Contentrain/ai

Minor Changes

• d617dab: fix(mcp): stop content_save unpublishing entries, and make bulk update_status persist every id

  Four publish-status bugs, all found on a live project against the CDN. Each one
  reported success while content quietly stopped being delivered.

  contentrain_content_save no longer resets an entry's status. It rebuilt
  meta from scratch on every write, so editing one field silently moved a
  published entry to draft — and the next CDN build served the collection as
  {}. Editing a field is not a publish decision, and per this repo's own split
  (MCP is deterministic infra; the agent is intelligence) MCP should never have
  been making it. An existing entry now keeps its status, approved_by and
  version; only a genuinely new entry starts at draft. source/updated_by
  still describe the current write. The same reset lived in a second copy behind
  contentrain_apply and scaffolding — both now share one mergeEntryMeta.

  contentrain_bulk update_status no longer drops entries. It launched one
  writeMeta per entry ID through Promise.all, and every call read the same
  snapshot of the shared {locale}.json and rewrote the whole file — so N-1
  updates were lost while the response reported all N as updated. It is now a
  single read-modify-write per locale file, and updated counts what actually
  persisted. copy_locale had the identical race and wrote 1 meta record instead
  of N. Neither had any test coverage; bulk now has a suite.

  update_status works on singletons and dictionaries. The entry_ids guard
  ran before the model-kind guard, so a singleton had no reachable path: omitting
  entry_ids failed with "requires entry_ids", supplying them failed with "only
  supported for collection models". Call it without entry_ids for these kinds.
  It also takes an optional locale now, instead of always rewriting every
  supported locale.

  Non-i18n models keep exactly one meta record. Content collapses to a single
  data.json while meta was still derived from the caller's locale, so one
  content file could end up with meta/{id}/tr.json and meta/{id}/en.json
  and readers disagreed about which was authoritative. metaFilePath now takes
  i18n and the default locale and pins the record there. This also fixes
  non-i18n collection deletes, which looked for meta/{id}/data.json — a file
  that never existed — and orphaned the meta entry.

  contentrain_doctor's SDK freshness check works again. It compared
  directory mtimes, but generate rewrites the client files in place (which never
  moves the directory's mtime) while a selective sync recreates model files via
  git checkout (which does). Once you had saved a model, it reported "Stale"
  permanently. It now compares the newest file mtime under each directory.

  contentrain_validate gained two checks for the class of failure above,
  since it reported 0 errors throughout: a notice for drafts sitting alongside
  published entries in one collection, and a warning for a non-i18n model whose
  meta layout disagrees with its content layout. Neither is auto-fixed —
  publishing is a content decision.

  MIGRATION — read before upgrading Studio. Projects that ran an affected version
  have singletons and entries sitting at draft that were never meant to be. That
  is currently harmless, because the CDN publishes singletons and dictionaries
  regardless of status. When Studio starts enforcing status for those kinds, that
  content will disappear from the CDN. Upgrade here first, run
  contentrain_validate to find the drift, restore it with contentrain_bulk update_status, and only then take the Studio change.