[kvlonge]

Why

Native compiler stages still re-ran the Python implementation after a failure, only to get Python's exact error. For declarations and scopes that re-run is pure overhead and, under the macro bridge, executes macros twice. This is the first error-family slice of moving those errors native.

Changes

• Declaration files: a native parse failure, read error or value error is raised directly; Python no longer re-parses the file to cross-check it. Files native defers still parse in Python.
• @enum/@const references: the native scan returns the references Python resolves plus the error that stops its walk (unclosed quote or block comment, malformed reference). The expansion raises Python's first error in walk order instead of handing the string back to Python, with shared messages.
• Exact errors stay CompileInputError (code P001, same text, same lifecycle error_type on every engine) and carry bridge_independent=True. This includes Python's own scan errors when the native scan defers.
• Macro bridge: the stage re-runs in Python only for failures after the bridge has scanned SQL that are not marked bridge_independent.
• Tests: native/Python parity on generated reference SQL, declaration files without the cross-check, bridge re-run rules, E2E compiles proving identical errors with macros run once (including a Python-scanned unclosed quote), the lifecycle error_type on all three engines, and five failure-corpus cases.

Verification

• cargo test --workspace --all-features, clippy and fmt with Rust 1.99.0
• Focused pytest for the macro bridge, model loop, authored SQL and declaration files, plus the E2E module
• make compiler-differential with CI flags (233 identical), make compiler-differential-shipped (227 identical), tests/e2e/scripts/compiler_differential
• make check-ci, make test, public-repository scan