fix!: pass macros to the last audit in load_multiple_audits

[markussagens]

Description

Fixes #6119.

load_multiple_audits passed module_path, macros and jinja_macros to load_audit inside its loop, but not in the final yield load_audit(...). So the last AUDIT in a file (the only one, for a one-audit file) was loaded with an empty python_env and an empty Jinja macro registry. Rendering a standalone audit that used a project macro then failed with MacroEvalError: Macro '<name>' does not exist.

The final call now passes the same three arguments as the call inside the loop. The other arguments already matched.

Fingerprint impact (!): an affected standalone audit now stores the macros it references, so its metadata hash and fingerprint change. After upgrading, sqlmesh plan will list it as a metadata-only change. These audits failed to render before, so nothing that previously worked changes behavior.

Test Plan

Three regression tests in tests/core/test_audit.py. All three fail without the fix and pass with it:

• test_load_multiple_single_standalone_with_macros: one-audit file using a project Python macro.
• test_load_multiple_last_standalone_with_macros: two-audit file where only the last audit uses the macro.
• test_load_multiple_last_standalone_with_jinja_macros: last/only audit using a project Jinja macro.

pytest tests/core/test_audit.py passes (46 tests). Also ran the issue's reproduction project end to end: Context(...).standalone_audits["only_allowed"].render_audit_query() now renders.

Checklist

• I have run make style and fixed any issues (ran ruff check, ruff format and mypy on the changed files)
• I have added tests for my changes (if applicable)
• All existing tests pass (make fast-test) (only tests/core/test_audit.py was run locally)
• My commits are signed off (git commit -s) per the DCO

🤖 Generated with Claude Code

[mday-io]

Thanks for the clear write-up and repro, this looks good.

• The fix is right: the final load_audit(...) call now gets the same module_path / macros / jinja_macros as the call inside the loop, so the last (or only) standalone audit in a file gets project macros like the others.
• I checked out the branch and ran the new tests against the old definition.py: all three fail there and pass with the fix. I also confirmed end to end that a one-audit file using a project Python macro now renders.
• On the !: agreed. python_env feeds into the standalone audit's metadata_hash, so affected audits show up as a metadata-only change on the next plan. There's no backfill and no migration, and only audits that couldn't render before are affected. This matches what we did in #5758.

Approving.

[mday-io]

for context of release notes: Standalone audits that are the last (or only) AUDIT in their file can now use project Python and Jinja macros. Affected audits show up as a metadata-only change on the next sqlmesh plan (#6120).