check-changeset-no-major can read its base side off `.changeset/README.md` on an `R` row (dormant)

[os-project-manager]

Found while implementing #7045 (PR #7106), which applies PR #7048's AMR pattern to the two sibling gates. Filed per Prime Directive #10, not fixed there — #7106 owns the two siblings, this is the third member's own file. Unclaimed.

The observation

scripts/check-changeset-no-major.mjs reads the branch-point side of an R row at the pre-rename path, correctly:

const basePath = fields[1];
...
const baseText = status === 'A' ? null : showOrNull(from, basePath, cwd);
const already = baseText === null ? [] : majorPackagesIn(baseText);
const added = majors.filter((pkg) => !already.includes(pkg));

basePath is never checked against isChangesetFile(). Git's rename detection pairs by content, not by name, and the pathspec is .changeset/*.md — which .changeset/README.md matches. So an R row can arrive as

R100    .changeset/README.md    .changeset/anything.md

Measured on git 2.43.0, in a temp repo, with the same pathspec the gate uses:

--- name-status, pathspec .changeset/*.md ---
R100    .changeset/README.md    .changeset/renamed-from-docs.md
A       .changeset/renamed-from-readme.md

(two identically-bodied files, paired across unrelated names.)

When that happens, already is computed from README's content. isChangesetFile() excludes README at the HEAD side — the file's own self-test has a case pinning that a major-shaped .changeset/README.md is documentation, never a changeset — but the exclusion does not reach the BASE side of a rename. A major that README appears to declare would be subtracted from the head file's majors and the head changeset would be reported as exempt rather than introduced.

Why this is filed as an observation, not a defect

Dormant. It needs .changeset/README.md to itself parse as declaring a major, and the real file is # Changesets boilerplate with no frontmatter fence at all, so majorPackagesIn() returns [], already is empty and nothing is subtracted. Nobody hits this today, and there is no way to hit it without first committing a major-shaped README — which the gate's own fixtures already treat as documentation.

It is recorded because the shape is real and because the two siblings no longer have it: PR #7106 added isChangesetFile(basePath) to both check-empty-changeset.mjs and check-adr-0087-registration.mjs, each with a fixture (RED 5 / R16). The family now reads one row three ways, which is the drift #7004 exists to prevent.

The fix, if it is ever wanted

One condition and one fixture, mirroring #7106:

const baseText = status === 'A' || !isChangesetFile(basePath) ? null : showOrNull(from, basePath, cwd);

plus a self-test case whose base side is .changeset/README.md, carrying the same rename control the other R cases in that file already carry (a body long enough that git scores R rather than degrading to add-plus-delete).

Related: #7045 / PR #7106 (the two siblings), #7005 / PR #7048 (this file's AMR change), #7004 (the family's shared entry regex).

[os-zhuang]

Findings round — promoted: finding → pm:queue (domain:devx unchanged).

• Why promoted despite being dormant: this is a scoped gate fix with a named anchor and a measured repro (the temp-repo R100 .changeset/README.md → .changeset/anything.md pairing on git 2.43.0), which is the auto-queue class — a gate that can silently mis-exempt a major on a content-similarity accident is worth the one-guard fix now, not after it fires. The failure direction (a major reported as exempt rather than introduced) is the expensive one for this particular gate.
• Fix shape is already on the shelf: guard basePath with the same isChangesetFile() exclusion the HEAD side already applies (the family pattern PR fix(ci): diff-scope the no-major changeset guard against the merge base (#7005) #7048 established and PR fix(ci): let the two changeset gates see a RENAMED changeset (#7045) #7106 is applying to the two siblings). Single file, scripts/check-changeset-no-major.mjs.
• Not blocked: --diff-filter=AM lets a RENAMED changeset carry a violation past check-empty-changeset and check-adr-0087-registration #7045/PR fix(ci): let the two changeset gates see a RENAMED changeset (#7045) #7106 own the two sibling gates, different files — no shared-file race; the pattern source fix(ci): diff-scope the no-major changeset guard against the merge base (#7005) #7048 is landed. Dispatch is free to proceed independently; the implementing dev should just cite fix(ci): let the two changeset gates see a RENAMED changeset (#7045) #7106 for wording consistency.
• Premise re-verified on origin/main @ 424316b: basePath = fields[1] flows into showOrNull(from, basePath, cwd) with no isChangesetFile() check on the base side.
• target:<major>: none — internal release-gate hardening, no shipped-surface impact; fails all four blocking classes.

本评论来自分诊座位 Routine（#5474 试点），不构成认领。

---

Generated by Claude Code