fix(docs): restore the intended bytes of the AGENTS.md heredoc example

[os-zhuang]

Fixes#5150

One line. The shared-scratchpad clause in AGENTS.md §9 recommends git commit -F - with a heredoc as the mechanism half of the rule, but the line landed carrying the single-quote-inside-single-quotes shell escape, leaked in by the write that created it. Copied verbatim, the heredoc's terminator is no longer EOF, so the closing line never matches and the command hangs with no error — pushing the reader back to the temp-file fallback that the very same paragraph argues against.

Surrounding prose, the numbered ordering, and the paragraph arguing why mechanism outranks discipline are untouched, per the card. The diff is 1 insertion, 1 deletion.

Byte verification (not rendering)

A diff view escapes quotes and looks fine either way, so these are read back from the committed object at 55ab4c9.

$ git show HEAD:AGENTS.md | cat -A | grep -n "EOF"
230: git commit -F - <<'EOF'$
234: EOF$

$ git show HEAD:AGENTS.md | grep -n "commit -F -"
227: 1. **首选机制:让内容根本不落共享盘。** 提交信息用 `git commit -F -` 配 heredoc(或多个 `-m`),PR 正文直接作为工具参数传(用 `gh` 就把正文写成进程内的 heredoc,别先写文件再 `--body-file`)。内容不落盘,就无从被顶掉。
230: git commit -F - <<'EOF'
239: 两条不是并列的两个建议:前缀与删除要求每个作者每一次都记得,记性会衰减,而衰减是静默的(见上:撞车不报错);`git commit -F -` 那种形式让撞车**不可能发生**,不依赖任何人的记性。所以能用形式解决的,就别退回到纪律。这一族目前**没有钩子**兜底(上面 worktree 与 stash 两条各有一个 PreToolUse 钩子),因此这条规则的全部效力就在于你选哪种形式。

A residue scan over the committed blob returns nothing (grep exit 1). The fix was applied with an editing tool, never a nested-quoted shell heredoc — that write path is what produced the defect in the first place.

Gates

Derived from the changed path (AGENTS.md) by inspecting each script's actual scan set, not from recall. All run at 55ab4c9, the final commit:

Gate	Exit
node scripts/check-control-bytes.mjs	0
node scripts/check-doc-links.mjs	0
node scripts/check-changeset-presence.mjs	0
node scripts/check-changeset-no-major.mjs	0

Included because AGENTS.md is genuinely in scope: check-control-bytes enumerates via git ls-files (all tracked text files), and check-doc-links lists { path: 'AGENTS.md', rule: 'disk' } explicitly. Excluded with reason: check-skills-paths has SCAN_ROOT = 'skills', so a root file is outside it; the remaining check:* scripts reference AGENTS.md only in header comments, not in a scan set.

Changeset

None, and none is owed — check-changeset-presence decides this itself and says so:

Compared the working tree with 51e65d423 (merge-base with origin/main): 1 file(s) changed,
0 of them under the src/ of a package the release covers, 0 under a package changesets ignores,
0 changeset(s) added.
✅ No source of a released package changed in this range, so no changeset is owed.

Its guarded surface is < pkg >/src/** for packages in the fixed group of .changeset/config.json; a root AGENTS.md edit is outside it. Note for the reviewer: no skip-changeset label was applied, because that mechanism does not exist in this repository — content/docs/guide/ci-cd-pipeline.md records it explicitly ("None of it existed"), and the exemption form here is an empty-frontmatter changeset, which this change does not need either.

Reverse verification

Predicted before running: nothing goes red. Observed: exactly that.

Re-introducing the corrupted bytes and re-running the full derived union left all four gates at exit 0. The re-introduction was byte-identical to origin/main — independently confirmed by check-changeset-presence then reporting 0 file(s) changed against the merge-base.

Reasons, all by design rather than oversight: check-control-bytes judges control bytes only, and ' (0x27) / " (0x22) are printable ASCII; check-doc-links parses links and paths, and its own header notes this file has no markdown links at all; the changeset gates never read file content.

The generalized hole — no gate anywhere validates whether a fenced shell example in any markdown file is executable — is filed separately, unassigned, as #5151 with the measurement.

---

Generated by Claude Code