[kvlonge]

Why

Under the preview engine, native reference extraction sent rejected calls (P012) back to Python and re-scanned to confirm P001 scan errors, which had no location.

Changes

• Native returns rejected calls with Python's exact P012 message, help, fix and start; only unsupported line-comment prefixes defer.
• Native P001 scan errors are raised directly, with no Python re-scan or cross-check.
• Behavior change, both engines: a model's P001 reference-scan error (unclosed parenthesis, quote or comment, empty argument) now starts with path:line:col:. The fault offset comes from the scan of the SQL that failed and is mapped back to the file through the body start and the declaration and macro expansion spans; nothing is searched. It is located, and raised with bridge_independent=True, only if it lies in authored text and the scan fails identically with every expansion output blanked. Otherwise, and for tests, audits, functions and @@var bodies (no map yet), the error names the file alone and keeps the bridge re-run. The type stays CompileInputError.
• Failure-corpus expectations are updated; no docs quote these messages.
• Tests:
  • Rust cases taken from Python.
  • Generated parity on all seven lexical syntaxes, including an escaped-quote header.
  • Source-map cases: header offset, two passes, a fault in macro output, an authored fault shaped by an unclosed macro quote.
  • E2E on all three engines: escaped-quote header, macro-output fault before an authored one (Python re-run kept), authored fault after a macro (no re-run), TEST/AUDIT/FUNCTION headers; lifecycle error_type CompileInputError.

Verification

• cargo test --workspace --all-features, clippy and fmt with Rust 1.99.0
• Focused reference, E2E compile and corpus-code tests
• make compiler-differential with CI flags, make compiler-differential-shipped, tests/e2e/scripts/compiler_differential
• make check-ci, make test, public-repository scan