fix(docs): repair zh-cn 15.8 markup, and run the lint that was added but never ran

[marevol]

Two follow-ups to #465.

Bold text that never opened, in the zh-cn 15.8 LLM guides

Inline markup cannot open against a CJK character, so 实现其**AI搜索模式 was never read as strong emphasis. The asterisks stayed in the text and reached the rendered page and its meta description.

The 15.7 copies of these three pages were repaired in #465. These are their 15.8 counterparts, left out at the time to avoid colliding with work in progress, and fixed here with the same escaped space this repository already uses for CJK-adjacent markup.

zh-cn/15.8/config/rank-fusion.rst carries the same defect and is deliberately not touched — it is already repaired on the branch that rewrites the 15.8 semantic search guides, and changing it here would only conflict with that.

15.8 is not published yet, so nothing live changes. This keeps the defect from shipping with the release.

CI

This repository had no CI at all. tools/check_headings.py was added in #465 to stop a defect class from coming back — a section title whose overline and underline differ is dropped by docutils, and the next section silently becomes the page title without anything failing the build — but nothing ran it, so it could only catch what someone remembered to check by hand. The same was true of tools/test_update_eol.py.

Two jobs, nothing beyond the standard library:

• tools — the unit tests, plus update_eol.py --check, which fails when an eol.rst has drifted from versions.json. Its date warnings are reported without failing the build, so time passing cannot turn the build red on its own.
• section titles — the heading lint, scoped to the versions versions.json calls current. Reading the version from versions.json means a release moves the scope by itself instead of needing an edit here, and it keeps the archived versions, which still carry mismatches and are consolidated by canonical anyway, from blocking unrelated work.

Verified green on the current tree, and red with the expected one-line report when a heading rule is shortened by a single character.

Noted, not changed here

update_eol.py reports that 14.19 reached end of life on 2026-08-01 while versions.json still lists it as supported. Whether a version is still supported is a decision recorded in versions.json, not something a tool should infer, so it is left for the maintainer.