v2 Stage 3: L3 intraprocedural CFG/CDG/DDG (syntactic)

[rahlk]

Part of #69.

Stage 3 — L3 intraprocedural CFG/CDG/DDG (syntactic)

Branch: feat/v2-stage3-l3

Deliverables:

• dataflow/identity.py — the (signature, int node_id) ↔ (can:// callable-id, ordinal-id) bijection; consumed by both emitters.
• Populate body with statement nodes + @entry/@exit; emit cfg/cdg/ddg edge lists on each callable, re-targeting the existing cfg/dominance/defuse kernels through the identity map.
• Split the current PDG rows into cfg/cdg/ddg; add a syntactic-only DDG mode (textual interference, oracle bypassed, prov:["ssa"]).
• CLI: --graphs cfg,dfg,pdg scoping; --graph-field-depth at L3+; sdg requires -a 4 (flag error otherwise).
• Neo4j: PyCFGNode keyed by ordinal id; PY_DDG gains prov.
• Migrate test/test_dataflow_{cfg,dominance,defuse,pdg,scc,cpg,slicing}.py to v2 ids.

Files: dataflow/{identity,builder,defuse}.py, core.py, __main__.py, neo4j/project.py, dataflow tests.

Gate: intraprocedural backward slice == hand-computed set exactly (over ordinal ids); assert_conformant(..., 3) incl. no dangling cfg/cdg/ddg; L2 ⊆ L3.

Expand into full bite-sized tasks with superpowers:writing-plans at stage start.

[rahlk]

Two forward-risks surfaced by the Stage 1 whole-branch review that Stage 3 must handle:

1. Cross-level cache body-leak (correctness). The analysis cache stores the full Analysis (bodies + edge lists) and _file_unchanged keys only on content hash / mtime / size — not on analysis_level, and populate_l1_body/assign_ids add to body but never clear it. Harmless at L1 (only call nodes populate body), but once Stage 3 fills body with statements + @entry/@exit + cfg/cdg/ddg, a subsequent lower-level run reusing the cache would leak higher-level content into lower-level output — breaking level purity / the L1⊆…⊆L4 monotonicity gate. Fix: key the cache by analysis_level (or clear regenerated body/edge collections on reuse).
2. Re-enable the Neo4j CPG overlay (_project_program_graphs is currently disabled + dead-references the dissolved program_graphs field) on the v2 tree, and close the latent dual :PySymbol constraint (id + signature) before Neo4j is exercised.

[rahlk]

Additional Stage-3 tracking (from the Stage 2 whole-branch review): when re-enabling the Neo4j projection, reconcile a classification-precedence asymmetry — reidentify_call_graph (JSON) resolves declared-first (sig_to_id wins) while _call_endpoint (Neo4j) resolves external-first (external_symbols wins). Inert today (a symbol is either project-declared or external, never both), but make them consistent so JSON and Neo4j can't disagree on an endpoint id if that invariant ever weakens. Track alongside the latent dual :PySymbol (id + signature) uniqueness constraint.