docs(ui): name the operator card the way the code names it

[argszero]

Summary

ui/README.md's "设置 / 管理 / 运营布局约定（v1.22）" section still describes the
operator overview's key card with the wording C2158 replaced:

…与上游 key 健康 .mini-list（key_health 按厂商聚合 total/on/off，三态 pill：健康 / N 个异常 / 全部失败）…

That sentence was true when it was written (7d9b56e, PR #174). C2158 (PR #269,
squash e5ee178) then renamed the card 上游 key 健康 → 上游 key 状态 and rewrote the
three pills 健康 / {n} 个异常 / 全部失败 → 全部启用 / {n} 个停用 / 全部停用, because the
data carries no health signal at all — /api/ops/runtime only returns total / on / off
(off = total − on), so a user pausing their own key (a normal action) was being rendered as a
red failure state.

#269 changed the key names, both language packs, the static fallback in ui/index.html and the
render site in ui/js/app.js — and added a new section to this very file that says the opposite
(:1080, "后端不产出任何健康信号"). It left this line untouched, and its edit sheet
(4 files / 7 edits) never listed it. So today the document contradicts itself: the old section
promises a card the code no longer ships.

Changes

• ui/README.md: in the operator-layout bullet, the card is named 上游 key 状态 and its
  three pills are 全部启用 / N 个停用 / 全部停用, with a pointer to the C2158 section below.

One line (+1 / −1). The other claims on that same bullet were re-checked and are still true:
version ← env!("CARGO_PKG_VERSION"), the five uptime_* fields, split_uptime,
fmtUptime's top-two-nonzero units, ops.uptime.{days,hours,minutes,seconds}, and
today_hours zero-filled 0–23.

The single remaining occurrence of the old wording (ui/README.md:1084) is the C2158 section
quoting it as history ("这张卡曾把「停用」渲染成…") and is deliberately left untouched —
editing it would erase the record of the defect. Hence a whole-line replacement rather than a
token-level one.

No gate. This is prose restating an implementation name, and the repository's standing
decision on that axis is to fix the data, not to guard the prose (precedents: PRs #309, #310,
#312–#315). No cache-bust token: ui/README.md is not a served asset.

Related Issue

Tests

• cargo test — 417 passed / 0 failed (unchanged baseline; the change touches no tested byte)
• cargo fmt --check — clean
• cargo clippy --all-targets -- -D warnings — clean
• Independent check that the corrected line agrees with the shipped values: the line contains
  ops.keys.title zh ("上游 key 状态") and ops.keys.allOn / allOff zh values, and no longer
  carries any of the three old pill words. ui/README.md is part of the corpus consumed by
  i18n_pack::every_pack_key_reaches_a_consumer, and that test still passes — the new wording
  introduces no i18n key token.

Checklist

• 分支命名符合约定（docs/）
• Commit message 使用 Conventional Commits 格式
• 单一职责，改动最小化

[argszero]

✅ LGTM (self-review — allow_self_merge is enabled for this task; GitHub does not let an author approve their own PR, so this is a plain comment).

What I checked:

1. The claim is reproducible on the branch tree. The corrected line now carries the shipped values:
   ops.keys.title zh = 上游 key 状态, and ops.keys.allOn / allOff zh = 全部启用 / 全部停用. The three
   old pill words (健康 / 异常 / 全部失败) no longer appear on that line.
2. Provenance is drift, not a choice. git log -L 385,385:ui/README.md shows the line last touched by
   7d9b56e (PR feat(ops): show service version and uptime on the operator overview #174) — not by C2158's PR fix(ui): stop reading a key's status count as a health verdict #269, even though fix(ui): stop reading a key's status count as a health verdict #269 did edit this file. fix(ui): stop reading a key's status count as a health verdict #269's edit sheet
   (4 files / 7 edits) never listed the line, so it was forgotten rather than rejected.
3. The one remaining occurrence is intentional. ui/README.md:1084 quotes the old wording as history
   inside the C2158 section. A token-level replacement would have erased that record; the change is a
   whole-line replacement, so only :385 moved. Verified: 上游 key 健康 count went 2 → 1.
4. The rest of the bullet is still true — version ← env!("CARGO_PKG_VERSION"), the five uptime_*
   fields, split_uptime, fmtUptime's top-two-nonzero units, ops.uptime.* (both packs), today_hours
   zero-filled 0–23. Not a stale bullet overall, just one wrong card name.
5. No gate, by the repo's own standing decision. doc-comment-claims axis = fix the data, not guard the
   prose (precedents docs(perf_gate): drop the stale file count from the gate header #309, docs(ui): stop restating the i18n counters and correct two stale gate claims #310, fix(admin): state the refusal a deleted model now meets, not the old zero bill #312–docs(ui): name only the bump sites, test and hint the code still has #315). ui/README.md is not a served asset, so no cache-bust token.
6. Corpus interaction checked. ui/README.md is part of the corpus read by
   i18n_pack::every_pack_key_reaches_a_consumer; the new wording adds no i18n key token, and that test
   still passes.
7. Suite on the branch tree: cargo test 417 passed / 0 failed (unchanged baseline), cargo fmt --check
   clean, cargo clippy --all-targets -- -D warnings clean. PR CI: test / fmt / clippy pass (2m28s) and
   msrv pass (26s).