[kvlonge]

Why

Protect warm compilation and small edits with diverse SQL and shared dependencies, rather than relying only on repeated synthetic query shapes.

Changes

• Generate a neutral 1k all-rules workload with windows, arrays, grouping, lateral joins, unions and scoped Python macros.
• Require 950+ operator shapes independent of identifiers/literals, a 900+ model component, bounded depth and substantial leaf coverage.
• Check warm, leaf, shared-model and macro edits against uncached semantic/artifact oracles and pinned fingerprints.
• Add a 4GiB CI step and real DuckDB integration coverage.

Verification

• 64-model cold/warm comparison and generated SQL tests passed in DuckDB.
• Python 3.12 1k calibration passed all 11 scenarios: 999 operator shapes, depth 52, 298 leaves; 5.40s warm and 7.28–8.13s edits.
• Independent read-only review found no correctness issues.
• make check-ci and the public-repository hygiene scanner passed.