[kvlonge]

Why

M3i (native rendering as the default) needs the per-model loop's config work to run natively. After M3e/M3f the preview engine still built model config and ran the model config validators in Python for every model.

Changes

• sqlbuild-model-config: native path-default selection, Python-exact Duration/ISO cursor-bound parsing, header storage overrides, and the happy path of all ten model validators (references, incremental/microbatch, capability, contract, non-incremental, snapshot, custom materialization, storage, migration, placeholders). Anything uncertain (non-ASCII text, unusual ISO forms, big ints, odd types) rejects.
• Bindings: NativeModelConfigBuilder layers defaults, path defaults and headers, expands ${...} templates, applies target namespace overrides and storage policies, and returns the reads Python records; NativeModelValidator accepts or rejects.
• Preview MODEL_CONFIG stage: the loop uses both. A deferred build runs build_model_config; a rejection runs the Python validators for the exact error. Shipped/python engines unchanged.
• Python validator sequence moved to run_python_model_validators (same order).
• Stays Python: SQL-analysis gate/opt-out, Python-hook checks, enum contract columns and named usages (object assembly until M3j), cursor-intrinsic scan, reusable config cache. Hook expansion is already native via the macro bridge.

Verification

• Seeded parity: 6,000 configs (native accepts 1,607 of 1,610 Python accepts, never one Python rejects); 6,000 builds (equal config and reads or defer, never hides a Python error).
• 19 new failure-corpus validator cases; DuckDB build E2E on python/native/native-preview.
• cargo test --workspace --all-features; clippy/fmt 1.99.0.
• make compiler-differential (CI flags): 243/243; -shipped: 237/237.
• tests/e2e/scripts/compiler_differential, make check-ci, make test (11,833 passed), hygiene scan.