feat(content): read the title from the field the model declares

[ABB65]

Closes T07b. Stacked on #191 (the migration) — merge that first; without it a bumped Studio fails validation on its own dictionary.

An integration group listed as i-lucide-bot, an article as its slug, a hero slide as f3a81c09d24e. Each is the same bug: Studio guessed which field was an entry's title, and the guess ranked slug alongside string and had no opinion about icon.

@contentrain/types@1.0.0 replaces the guess with a declaration.

package	from	to
@contentrain/types	0.9.0	1.0.2
@contentrain/mcp	2.3.0	3.0.2
@contentrain/query	7.0.1	7.0.4

One resolver, two call sites

app/utils/entry-title.ts. The entry list and a relation's label need the same answer — a hero slide naming an article one way while the article list names it another is how this started.

The old inference survives one step down, minus its two bad rankings (slug demoted, icon/color excluded). Hand-edited models, the migration window and older CLI output all still exist, and Studio has to render them rather than 500.

1. model.title_field          ← the contract
2. a name-like key            title / name / label / heading
3. a required text field
4. any text field that is not icon / color / slug / code / url
5. the first field            so something renders

updated_at had to be written, not just read

The brief asked only that Studio read the new field. Reading it would have returned nothing forever:

• applyStudioMetaOverrides replaces MCP's meta wholesale to keep Studio's autoPublish / updated_by semantics
• updateEntryStatus builds meta by hand

MCP mints the stamp; both of those threw it away. Both stamp it now — one timestamp per call, so a bulk status change reads as the single operation it was. Without this, T06's "sort by recently edited" would have had no data for precisely the entries Studio users create.

The picker writes directly, not through chat

Every other action in that header goes through sendChatPrompt. This one does not: an MCP write takes ~17-18s, and a radio button that stays unconfirmed that long reads as broken.

[modelId].patch.ts gains a titleField branch, deliberately outside the forms.enabled gate — which field titles an entry is part of the model contract, not a forms feature, and gating it would leave listings unreadable on every plan without forms. There is a test that fails if someone moves it back inside. Owner/admin is kept: it writes a model definition and produces a git commit.

The chat path needs nothing — the agent already calls contentrain_model_save with the full definition.

What the bump moved, beyond title_field

The standing rule after an MCP major is to diff its git internals rather than trust a green suite (1.8.0's mergeBranch silently started deleting the source branch). Doing that:

• GitHubProvider is byte-identical bar import hashes. Studio's write path is unchanged.
• "a document save no longer replaces what it did not send" is neutral here. Studio already merges frontmatter and body before handing the entry to planContentSave, for both documents and collections — the two merges compose.
• planVocabularySave / planVocabularyDelete are new. Studio's vocabulary PATCH still does its own read-modify-write; not switched in this PR, because the retry loop from fix(vocabulary): stop concurrent saves from dropping each other's terms #186 guards a lost-update race the plan helper does not. Worth revisiting.

Verified

• pnpm lint — 0 errors (7 pre-existing warnings)
• pnpm typecheck — clean, including against the now-required ModelDefinition.title_field
• pnpm test — 143 files / 1225 tests
• 17 unit tests on the resolver (each of the three reported miss-titles has one), 4 more on relation labels, 7 integration tests on the PATCH branch, 2 on the updated_at stamps

Not verified — needs staging

The acceptance criterion the brief calls non-negotiable for an MCP major:

• a real entry saved on staging. Mocked git tests hide write-path behaviour changes; the diff above is evidence, a real save is proof.
• the picker end to end, and the list re-rendering after it
• "make the title field title" through chat, and the list refreshing
• the three reported models actually reading correctly (Integration Groups, Articles, Hero Slides)