Class-attribute literal values are dropped; Odoo model identity/composition (_name/_inherit) unrepresentable (CLDK-004)

[rahlk]

Target: canonical schema v2 (epic #69), with the model-composition layer routed through design first — it is new contract surface. From the external v0.3.0/Odoo quality audit, finding CLDK-004 (severity Critical).

Audit evidence (Odoo, native export)

• 1,235 _name and 2,663 _inherit class attributes exist as :PyAttribute nodes — with zero captured literal values (no value/initializer property).
• Zero model-composition relationship types exist in the schema (nothing matching ODOO/MODEL/INHERIT).

So _name = 'account.account' is present as a node named _name whose defining string is unrecoverable from the graph, and Odoo's registry-based model composition (multiple classes contributing to one model via _inherit) has no representation at all.

Root cause — two independent gaps

1. Class-attribute values are dropped at capture. PyClassAttribute (codeanalyzer/schema/py_schema.py:293-300) has no value/initializer field, and _class_attributes (symbol_table_builder.py:356-432) extracts only the name and a Jedi-inferred type, never the assignment's RHS. This is an asymmetry: PyVariableDeclaration (module/local variables) does keep initializer.
2. No framework-semantics layer exists. The analyzer is fully framework-agnostic; the only inheritance edge is PY_EXTENDS over lexical base classes, which cannot express registry composition declared through _inherit strings.

Fix direction

• v2 schema: give class attributes their initializer (general fix, benefits every consumer — the audit found the missing values via _name/_inherit but the drop is universal).
• Design-level (via designing-cldk-changes): a model identity/composition layer so literal registry lookups can traverse contributor classes — every literal model identity and contributor class represented, including multi-class composition (audit acceptance gate). This is also the second half of Callee inference anchors at call-expression start: attribute calls resolve to receiver type, 99.9% of Odoo env calls bind to enclosing class (CLDK-001 + CLDK-002) #80: after the callee-anchoring fix, env['model'].search(...) dispatch still needs this layer to resolve to the right model implementation.

[rahlk]

Status + design sketch for the remaining half of this issue

Landed (v1 batch, feat/audit-remediation-v1): class-attribute initializer capture. _name = 'account.account' / _inherit = ['mail.thread'] are now recoverable from both analysis.json (PyClassAttribute.initializer) and the graph (PyAttribute.initializer), which satisfies the audit's identity gate: a consumer can query every model identity and every _inherit contribution as literals.

Still open (this issue stays open for it): the model-composition layer — making registry dispatch traversable without consumers re-implementing Odoo semantics in Cypher.

Proposed overlay design (for a future designing-cldk-changes round)

• Opt-in framework pass, e.g. canpy --framework odoo, running after the symbol table is built. The core analyzer stays framework-agnostic; overlays are additive passes that only read captured facts (the attribute initializers).
• Derivation, literal-only: evaluate _name / _inherit / _inherits initializers when they are string or list-of-string literals (ast.literal_eval on the captured text). Dynamic values stay underived — recorded with an unresolved_reason rather than guessed.
• Proposed graph shape (additive, MINOR bump when built):
  • :PyOdooModel {name} — one identity node per registry model name.
  • PY_DEFINES_MODEL (:PyClass → :PyOdooModel) from _name.
  • PY_EXTENDS_MODEL (:PyClass → :PyOdooModel) from each _inherit entry.
  • Model composition = all contributor classes of a model, reachable in one hop; env['account.account'].search(...) dispatch = contributors of account.account filtered by a search method (PY_HAS_METHOD), MRO ordering left to the consumer or a later pass.
• JSON mirror: a top-level frameworks.odoo.models map (model name → {defining_classes, extending_classes}), so both projections stay in lockstep.
• Non-goals for the first cut: _inherits delegation semantics, _auto/abstract flags, addon manifest awareness, cross-addon override ordering. Each can layer on later without changing the two edge types.

The audit acceptance gate this closes: "Every literal model identity and contributor class is represented, including multi-class model composition used by registry dispatch." Related: the callee-binding fix (#80/#89) already stopped env[...] calls from wrongly binding to the caller's class — with this overlay they become correctly resolvable instead of unresolved.

[rahlk]

Progress: part 1 (initializer capture) shipped in v1.0.0 — PyClassAttribute.initializer captures the assignment RHS (landed via the #95 audit batch and carried into the v2 tree in the v0.3.1→v2 reconciliation merge; test_class_attribute_initializers_are_captured runs on the v2 suite). _name = 'account.account' is now recoverable from both projections (PyAttribute.initializer in Neo4j).

Part 2 — the model identity/composition layer (registry-based _inherit composition, env['model'] dispatch) — remains open and is design-rung scope (designing-cldk-changes): it is new cross-language contract surface, per this issue's own routing note. Keeping this issue open to track it.