Design note: a weighted module inside policyengine.py, replacing the microdf dependency #528

Tracking issue for docs/weighted-module-design.md, the decision record for replacing policyengine.py's microdf dependency with an internal policyengine.weighted module: pure numpy estimator kernels (shared with microcosm.frame.accounting), weights as an ordinary <entity>_weight column, and a closed df.weighted(...) accessor that exposes only weighted operations.

Context: the September 2026 review of microdf (PolicyEngine/microdf#315, #333) found that a pandas subclass cannot fail closed, and #527 found outputs/inequality.py reimplementing Gini and top shares with a different tie rule from microdf. microcosm-frame already declares itself microdf's successor.

The note fixes the three-layer design, pins every estimator definition against a named reference, and stages the migration: Stage 1 (module lands, inequality.py, decile_grouping.py and labor_supply_response.py move onto it), Stage 2 (MicroDataFrame leaves the public data model; major version), Stage 3 (engine return types; a separate decision).