Story: Refactor conventions onto families and bundles

A convention is declared once, in its own entity, and carries only its own fields. When this story closes, nothing in the code holds every family at the same time, so a change to one family cannot break another family's build. An import applies the families it carries as one unit of work, and a failure names the family that failed. The ORE conventions file is a package the boundary maps, not a type the components share.

The story closes on the tasks it finished:

• Model each convention family on its own.
• Generate a protocol and handler per family.
• Declare the conventions bundle.
• Rewire the ORE path to families and packages.
• Replace the web family list with the registry.
• Reconcile the record with the merged work.
• Split the ORE side into per-family units.

• Closed at sprint 28 close with its 7 done tasks. The 3 open tasks and the original acceptance moved to Refactor conventions onto families and bundles — the remaining work in the product backlog inbox.
• A bundle is the largest atomic unit. The import stages one row per family and applies them in one unit of work, so a reader sees either all of the bundle or none of it. Nothing smaller is atomic and nothing larger is offered.
• The family list is generated once and every consumer derives from it. A hand-maintained list beside it is a second source of truth and drifts.
• The registry cannot dispatch by itself. A table of strings cannot call a typed repository, so the same generation emits the per-family apply, and the service calls generated functions instead of listing families by hand.
• The ORE FX convention is a projection, not an entity. It is a currency pair, that pair's convention, a spot lag and the pair's advance calendars, all of which are entities, so fx stays out of the registry and the boundary mapper derives it.
• The tenor basis swap convention spans the shared convention row and its two part families, because that is what the document carried. Collapsing it to one registry entry is a later decision.
• The MSVC failure is accumulated context in the handler translation unit, not one over-wide literal. The investigation in Fix the tenor basis swap convention model and the document TU measured it, and task 8 follows that finding.
• The conventions import owns no state of its own beyond the staging rows. A retried begin names the import it already opened, and a commit on an import that is not staging is refused, so a crash mid-import leaves an import that can be read and committed. Its expires_at is stamped at begin, 24 hours ahead by default, though nothing sweeps an expired staging import yet.

• Collapsing the tenor basis swap family into one registry entry.
• Changing the ORE XML model. The generated conventions.hpp stays as it is. This story changes what crosses a component boundary, not the file format.