What

An OnDemandFeatureView's UDF (pandas and python modes) was receiving every feature column in the online response, including features from feature views that the ODFV did not list in its sources. This allowed a UDF to silently depend on an undeclared source.

This filters the UDF input down to the ODFV's declared sources before the transform runs. Join keys, request-source fields and the declared features are kept; only columns from undeclared feature views are removed. substrait already restricts its inputs through its query plan, so it is left as is.

Fixes #6158.

Tests

Three tests were added in test_on_demand_pandas_transformation.py:

• pandas mode: the UDF sees its declared feature and the entity join key, but not another requested feature view's column.
• python mode: the same check with the dict input.
• two ODFVs: one ODFV's UDF does not see the other ODFV's source column, even when that feature view is present only as the other ODFV's input.

Each of these fails on the current code and passes with this change.

Functional test results:

$ pytest tests/unit/test_on_demand_pandas_transformation.py \
    -k "undeclared or other_requested_odfv"
3 passed in 0.52s

ruff and mypy are clean.

Scope

This covers the online path (get_online_features), which is what the issue reports. The offline path (get_historical_features) has the same shape and can be addressed as a follow-up.