F5: Neo4j backends → schema-v2 projection (CanNode reconstruct, contract-driven)

[rahlk]

PROBLEM

The Neo4j backends and reconstruct.py read the old node labels and properties. Schema v2 projects a different graph (:CanNode families, containment as HAS_*/DECLARES edges, overlay relationships), published as a machine-readable schema.neo4j.json per analyzer.

SCOPE BOUNDARY

TS + Python graph backends; Java's is the Java-lane child (#228). Read-only backends — ingestion stays the analyzers' --emit neo4j job (#160's split stands).

GOALS

1. Reconstruct the cpg models/views from the v2 node/relationship families, driven by each analyzer's published schema.neo4j.json (not hand-maintained label lists).
2. Read capabilities off the :Application node (schema_version, max_level) — same capability model as the local backends.
3. Backend parity vs the F1 fixtures "modulo documented lossiness": body text is absent when source was not projected — documented, never silently absorbed.
4. Closes the TSNeo4jBackend.get_call_graph() omits synthesized anonymous-callback nodes — backend parity gap #176 class of parity gaps by construction (one reconstruct path from one schema contract).

CAVEATS AND KNOWN RISKS

• The graph is always full-depth per the keystone (levels gate the JSON path only) — do not level-gate graph reads.
• Requires a seeded Neo4j fixture in CI (container or recorded sessions); coordinate with F1's harness.

DEFINITION OF DONE

• Local vs Neo4j answers identical on the fixtures, modulo the documented lossiness list.
• Conformance test against each analyzer's schema.neo4j.json green.
• TSNeo4jBackend.get_call_graph() omits synthesized anonymous-callback nodes — backend parity gap #176 closed.

Part of #238