codeanalyzer-java: Neo4j v2 projection — L4 SDG overlay (after L4; base+L3 landed via #198)

[sinha108]

Status (updated 2026-08-28)

Base relabel (L1/L2) + L3 overlay landed via #198 (PR #201); only the L4 SDG overlay remains here. The projection was de-consolidated and shipped pre-L4: #198 projected the base v2 families and the L3 dataflow overlay (J_CFG_NEXT/J_CDG/J_DDG), bumped the graph catalog to 2.0.0 (with the L4 families declared but not projected), and performed the v2 default-output flip (--schema defaults to v2, released as 3.0.0). This issue now tracks only the remaining L4 SDG overlay (J_PARAM_*/J_SUMMARY projection), to land after L4 (#184).

Original consolidation plan (2026-08-24, superseded)

This issue originally covered all Neo4j v2 projection work as one consolidated pass sequenced after L4 lands (base + L3 + L4), to do the Cypher/schema mapping once against an L4-stable schema, with the default-output flip moving to post-L4. #198 revised that: base+L3 (and the flip) shipped before L4, so only the L4 overlay remains here — two catalog touches (2.0.0 now, an L4 bump later) rather than one.

Spec

https://github.com/codellm-devkit/codeanalyzer-java/blob/main/docs/design/specs/schema-v2-l3-l4-design.md (§11 Neo4j projection)

Problem

neo4j/SchemaCatalog is at v1 1.0.0 with old node families (JCompilationUnit, JType is_* booleans, JCallable with code) and J_CALLS carrying type/weight/source_kind/destination_kind. It must project the full v2 CPG (L1 through L4).

Scope — v2 Neo4j families/relationships (base + L3 landed via #198; L4 overlay remaining)

Reuse the existing GraphProjector/RowBuilder/GraphRows/CypherWriter/BoltWriter machinery; keep the dual-label JSymbol merge pattern and _module provenance.

• Base (L1/L2): node families JApplication, JModule, JSymbol (+ specific type-kind label), JCallable, JField, JParameter; J_-prefixed relationships (D9) J_HAS_MODULE, J_DECLARES, J_HAS_CALLABLE, J_HAS_FIELD, J_RESOLVES_TO, J_CALLS(weight,prov), J_EXTENDS, J_IMPLEMENTS.
• L3 overlay: body/CFG nodes (merge key @line:col/@tag, props kind/start_line/end_line/_module) + J_HAS_BODY_NODE, J_CFG_NEXT(kind), J_CDG, J_DDG(var,prov).
• L4 overlay: formal_in/out + actual_in/out vertices + J_PARAM_IN, J_PARAM_OUT, J_SUMMARY.

Goals

• Base v2 node families keyed by can:// id + the base J_ relationships (above) — feat(neo4j)!: schema v2 graph projection; v2 becomes the default schema (3.0.0) #198 / PR feat(neo4j)!: schema v2 graph projection; v2 becomes the default schema (3.0.0) #201
• SchemaCatalog major version bump; --emit schema reflects every family through L4 (2.0.0; L4 families declared) — feat(neo4j)!: schema v2 graph projection; v2 becomes the default schema (3.0.0) #198 / PR feat(neo4j)!: schema v2 graph projection; v2 becomes the default schema (3.0.0) #201
• L3 overlay families/relationships — feat(neo4j)!: schema v2 graph projection; v2 becomes the default schema (3.0.0) #198 / PR feat(neo4j)!: schema v2 graph projection; v2 becomes the default schema (3.0.0) #201
• L4 overlay families/relationships (J_PARAM_*/J_SUMMARY) — remaining, after L4 (codeanalyzer-java: L4 SDG — param_in/out + semantic DDG #184)
• --emit neo4j runs at full implemented depth; -a/--graph-field-depth alongside --emit neo4j is an explicit non-zero error — feat(neo4j)!: schema v2 graph projection; v2 becomes the default schema (3.0.0) #198 / PR feat(neo4j)!: schema v2 graph projection; v2 becomes the default schema (3.0.0) #201

Caveats and known risks

• Families rename = major graph-schema change; moves in lockstep with the JSON schema.
• Deferred-edge (no-dangling) discipline; unresolved-string fallback rides on the source node's props.
• Originally planned as one post-L4 schema bump; feat(neo4j)!: schema v2 graph projection; v2 becomes the default schema (3.0.0) #198 instead shipped base+L3 now (2.0.0), leaving one further catalog bump for the L4 overlay.

Definition of done

• --emit schema produces a schema.neo4j.json listing every v2 node family the JSON emits, through L4.
• --emit neo4j (no URI) writes a graph.cypher that loads clean against an empty Neo4j and is idempotent.
• Cross-projection gate: Neo4j node/edge counts at full depth match the JSON at max_level (modulo explicit HAS_* containment); no dangling relationships.