Migrate to canonical schema v2: can:// ids, statement-level CPG, identity-only edges

[rahlk]

Parent: codellm-devkit/.github#35 · Roadmap: docs/design/roadmap.md (candidate 9)
Identity grammar spec: can-uri-service-segment.md · codellm-devkit/.github#39

Skeleton — bodies to be filled in by the maintainer.

Problem

codeanalyzer-java is the only analyzer still on schema 1.1.0, and the only one whose shape canonical v2 explicitly forbids rather than merely differs from.

• No can:// ids. Python and TypeScript both emit them; Java does not.
• No statement-level nodes in the Neo4j projection. schema.neo4j.json declares 16 labels — JApplication, JCompilationUnit, JType, JCallable, JCallSite, JField, JParameter, JVariable, JAnnotation, JComment, JCrudOperation, JCrudQuery, JEnumConstant, JInitializationBlock, JPackage, JRecordComponent — and none is a body/statement node. Deferred deliberately in .claude/SCHEMA_DECISIONS.md Callable signature and patch for expression resolution failure #9: "Statement-level CPG (CFGNode etc.) stays a follow-up." This is that follow-up.
• Dependence edges are callable→callable. J_CONTROL_DEP / J_DATA_DEP / J_HEAP_DATA_DEP connect JCallable to JCallable (SCHEMA_DECISIONS.md Callable signature and patch for expression resolution failure #9), where canonical v2 places cfg / cdg / ddg at statement→statement on the callable.
• Rich edges. The emitted shape drives python-sdk's JGraphEdges, whose source and target are full JMethodDetail objects rather than node ids (cldk/models/java/models.py:517-523). Canonical v2's expansion rubric: "Never: … add a rich-edge variant."

Scope boundary

Brings the analyzer's emitted contract to canonical v2. Does not change python-sdk's Java models — that is codellm-devkit/python-sdk#309, and it is blocked on this issue, not the reverse. Does not settle the entrypoint vocabulary (JEntrypoint marker + is_entrypoint), which is a separate design session.

Goals

• can://<service>/<lang>/<file>/<type>/<sig> ids, with Java's signatureOf() pinned (generics, overloads, <init>) — note the service segment is outermost and there is no <app> segment (spec D1/D2); the app-name CLI option becomes --service, same default (input directory name)
• Statement-level body nodes in both projections, on the settled body-node model
• cfg / cdg / ddg / summary re-anchored from callable→callable to statement→statement
• Identity-only edges — endpoints are ids, no nested callables
• SCHEMA_VERSION 1.1.0 → 2.1.0 (not 2.0.0 — Java emits no can:// id today, so it lands directly on the service-segment grammar and never has to migrate off 2.0.0), with .claude/SCHEMA_DECISIONS.md updated
• Passes the shared conformance suite (Shared conformance suite: check every analyzer against canonical v2 .github#38) and the projection-parity gate (Projection-parity gate: prove analysis.json and the Neo4j graph agree .github#37)

Caveats and known risks

• WALA node identity is SSA instruction order, not AST source spans (SCHEMA_DECISIONS.md Update excluding signature-related files #5): node_id 0 is a synthetic ENTRY, then SSA instructions by iindex. Canonical v2 addresses sub-callable nodes by @line:col. These do not map cleanly — WALA nodes are SSA instructions, not AST regions, and source lines come from ECJ/CAst positions with a -1 sentinel when unavailable. This is the hard part of the issue and may force either a Java-specific local-id form recorded in SCHEMA_DECISIONS.md, or a change to the canonical grammar. It should be raised in the contract spec (Canonical Neo4j projection contract: merge labels, body-node model, can:// grammar .github#36), not discovered during implementation.
• CFG edge kinds are derived, not labelled by WALA (SCHEMA_DECISIONS.md Update main.yml #6) — true/false by successor order, loop_back by iindex comparison. Deterministic but approximate; carry the documentation forward rather than silently absorbing it.
• No SUMMARY edges today (SCHEMA_DECISIONS.md Add thrown exceptions to callable info in symbol table #7) — WALA computes HRB summaries lazily inside Slicer.
• This is a breaking change for every python-sdk consumer of JGraphEdges. It must ride one migration with codeanalyzer-python 3.0.0 and the SDK, per the epic's release plan.
• TODO

Definition of done

• TODO — exact conditions. At minimum: the conformance suite and parity gate both pass, and no JGraphEdges-shaped rich edge remains in the emitted contract.

[rahlk]

Folding the can:// service segment into this migration rather than filing a separate issue,
since Java has not emitted a can:// id yet and would otherwise implement the old grammar and
immediately migrate off it.

Spec: can-uri-service-segment.md · coordination codellm-devkit/.github#39 · siblings codellm-devkit/codeanalyzer-python#123, codellm-devkit/codeanalyzer-typescript#91.

What changed in the body above — two goal lines:

- can://<lang>/<app>/<file>/<type>/<sig>
+ can://<service>/<lang>/<file>/<type>/<sig>      # service outermost, <app> collapsed into it

- SCHEMA_VERSION 1.1.0 → 2.0.0
+ SCHEMA_VERSION 1.1.0 → 2.1.0

Java is the cheapest repo in this change: Python and TypeScript have to rewrite live ids
(83 and 34 references), while Java has none, so it lands on the final grammar first time. The
app-name CLI option becomes --service, defaulting to the input directory name.

Two things this does not touch, flagged for the maintainer rather than silently edited:

1. The caveat "It must ride one migration with codeanalyzer-python 3.0.0 and the SDK, per the
   epic's release plan" now looks stale. There is no codeanalyzer-python 3.0.0 — the
   service-segment change was taken as an in-place amendment to v2 (2.0.0 → 2.1.0, spec D3),
   on the grounds that no consumer holds a can:// id today. Whether Java's migration still
   needs to ride a single train with Python and the SDK is a release-plan question that predates
   this spec and is yours to settle.

2. The WALA local-id problem in the caveats — SSA iindex order versus canonical @line:col —
   is untouched and remains the hard part of this issue. That caveat says it should be raised in
   the contract spec Canonical Neo4j projection contract: merge labels, body-node model, can:// grammar .github#36. It still should; the service segment sits above
   the callable line and does not interact with sub-callable identity either way.