Delete a branch asynchronously (ARD-1453), in parity with branch create.

Returns 202 + an operation handle; the CLI polls GET /v1/operations/{operation_id} for staged progress and the terminal outcome. The teardown runs in the branch_delete.v1 DBOS durable workflow (mark deleting -> external cleanup -> finalize), so a crash mid-teardown is recovered by DBOS rather than stranding the row in deleting with no retry driver. Audit ordering and the partial-cleanup terminal-failed semantics are preserved inside the workflow.

Idempotency:

• A concurrent or retried delete of the same branch coalesces onto the active operation via the (type, resource_id) partial unique index — no second workflow is enqueued.
• A re-delete of an already-gone branch polls the prior delete operation to completion (clean no-op) instead of 404ing.