GeezerVison

Refs #1865.

When more nodes or edges match than a preview can display, the database's storage order decides which rows are shown. Rebuilding the same graph in a different file order can therefore change the search, call graph, or override preview even though the code is identical.

Sort those rows by source identity before applying their existing limits. For call edges, also use the destination and call-site metadata. Typed searches retain a label-specific match and use properties available on that label in the embedded schema.

The regression builds the same 200-file graph in opposite insertion orders. All six cases (name, pattern, type, incoming calls, outgoing calls, and overrides) fail on the base and pass with the fix. The focused Kuzu run passed 38 tests, and all six regressions passed on Ladybug. Ruff, Black on the new tests, and git diff --check passed. This covers the six preview branches above. Call-chain, dependencies, inheritance, the offline default and other query paths listed in #1865 are outside this change. Neo4j and FalkorDB were outside local validation.