D304 — Map titles are reviewed as one set, and renamed through map_update

Decision. kb_review emits one map_naming item when a KB's map titles (three or more) mix a subtitle with none, or Title Case with sentence case. It names every map as <map>/_map and lists each folder and title as evidence. map_update takes a title, which renames the map in _map.md and index.md, H1 included, and never touches the folder.

Why. On a real KB the rail listed maps named in two languages, with and without a subtitle, in two capitalisations. That is noise every reader pays for, and nothing reported it. Nor could an agent fix it: a map's title lives in four places (_map.md and index.md, frontmatter and H1), and map_update changed only the contract. The server measures the two properties it can measure in any language: the shape and the capitalisation of a word. Whether titles share one language, or match their folder, is a judgement the item hands to the agent by listing every title. The cost is one more review kind, and a pseudo-ID for a map that is not a concept.

Alternatives rejected. - One item per map in the minority shape: renaming is one scheme for the whole set. Split into items, the operator would agree a scheme once per map. - Detecting language mix or folder mismatch: no generic test separates a folder id whose title is a translation from one whose title is a different subject. A word-overlap test would flag every KB whose folders are English ids and whose titles are not. - Renaming the folder too: every link and every scope rule names the folder. A title is free to change, a folder is an ID.

Consequences. map_naming is whole-graph, so a caller restricted to part of the KB never sees other maps' titles. It is dismissed by lint_ignore in a named map's _map.md, which the review reads for that purpose alone. The kb-doctor skill proposes the scheme, and renames only after the operator agrees.