafted, whether or not that parent is in the store. New export isGraftBoundary(cwd, sha, parents?) compares the two lists; touchIsProvable fails leg 1 on it and returns boundary: true with a reason that names the outlived registration.

The card named two candidate discriminators (shallow-file membership, or rev-list --max-parents=0). Both were measured; SHA^@ is the same traversal fact, asked per sha:

A real root still passes leg 1 (its object names no parent, so the two lists agree) and an ordinary commit still answers — the case a --max-parents=0 count on its own cannot tell apart, which is why this compares two lists instead of counting one.

No behaviour change beyond the flag and the sentence

The refusal itself, the --deepen doubling to --unshallow, exit 2 and the empty stdout are untouched. A grafted commit's diff-tree prints nothing in every case, so no reading that used to be provable: true can become a refusal here.

Tests

• node scripts/pm/git-history.mjs --self-test — 62 cases, exit 0 (was 54). Eight new cases in the touch battery, floor 17 to 25.
• node scripts/pm/check-harness-current.mjs --self-test — the consumer, 14 cases, exit 0, unedited. It prints boundary where it used to print unprovable for this state, which is the point of the card.
• Fixture, constructed not described: clone --depth=5, then a plain git fetch origin PARENT_SHA (no --deepen, so nothing rewrites the shallow list) with uploadpack.allowAnySHA1InWant on the upstream. The case READS THE STATE FIRST — shallow list still names the boundary, parent object present, walk still roots there, clone still shallow — because a case asserting on a state that did not form proves nothing.
• Ablation (leg reverted to parent-presence only, on the committed tree, restored with git checkout HEAD -- under a trap, blob hash compared both ways): 3 cases red, then byte-identical restore and green again.

PRE : hash-object=1239050079c6d4507a65e8cf3af16f3f2cdfc5d7 == HEAD:scripts/pm/git-history.mjs
POST: removed-marker=0 removed-leg=0 injected=1 hash=188314847512759d24fa11310b86c5d3d4bdb815
ABLATED SELF-TEST EXIT=1 · git-history --self-test: 3 FAILED.
RESTORED: hash=1239050079c6d4507a65e8cf3af16f3f2cdfc5d7 (HEAD=1239050079c6d4507a65e8cf3af16f3f2cdfc5d7) git-diff-HEAD-empty=yes
RESTORED SELF-TEST EXIT=0 · all cases passed.

Acceptance notes

• The first spelling of the CLI-refusal case tested stderr for graft boundary and stayed GREEN under ablation: the refusal boilerplate prints that phrase on every refusal, so the case was pinning a constant. The ablation is what found it; the case now pins the reason clause and the absence of the misleading sentence, and its own history is written into the case name.
• One wording-only second commit: a neighbouring case title still called the predicate "the parent's presence". The assertion is unchanged.
• skip-changeset: scripts/pm/** is not in any package's files[] — nothing publishes.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Bz6hxDBqK62NP2W1LATvnt

Generated by Claude Code