docs(pm): record three measured auto-merge/label platform readings (#17523)

Three deduplicated readings land in the pm-dispatch platform readings
table, in the queue-membership and auto-merge section:

- arming auto-merge on a PR whose checks are all complete enqueues at
  once (28-60 s from the enable call to `added_to_merge_queue` here),
  so no window stands between arming and landing;
- a call made before mergeability settles arms classic auto-merge
  without enqueueing; the settled call is the one that enqueues;
- a GitHub label description is capped at 100 characters, longer ones
  refused with 422.

A fourth candidate reading — that an empty-method success line from
`enable_pr_auto_merge` is a real arming rather than the documented
graceful no-op — is already carried on the tree and lands no line.

The line ceiling for the readings file moves 422 -> 425 under the
standing one-file exception, recorded as a `ruledRaises` record that
quotes it.

Claude-Session: https://claude.ai/code/session_01YKEjmbYNvYWJvWGSWx26zK

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

Author: os-litant

.claude/skills/pm-dispatch/references/platform-readings.md

@@ -81,8 +81,11 @@
 - 入队事件与队列 ref 迟 1–3 分钟才出现 ⇒ 轮询预算按 3 分钟,⛔ 不按 1 分钟判没挂上。
 - 队列窗口有界:满窗条目 ref 与 `merge_group` run 双缺席,ref 随前一条落地才现,不按计时器。
 - ready 翻转触发检查重跑 ⇒ 入队落在翻转之后约一分钟,那段空窗不是挂载失败。
+- 检查全部完成的 PR 挂 auto-merge 即入队,本仓 28–60 秒 ⇒ 挂载与落地之间无窗口。
+- `mergeable_state` 未落定时挂上的是经典 auto-merge、不入队,落定后再挂才入队。
 - `behind` 的 PR 照常入队:落后于 main 不是入队否决,⛔ 不为它先跑 update-branch。
 - `check_suite.completed` 会命名过期 head,check-run 也只属最后一次 push ⇒ 用前先重读当前 head。
+- 落地相邻的写侧一则:GitHub 标签描述上限 100 字符,超长写回 422,133 字的原文即不可存。
 
 ## API 配额
 

scripts/pm/check-skill-line-ratchet.mjs

@@ -733,7 +733,40 @@ export const CEILINGS = new Map([
   // line-neutral folds among this file's adjacent rule pairs, and re-wrap
   // funding is refused per the 2026-08-17 rule in any case. Landed count,
   // headroom 0, same convention.
-  ['.claude/skills/pm-dispatch/references/platform-readings.md', 422],
+  // Raised 422 → 425 by the TENTH readings increment, again under the STANDING
+  // one-file exception quoted above rather than a fresh decision card, and again
+  // recorded as a `ruledRaises` record citing it. Spent at ONE line per
+  // deduplicated reading, each written in this file's voice: arming auto-merge on
+  // a PR whose checks are ALL complete enqueues AT ONCE — 28–60 s from the enable
+  // call to `added_to_merge_queue` measured here, 17–20 s enqueue→merge across a
+  // dozen PRs on the sibling repo the carrier card was filed from — so no window
+  // stands between arming and landing (+1); a call made before mergeability
+  // settles arms CLASSIC auto-merge WITHOUT enqueueing, and the settled call is
+  // the one that enqueues (+1); and a GitHub label description is capped at 100
+  // CHARACTERS, a longer one refused with `422`, measured on a maintainer ruling
+  // fixed verbatim at 133 characters and therefore unstorable (+1) = +3 exactly.
+  //
+  // ⚠️ The candidate set was FOUR, and the cut is the exception's own dedup
+  // condition doing its work rather than an author trimming. The fourth reading —
+  // that an `enable_pr_auto_merge` success line carrying an EMPTY method and
+  // timestamp is a real arming and not the documented graceful no-op — is already
+  // on the tree in a stronger and more general form, at the two rows reading
+  // 「`enable_pr_auto_merge` 对已 `mergeable_state: clean` 的 PR 照样成功,与工具描述
+  // 的优雅失败相反」 and 「回显两向不可靠,空回显不等于未挂上」. The carrier card's own
+  // measurement was taken against `:13–:14` alone, which carry the FIELD's
+  // instability rather than the ECHO's, so the reading it drew — that the
+  // disambiguation is absent — was falsified on the tree before any line was
+  // written. Candidates 4 / landed 3 / already present 1 / refused 0.
+  //
+  // ⛔ The carrier card's who-flips half lands NO line here in any form: it is
+  // closed by `references/landing-operations.md`, which already holds it, and a
+  // second copy is what the consolidation ruling above exists to refuse.
+  //
+  // Nothing was paid in place: the fourth increment MEASURED zero line-neutral
+  // folds among this file's adjacent rule pairs, and re-wrap funding is refused
+  // per the 2026-08-17 rule in any case. Landed count, headroom 0, same
+  // convention.
+  ['.claude/skills/pm-dispatch/references/platform-readings.md', 425],
   // Per-operation REST/GraphQL/git channel mapping — which fleet operation has
   // a REST twin (each row executed in a real session, provenance date carried
   // per row), the handful that are GraphQL-only, and the queue-routing
@@ -1286,6 +1319,23 @@ export const CROSS_FILE_MOVES = new Map([
           date: '2026-09-10',
           delta: 3,
         },
+        {
+          // The TENTH increment, under the same STANDING exception — the same
+          // words again, for the same reason: a record that quotes no ruling is
+          // RED and each record stands alone. The +3 is accounted for line by
+          // line beside this entry's ceiling above, and the exception's own
+          // conditions (per-item verification and the candidate / landed /
+          // already-present / refused counts) are carried by the raising PR's
+          // dedup table and the seat's ACCEPT.
+          ruling:
+            'the standing one-file exception for'
+            + ' `.claude/skills/pm-dispatch/references/platform-readings.md` — pm-dispatch'
+            + ' SKILL.md, verbatim and untranslated: 「唯一例外:`platform-readings.md`'
+            + ' 增量抬上限到落地行数,免决策卡,记 `ruledRaises` 引常设裁决。条件:席位验收评论'
+            + '逐条核实、去重计数(候选/落地/已有/拒收)、一事一行、不计重排」',
+          date: '2026-09-10',
+          delta: 3,
+        },
       ],
       sources: [
         ['.claude/skills/pm-dispatch/references/lanes/cli.md', 35],