[codex] privatize swift query selection

[vvv214]

Summary

This draft PR makes pipeline SWIFT query selection use DP-noised error scores.

• split SWIFT budget requests into selection-score and measurement budgets
• add Gaussian noise to the vector of candidate errors before swift.select_queries
• use the measurement budget only for final selected marginal measurements
• avoid writing exact errors into diagnostic output
• add focused tests for score noising and invalid selection-budget fractions

Why

The pipeline path computed exact candidate marginals, converted them into exact error scores, and passed those errors into clique-tree/query selection. The budget request named Swift Select Queries was being used as a source for measurement allocation rather than privatizing the selection scores.

Validation

• python3 -m py_compile dpsynth/pipeline_transformations/swift.py tests/pipeline_transformations/swift_test.py
• rg -n "Swift Select Queries|compute_errors\\(|select_queries\\(" dpsynth/pipeline_transformations/swift.py tests/pipeline_transformations/swift_test.py
  • confirms the old budget name is gone
  • confirms exact errors are still computed as an intermediate but swift.select_queries is reached after _add_noise_to_errors

Full pytest was not run locally because this checkout's test dependencies (pipeline_dp, mbi, dp_accounting, Beam/JAX stack) are not installed in the current environment, and the local Python is 3.14 while the project declares <3.14.