[kvlonge]

Why

Header columns, audits and ${...} templates are parsed in Python for every model. This moves that work native behind the opt-in native-preview engine, byte-identical to Python or deferring to it.

Changes

• New sqlbuild-model-config crate: MODEL header columns and model audits, ${...} template expansion (if, eq, ne, coalesce, ENV:, CTX:, variables) with its environment and context reads, and template and macro presence scans. Anything else defers to Python.
• The model loop parses all header metadata in one native batch and expands config templates natively when the new model_config stage is enabled. It is registered at the preview tier, so the shipped native default is unchanged.
• Errors stay in Python: a native rejection re-runs the Python path, which raises its exact error; if Python succeeds instead, a mismatch error names SQLBUILD_COMPILER_ENGINE=python.
• Compiler differential: seeds generate templated targets and audited, typed header columns; render coverage requires three new kinds; the failure corpus adds seven model config errors. Generated seeds shift, so both shared-analysis seed tests move to seed 34 and now assert a clean compile and real reuse.
• Layering, path defaults, enum contract columns and validators stay in Python until the Python render stage is removed.

Verification

• Seeded parity: 80k generated headers and 240k template values over every resolver flag set, 0 mismatches.
• Engine-tier integration tests (python, native, native-preview), forced-rejection mismatch tests, and an E2E CLI compile comparing python with native-preview for success and error output.
• make compiler-differential with captures and coverage: 151 identical, render 66/66, discovery 66/66; make compiler-differential-shipped: 145 identical; dense 3k and a large private project identical.
• cargo test, clippy and fmt (1.98.0), make check-ci, make test (11495 passed), tests/e2e/scripts/compiler_differential.