docs(pm-dispatch): a decision analysis recommends exactly one option (#24520)

Fixes #24444

Clause-②: no

## What

Two lines of
`.claude/skills/pm-dispatch/references/decision-analysis.md`, nothing
else (1 file, +2 / −2):

- **line 9** (the director's pre-batch increment): 「推荐/回退/口径」 → 「推荐/口径」,
the card's wording verbatim.
- **line 26** (the recommendation item of the card-face template): 「推荐 +
回退 + 置信缺口:荐一项、给回退项、明说看不见什么;⛔ 不沿用他席推荐。」 → 「推荐 + 置信缺口:只荐一,余各一句不取理由;盲区明说,⛔
不因此另荐;⛔ 不沿用他席推荐。」

The four facets, the 自检 line and the 裁后执行 paragraph are untouched;
`SKILL.md` is untouched.

## Why

The maintainer, 2026-10-19: 「你的决裁总是有一个回退合理吗？是完全不同的开发方向。最终结论应该只有一个吧？是
skills 的问题吗」, then 「同意」 on the card's change. Ruling of record: director
ledger comment `6259199455` (on #12708).

## Route A: compressed in place, line count unchanged

The card's line-26 text measures 150 bytes; the ratchet's
`MAX_LINE_BYTES` is 120 (`scripts/pm/check-skill-line-ratchet.mjs`), and
the file sits at 50 lines against its 50-line ceiling (headroom 0). So
the ruled wording cannot land verbatim on one line. Route A (compress
the wording, keep every clause, add no line) loses no clause, so route B
(fold the verbatim text into two lines and raise the ceiling 50 → 51
under the ruling) was not taken and the `CEILINGS` row is unchanged.

Clause survival, ruled text → landed fragment:

| Ruled clause (card text) | Landed fragment | Change |
|---|---|---|
| item name 推荐 + 置信缺口 | 推荐 + 置信缺口 | verbatim |
| ① 只荐一项 | 只荐一 | measure word 项 dropped |
| ② 其余选项各一句不取理由 | 余各一句不取理由 | 其余选项 → 余 (the remaining options, set
against 一) |
| ③ 看不见什么明说 | 盲区明说 | 看不见什么 → 盲区 (blind spots) |
| ④ ⛔ 不因此给第二推荐 | ⛔ 不因此另荐 | 给第二推荐 → 另荐 (recommend another); 因此 and ⛔ kept
|
| ⑤ ⛔ 不沿用他席推荐 | ⛔ 不沿用他席推荐 | verbatim, byte-identical to the old line's
tail |

Nothing is added: no new clause, no new term, each clause keeps its own
separator and its own ⛔. A shared-⛔ variant that kept 只荐一项 whole
(`…盲区明说,⛔ 不因此另荐,不沿用他席推荐。`, 119 B) was weighed and refused, because it
lets 因此 read across clause ⑤.

## Line budget

| | before | after | ceiling |
|---|---|---|---|
| file lines | 50 | 50 | 50 |
| line 9 bytes | 119 | 112 | 120 |
| line 26 bytes | 110 | 120 | 120 |

Control: with the card's verbatim text placed at line 26, the gate's own
`scanLineLengths` reports `[{"line":26,"bytes":150}]`; on the landed
tree it reports `[]`.

## Premise checks on `origin/main` `f4755a68a`

- The card's line numbers 13 and 30 are off: the two strings sit at
lines 9 and 26.
- `git grep -n "回退" f4755a6 -- .claude/skills/pm-dispatch/
.claude/agents/` gives 15 hits. Only these two are decision-analysis
uses; the rest (os-dev wording ×4, `SKILL.md` line 249's `??` fallback,
`dispatch-runbook.md`, `execution-duties.md`, `lanes/triage.md`,
`platform-readings.md` ×3, `reading-discipline.md`) are untouched.
- No script, test or doc pins the old wording: `git grep` for 回退/口径,
给回退项 and 推荐 + 回退 finds only the two lines themselves.

## Verification, all at head `ddf97945b`

The 17 families `node scripts/pm/dispatch-gates.mjs --commands --repo
objectstack-ai/objectstack` derives for this diff, each exit code
captured before any pipe:

| Command | Exit |
|---|---|
| `node scripts/check-closing-keyword-parity.mjs` | 0 |
| `node scripts/check-closing-keyword-parity.mjs --self-test` | 0 |
| `node scripts/check-comment-mask-corpus.mjs` | 0 |
| `node scripts/pm/check-harness-current.mjs --self-test` | 0 |
| `pnpm --filter @objectstack/lint run check:doc-formula-expressions` |
0 (first run 3, PREREQUISITE NOT MET: `@objectstack/formula` and
`@objectstack/lint` unbuilt in the fresh worktree; built under the
verify lock, rerun 0) |
| `pnpm check:agent-test-spelling` | 0 |
| `pnpm check:cross-package-test-inputs` | 0 |
| `pnpm check:doc-authoring` | 0 |
| `pnpm check:driver-memory-census` | 0 |
| `pnpm check:gitlink-declared` | 0 |
| `pnpm check:nul-bytes` | 0 |
| `pnpm check:pm-governed-merges` | 0 |
| `pnpm check:pm-skill-id-lint` | 0 |
| `pnpm check:pm-skill-ratchet` | 0, `decision-analysis.md is 50 lines
(ceiling 50; headroom 0)`, no line-length offender |
| `pnpm check:refd-timer-probe` | 0 |
| `pnpm check:skill-frame-sync` | 0 |
| `pnpm check:watch-hint-literal` | 0 |

`dispatch-gates --ran` with the exit-coded record: `17 derived, 17 run,
0 NOT-MEASURED, 0 UNRUN` (a derived zero). Also run: `pnpm
check:pm-settings-deny-roster` exit 0 (its roster lives under
`.claude`). `node scripts/pm/check-governed-merges.mjs --branch` answers
GOVERNED, Tier S (exit 3, the governed code). Text-only change: no
tests, no ablation. The type-check lanes, `test:repo` and `pnpm lint`
belong to CI.

## Acceptance notes

- Tier S (`.claude/**` only). This PR stays a draft; ready and queue
belong to the owning seat after an isolated contract-tier review.
- No changeset: `.claude/**` publishes nothing; `skip-changeset` is
declared with this PR's creation.

## 维护者速读(草稿)

- **改了什么:** 决策分析写法删掉「回退项」。总监批前增量从「推荐/回退/口径」改为「推荐/口径」;卡面推荐条改为「推荐 +
置信缺口」:只荐一个方案,其余方案各用一句话说明为何不取,看不见的地方明说,但不因此再给第二个推荐。只动这两行,文件仍是 50 行。
- **为什么改:** 您指出「最终结论应该只有一个」。旧模板要求每次附一个回退项,实际常是另一条完全不同的开发方向,等于把决定又推回给您。
- **风险与代价(含回滚):** 纯文本,不发布,不影响产品运行。第 26 行为了压进每行 120
字节的上限做了措辞压缩(「其余选项」→「余」、「看不见什么」→「盲区」、「给第二推荐」→「另荐」、「只荐一项」→「只荐一」),五个分句全部保留,但不是卡上原文逐字。回滚
= revert 本 PR 的一个提交。
- **席位意见:**
- **你要做的:** 无。本 PR 属 Tier S,由席位隔离复核通过后入队落地。

---
_Generated by [Claude
Code](https://claude.ai/code/session_01HEdP5tPhqQfzwRmKDxwn85)_

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

Author: objectstack-fleet[bot]

.claude/skills/pm-dispatch/references/decision-analysis.md

@@ -6,7 +6,7 @@
 
 - 每批恰好 5 张,末批不足照呈;呈完即停等回批,⛔ 永不一条消息投多批。
 - 批序:在飞被阻塞 > 运营阻塞 > 用户可见 > 结构性;卡面分析只一层 = 立卡席六项 + 四棱。
-- 总监批前卡评论是增量:命令复核前提,逐棱同异,推荐/回退/口径;不同意字母才展开四维。
+- 总监批前卡评论是增量:命令复核前提,逐棱同异,推荐/口径;不同意字母才展开四维。
 - 回批一行式逐卡对应字母选项;带的附加条件随执行落地,⛔ 不为它另起请示轮。
 - 呈递详解 = 逐卡散文按正文六项;⛔ 表格行、判词、栅格不构成详解,压缩批次是违规。
 - 受管草稿批例外:一 PR 一行(改了什么/为什么/席位意见/一个动作),⛔ 不套六项;速读住 PR。
@@ -23,7 +23,7 @@
 - 业务含义直译:每选项配一句业务上等于什么,允许行业类比。
 - 四轴从业务立场论证(含四棱行):论据是客户、产品或事故后果,机制名词只作括号补充。
 - 防 AI 犯错轴写清出错时谁看到什么:响亮拒绝,还是静默泄露。
-- 推荐 + 回退 + 置信缺口:荐一项、给回退项、明说看不见什么;⛔ 不沿用他席推荐。
+- 推荐 + 置信缺口:只荐一,余各一句不取理由;盲区明说,⛔ 不因此另荐;⛔ 不沿用他席推荐。
 - 推荐先只按①得出:终态句写两年后该能力的平台样子与主流平台建模,点名参照。
 - ②③④ 与工作量只调时序分期 ⛔ 不翻字母;契约迁就实现、分期半量 ⛔ 不作推荐。
 - ①读数相近而工作量差 ≥3× 推荐须点名;翻字母只凭维护者卡上另裁,唯④新增门禁例外。