feat(codegen): struct field access + container equality — TASK-29 Phase B (Tasks 3–4 of 4)

[ahrzb]

Summary

Completes Phase B of the codegen deferred surface: struct field access and struct/list equality. With Tasks 1–2 (merged in #4) this retires every container differential skip except UNNEST — the 5 remaining skips are all UNNEST, deliberately left for Phase C.

Plan: docs/superpowers/plans/2026-07-22-codegen-deferred-phase-b-containers.md (Tasks 3–4).

What's in this PR

Commit	What
bee2b39	Task 3 — struct field access. New FieldAccess IR node threaded through _convert_expr/infer_type/_validate_expr/_emit_expr + runtime.getfield. _convert_expr layers FieldAccess over multi-part columns (catalog/db/table qualifiers + leaf); the 2-part s.x stays a Column and is rewritten to FieldAccess in _validate_expr once schemas show the qualifier isn't a relation alias (mirrors native plan.rs). _validate_expr now returns the (possibly rewritten) node; every caller reassigns. Covers s.x, nested s.a.b, qualified t.s.x.
7a502a9	Task 4 — struct/list equality + container guard. runtime._veq (type-tagged deep equality: structs by key order+values, lists elementwise, scalars via val_eq so Int(1) != Float(1.0)); rt.eq/neq route containers through it. infer_type's BinaryOp arm checks containers first — eq/neq → BOOL, every other op defers — which also closes the Phase-A carryover where struct || x mistyped as STR.
795e3e9	Review fixes: fold an unquoted struct-column qualifier (S.x → s.x, matching DataFusion); comment the deliberate drop of the 4-part catalog raise (lets t.s.a.b resolve); add list-inequality coverage.

Testing

uv run pytest → 530 passed, 5 skipped (all UNNEST/Phase C), 5 xfailed. All field-access and equality shapes are value-asserted against the DataFusion oracle on both backends (native + codegen).

• Field access: s.x, nested s.a.b, qualified t.s.x (all pass on codegen).
• Equality: struct eq true/false (deep, type-tagged), list eq/ineq.
• xfail_on_native (codegen matches oracle, native has the gap): mixed-numeric list widening + the new uppercase-qualifier fold.

Scope limited to sql_transform/_codegen/ and tests/ — no native-engine changes.

Known limitations (out of Phase B's committed surface — flagged for tickets)

• Native parity gaps surfaced by the diff tests: native doesn't implement struct/make_array construction or fold struct-column qualifiers. Documented via xfail_on_native.
• struct = scalar returns False on codegen where DataFusion type-errors at planning.
• Field access on a NULL struct: DataFusion returns 0 (not NULL); neither engine matches. Worth its own investigation.

Review

Ran requesting-code-review on Tasks 3–4 (including independent vetting of the inherited Task 3 code) → ready to merge, no Critical/Important issues. The 795e3e9 fixes address the Minor findings; the remaining two are the noted out-of-scope limitations above.

🤖 Generated with Claude Code

[ahrzb]

/review

[claude]

Claude Code is working…

I'll analyze this and get back to you.

View job run