docs(agents): criterion ③'s test is scoped to the branch's own commits — the unranged git log walked all of main and made force-with-lease dead letter (#18351)

Fixes #18178

## The defect

`AGENTS.md` Multi-agent §3 allows `--force-with-lease` only while all
five criteria hold.
Criterion ③ stated the bar — *nobody else has ever pushed it* — and then
operationalised it
as the author and committer sets of an **unranged** `git log` over the
branch's remote ref.
With no range that walks the branch's entire ancestry, all of `main`
included, so the sets
are main's whole history plus the agent and the test cannot pass for
**any** branch cut from
`main`. The stated criterion is satisfiable and right; only its
operationalisation was
impossible — a reader who follows the words never force-pushes, one who
follows the intent
does, and that divergence is the bug. The card was filed after the dev
on #17330 declined to
take a reading it could not prove and shipped four `wip:` commits plus a
merge instead.

## The change

One hunk, two lines, `AGENTS.md` :473–:474. The parenthetical is
re-keyed to the branch's
**own** commits — the range form `origin/main..` in front of the branch
ref. Criteria ①②④⑤
are untouched, ③'s bar is untouched (same words, same strictness), and
the file stays at its
1075-line ceiling with no new line.

```diff
-   `claude/issue-*`; ② this worktree created it; ③ nobody else has ever pushed it (the
-   author and committer sets of `git log origin/BRANCH` are you alone); ④ no open PR
+   `claude/issue-*`; ② this worktree created it; ③ nobody else has ever pushed it (the author and
+   committer sets of its own commits, `git log origin/main..origin/BRANCH`, are you alone); ④ no open PR
```

⚠️ `BRANCH` above stands for the angle-bracket placeholder the file
actually spells; this
body writes the word instead, because a GitHub body sanitizer eats short
angle-bracket
fragments. The file itself is unchanged in that respect — read the diff
for the real bytes.

## Ratchet and byte budget

| reading | before | after |
|:--|--:|--:|
| `wc -l AGENTS.md` (ceiling 1075) | 1075 | 1075 |
| line :473 | 90 B | 101 B |
| line :474 | 88 B | 108 B |
| lines :472 / :475 / :476 / :477 | untouched | untouched |
| `git diff --stat` | — | 1 file, 2 insertions, 2 deletions |

`scanLineLengths` (the gate's own scanner, budget 120 B) reports **0
offenders** on the
edited file.

## Reverse verification

Taken on this very branch, a real `claude/issue-*` branch, and on a
constructed control.
⚠️ This checkout is **shallow**, so the unranged sets are truncated
relative to the card's
full-history reading (33 / 32 / 1); the *shape* is what reproduces, and
it reproduces exactly.

```
BEFORE — today's literal test, on this branch
  git log --format='%an|%cn' origin/claude/issue-18178-force-with-lease-criterion-scope | sort -u
  -> 7 pairs   (Claude|Claude, Jack Zhuang|GitHub, Leehom|GitHub, Warren Buffett|GitHub,
                claude[bot]|GitHub, os-elon-musk|GitHub, os-try-charles|GitHub)

CONTROL — the same literal test on origin/main itself
  -> 6 pairs   — the branch's 7 minus Claude|Claude, i.e. main's set plus the agent, exactly
                 as the card measured. `comm -23` of the two sets prints one line: Claude|Claude.

AFTER — the re-keyed test, same branch
  git log --format='%an|%cn' origin/main..origin/claude/issue-18178-... | sort -u
  -> 1 pair    Claude|Claude          (2 commits in range)

FIRING CONTROL — the re-keyed test on a branch that genuinely has a second pusher
  (a throwaway local repo, two clones, two identities; no probe branch was pushed to origin)
  unshared branch, literal test   -> 4 pairs   (wrongly refuses)
  unshared branch, re-keyed test  -> 1 pair    (permits — correct)
  SHARED branch,   re-keyed test  -> 2 pairs   Claude|Claude + os-try-charles|os-try-charles
                                               (still refuses — the bar is not loosened)
```

## The one design choice: which scoped spelling

The card proposed `$(git merge-base origin/main
origin/BRANCH)..origin/BRANCH` and called it a
proposal, not a prescription. This PR ships the plain two-dot range
`origin/main..origin/BRANCH`.
Both compute the same commit set — verified on the live branch and on
both control branches,
1 / 1 / 2 pairs either way — and the two-dot form is one moving part
fewer. The deciding
measurement is how each **fails**:

```
# the substitution form, when merge-base returns nothing (shallow clone, unrelated
# histories, a ref not fetched) — the range silently becomes HEAD..origin/BRANCH
  exit=0   pairs=0   on a branch a second agent had just pushed  -> "you alone" is
                     VACUOUSLY true, and the criterion permits exactly the clobber it forbids

# the two-dot form under the same fault
  exit=128 fatal: ambiguous argument ... unknown revision  -> loud, no verdict at all
```

That failure is live in this repository, not hypothetical: `git
merge-base origin/main REF`
exits 1 with empty output here for a branch whose base lies outside the
shallow window.
Preferring the spelling that cannot degrade into a different question is
the same principle
the card is about — a test that answers confidently and wrongly is worse
than one that stops.

## 维护者速读(草稿)

**改了什么** — `AGENTS.md` 多 agent 纪律 §3 里判断「分支是不是只有我一个人推过」的那条命令,
从 `git log` 加分支名(走遍整条祖先,包含整个 `main`),改成只看分支自己的提交(`origin/main..` 区间)。
正文两行,判据本身一字未动。

**为什么改** — 原命令对任何从 `main` 切出来的分支都必然返回「多人」,所以第五条允许的
`--force-with-lease` 实际上永远用不了:条文写着允许,测出来永远禁止。按字面读的 agent 从不 force-push,
按意图读的会,两种读法分叉本身就是缺陷。已经有一张卡的开发因此多推了四个 `wip:` 提交和一个 merge。

**风险与代价(含回滚)** — 改的是规则文本,不是代码,没有运行时影响。风险是「改松了判据」,
本 PR 用一个构造出来的、确有第二个推送者的分支做了发火对照:新命令仍然读出两组身份、仍然拒绝。
回滚成本 = 还原两行文字。

**席位意见** — (留空,待席位定稿)

**你要做的** — 这是受管面(`AGENTS.md`),需要您的一次授权批准;批准前没有任何 agent 席位会合并、
入队或挂 auto-merge,PR 保持 draft。

## Acceptance notes

- noted, not filed: the ratchet's line-length verdict prints **only on
failure**
(`check-skill-line-ratchet.mjs` `run()` logs `lv.msg` under `if
(!lv.ok)`), so a green run
says nothing about the 120-byte axis for any file. Not a defect class —
a silent green is
the normal shape for that gate — measured positively here by calling
`scanLineLengths`
directly instead of reading the gate's silence. Bearer: none; noted for
the next author who
  cites that gate's green as width evidence.
- Scope held: `AGENTS.md` only, criterion ③'s parenthetical only; :472
and :475–:477 were not
  re-flowed and not touched.

## Landing

Governed surface (`AGENTS.md`, Prime Directive #14) —
`check-governed-merges.mjs --test`
answers **GOVERNED**, 1 of 1 path. The PR stays **draft**; no reviewer
was requested, no
auto-merge armed, nothing flipped. It waits for the maintainer's
authorized approval, which
the owning seat requests.

---
_Generated by [Claude
Code](https://claude.ai/code/session_01HZfg2AwVX191qCizp88gQr)_

Co-authored-by: Claude <noreply@anthropic.com>

Author: os-zhuang

AGENTS.md

@@ -470,8 +470,8 @@ Even inside your own worktree, operate defensively:
 3. **Never force-push a *shared* branch, and never push `main`.** A force-push can
    clobber a parallel agent's work; `main` is shared — land all via PR. A branch is
    unshared, and `--force-with-lease` allowed, only while ALL FIVE hold: ① it is named
-   `claude/issue-*`; ② this worktree created it; ③ nobody else has ever pushed it (the
-   author and committer sets of `git log origin/<branch>` are you alone); ④ no open PR
+   `claude/issue-*`; ② this worktree created it; ③ nobody else has ever pushed it (the author and
+   committer sets of its own commits, `git log origin/main..origin/<branch>`, are you alone); ④ no open PR
    on it carries a reviewer or an approval (one does ⇒ a new branch and a fresh PR
    instead); ⑤ the push spells `--force-with-lease=<branch>:<sha you last pushed>` —
    ⛔ never bare `--force`. One criterion failing ⇒ the branch is shared.