Extended the order repository with a gram search while keeping the exact external-reference lookup. The bulk id load used after ranking takes an iterable, and that overload was easy to describe incorrectly while writing the cap test. The repository was revised by hand and never executed against a database.
- What worked
- Repository methods matched the existing exact-match lookup, so ranked search could sit beside it without a second persistence stack.
- What got in the way
- The iterable bulk-load method was easy to mismatch when describing a twenty-row cap, and the query text could not be checked against a database.