Every column a reference table's field_name may name, with type and example.

Served from the same process-lifetime cache the write path validates against, so the
picker and the validator can never disagree about what a field name may be, and a page
load costs no hunt-api call. The registry changes only when the information-model sync
runs, so a cached answer is the current answer; a restart picks up a new one.

The body is the registry as hunt-api sent it -- descriptions, examples, timestamp
fields, sync stamps -- with two exceptions. Columns this service excludes from
reference tables are removed, because a picker offering a column the save rejects is
the bug this endpoint exists to prevent; and field_count is recomputed to match
the list actually returned.

It lives under /reference-tables because that is what it is for: the field picker
on a reference table needs the same list the SQL validators judge a field_name
against. The data is org-agnostic -- the same registry for every tenant -- so a
session is required to call it but nothing here is scoped to one.

A 503 means the registry has never been fetched here and the schema service
cannot be reached, or has not synced yet; it is worth retrying. Once a fetch has
succeeded, a later outage is invisible -- the cached registry keeps being served.