python-sdk v2 model layer: one Node/Edge/Application

[rahlk]

Parent: codellm-devkit/.github#35 · Spec: docs/design/specs/2026-09-03-python-sdk-canonical-v2.md (#318)

Problem

python-sdk models the schema four times, once per language (cldk/models/{java,python,typescript,c}). All three analyzers now emit canonical v2 — codeanalyzer-python 1.4.0, codeanalyzer-typescript 1.2.0, codeanalyzer-java 3.0.1 — while the SDK pins 0.3.1 / 0.4.3 and bundles the 2.4.1 JAR, so it is the last consumer still parsing v1.

The spine is common to all three: id / kind / span, source once on the module node, named-map containment, a body{} map keyed by ordinal id, split cfg / cdg / ddg / summary lists on the callable, and identity-only edges at application scope. That agreement is what gets modelled once.

Scope boundary

The shared model layer and the Python leg that validates it — they land together, because a spine designed with no consumer is a spine designed against the spec rather than against a real payload. TypeScript is leg 2 (#307), Java is leg 3 (#310), the query layer is leg 4 (#311).

Does not collapse the per-language facades (PythonAnalysis / JavaAnalysis / TypeScriptAnalysis / CAnalysis) into one — explicitly out of scope for the parent epic. Does not change what any analyzer emits.

Goals

• cldk/models/cpg/ defines the spine once: envelope (Analysis, AnalyzerInfo), containers (Application, Module, Type, Callable), leaves (BodyNode, Field, Parameter, Decorator, Import, Comment, Span), edges (CallEdge, CfgEdge, CdgEdge, DdgEdge, IdEdge), and the repository-artifact layer
• Per-language subclasses narrow the container maps by redeclaring them and add typed extras — no generics, no untyped extra="allow" dicts (spec D2)
• No language model package redefines a spine field
• Python leg: pin codeanalyzer-python==1.4.0, subclasses, view layer, Neo4j backend, facade — pin bump and model migration in the same PR, since v1 models cannot parse v2 output and vice versa
• The PyCallableOverview projection pattern (#180/feat(python): bulk/projected accessors for the Python facade (#180) #181) generalises into the view layer, which is where API stability is bought: .code and .call_sites become computed views over body / source
• The nine-entry divergence register (spec § 5) filed onto Extract and freeze the canonical schema v2 contract from codeanalyzer-python .github#36 as the extraction input
• Ships as 2.0.0-rc.1; closes Incorporate codeanalyzer-python v1.2.0 #315, Incorporate codeanalyzer-python v1.3.0 #316, Incorporate codeanalyzer-python v1.4.0 #317, and Python get_method_bodies lacks the None-code omission filter (twin of the TS #298 rule) #301

Withdrawn from the previous revision

• The Java blocker. The earlier body said the SDK "cannot move to identity-only edges while its producer emits rich ones", and marked the Java half blocked on Migrate to canonical schema v2: can:// ids, statement-level CPG, identity-only edges codeanalyzer-java#179. codeanalyzer-java 3.0.1 emits JCallEdge{src, dst, prov, weight}, JIdEdge{src, dst} and a canonical body map. JGraphEdges and _CALLABLES_LOOKUP_TABLE survive only in the SDK's own v1 models — SDK-side legacy retirement, not an upstream dependency.
• "Invisible to users until the pins move off the 0.x line." True when written; the pin bump is now part of leg 1 rather than a follow-up, so nothing about this work is deferred behind a separate pin-moving issue.
• "May need promotion to its own epic." Epic fixing caller method #35 keeps coordinating (spec D5). Per-leg work items are filed at pickup.

Caveats and known risks

• The contract this models is not frozen: Extract and freeze the canonical schema v2 contract from codeanalyzer-python .github#36 is open and codeanalyzer-schema holds only a README. The SDK models what the three analyzers actually emit and records every divergence rather than blocking (spec D1) — which means today's divergence is absorbed into the SDK's view layer until COBOL support #36 retires it.
• Removing _CALLABLES_LOOKUP_TABLE is behavioural, not a refactor: today a miss silently fabricates a callable with is_implicit=True and -1 sentinels rather than failing, so consumers may depend on edges resolving to synthetic nodes. That lands in leg 3 (python-sdk: Java models → schema v2 (two-layer views) #310), not here.
• Four of the nine registered divergences are parity-clause breaches on the Python side (config_reads_unresolved vs config_reads, base_classes vs base_types/interfaces, attributes vs fields, and v1 line fields retained alongside span). The SDK absorbs them; each is a line item COBOL support #36 can retire.

Definition of done

• cldk/models/cpg/ exists and no language model package redefines a spine field.
• The Python facade loads codeanalyzer-python 1.4.0 output, and every public accessor's name, signature and return type is unchanged — proven by the existing Python facade tests running unmodified.
• Both projections agree: analysis.json and the Python Neo4j backend return the same facts for the same run.
• pyproject.toml pins codeanalyzer-python==1.4.0.
• The divergence register is filed on .github#36.
• 2.0.0-rc.1 cut; Incorporate codeanalyzer-python v1.2.0 #315, Incorporate codeanalyzer-python v1.3.0 #316, Incorporate codeanalyzer-python v1.4.0 #317 and Python get_method_bodies lacks the None-code omission filter (twin of the TS #298 rule) #301 closed.