Specializer raw-dict output mode — output='dict' opt-in, python_dict gap closed to 0.74-1.11x (TASK-51)

[claude-agent-ahrzb]

Raw-dict output mode — output="dict" opt-in closes the python_dict gap (TASK-51)

The serving report's standing caveat was that a handcrafted server returning plain dicts beat the specializer's typed output by 1.3–2×, because pydantic model construction dominated our cost. This ships the opt-in lever:

fn = DuckDBInferFn(sql, row_tables=..., static_tables=...)              # typed models (unchanged default)
fn = DuckDBInferFn(sql, row_tables=..., static_tables=..., output="dict")  # plain dicts

Design

• The generated marshaller already builds the exact per-row dict the model would consume — dict mode returns it and skips construction, so the modes agree field-for-field by construction. The serving parity gate now enforces dict == typed before any timing.
• Strictly opt-in: the typed default is untouched; output="dict" is mutually exclusive with a supplied output_model; unknown values reject; an output getter exposes the mode.
• All four emit paths honor it: generated marshaller, generic env-pinned baseline, the reentrancy fallback, and the constant engine (which copies per call so caller mutation can't leak into the next response).

Measured (n=1024 p50, 4 Kaggle scenarios)

• 25–35% off the typed path.
• vs the handcrafted python_dict floor: 0.74–1.11× — spec_dict outright wins house_prices and sits within ~35% elsewhere, down from the 1.3–2× caveat. The remaining gap is input marshalling, not output.

Gate green (759 passed + 13 xfail); 6 new tests + the parity-gate extension + a spec_dict bench engine row. TASK-50's Done-flip rides this branch.

🤖 Generated with Claude Code