[kvlonge]

Why

The preview engine re-ran Python whenever a native model config stage failed, only to get the exact error. That kept a Python fallback on every error path and could run user macros twice. Native stages now report Python's exact errors themselves.

Changes

• Native config build, model validators, MODEL header columns/audits, ${...} templates and path-default selection return the error Python raises: same class, code, message and help, in the same order. Errors from stages that never read SQL (config build, templates, header metadata, path defaults) raise as CompileInputError(..., bridge_independent=True). Validator errors are not marked, because some read macro-expanded SQL; the macro bridge decides whether to re-run them. Other errors raise as ResourceIdentityError, DiscoveryConflictError or ConfigValueTypeError where Python uses those.
• Removed the NativeStageMismatchError paths for header metadata and templates, the path-default re-run, and the validator re-run on errors. The Python validators now run only when native defers.
• Native still defers inputs that need Python's exact rules: non-ASCII or unreadable text, very large integers, unusual ISO timestamp forms, sub-second cursor-bound ordering, Decimal special forms, dbt references, audit thresholds, non-string metadata keys and unhashable values.

Verification

• Parity tests against Python with generated multi-fault configs (validators and builds); new failure-corpus cases for templates, header metadata, references and several faults; and an E2E test showing preview errors match Python with no Python fallback and macros run once; a reference error from macro SQL is re-run by the bridge.
• cargo test, clippy/fmt on 1.99.0, make check-ci, make test, both compiler differentials with coverage, and the differential E2E suite.