[kvlonge]

Why

The release PR's "Compare with the previous release" job failed before measuring: the previous release's benchmark generator ran with the head interpreter and only the baseline source on PYTHONPATH, so it imported the head sqlbuild package. Once head removed a constant the baseline generator still imports, generation failed with ImportError. The compile ratio guard had already been fixed for the same class of bug.

Changes

• Probe the baseline environment's interpreter for its package directories and put them on PYTHONPATH after the baseline source and ahead of the head packages, so sqlbuild, its native module and its runtime dependencies resolve to the baseline install. The head interpreter still supplies test-only generator dependencies (pytest) that the baseline runtime environment lacks. Applies to both the pristine and dense baseline generators.
• Fail clearly if the baseline environment uses a different Python minor version.
• Widen the workflow path filter to the other comparison helpers the job imports.

Verification

• New tests: a baseline generator importing a sqlbuild symbol absent from this checkout succeeds (pristine and dense); both fail without the fix.
• Ran the full comparison locally against baseline 0.139.0 with small model counts: baseline generation and every measured command complete; generator imports of sqlbuild, polyglot_sql, duckdb and scripts resolve to the baseline install/source.
• make check-ci, focused release/compile performance tests, public-repository scan.