[kvlonge]

Why

A runtime not_null audit verifies data after execution; it does not declare that a model's output schema is non-nullable. Treating the audit as schema metadata caused false K004 diagnostics whenever static analysis correctly proved that the SQL expression could return null.

Changes

• Make K004 depend only on explicit nullable false contract metadata.
• Preserve runtime audit attachment, severity, and execution behavior.
• Add unit coverage for audit-only and explicit-nullability declarations.
• Add real CLI integration coverage proving audit-only nullable output compiles while nullable false still fails with K004.

Verification

• uv run pytest tests/unit/src/sqlbuild/compiler/contracts/test_validate.py tests/integration/src/sqlbuild/cli/commands/main/test_compile.py -n auto --dist loadfile
• Focused Ruff formatting/lint and Ty checks
• uv run fensu check
• uv sync --all-extras
• make check-ci
• Independent read-only diff review: no findings