fix(schema_loader): quote case-sensitive column names in external models

[devtechedge]

sqlmesh create_external_models serializes column names from information_schema verbatim, while the table name is already dialect-serialized. On case-folding dialects like postgres, a column created as "ID" was written as the unquoted key ID. On load, normalize_identifiers folds that key to id, so the generated external_models.yaml no longer describes the table and plan/lint fail with ambiguousorinvalidcolumn.

This adds a _serialize_column_name helper in sqlmesh/core/schema_loader.py that quotes a column name only when the dialect's identifier normalization would rewrite its unquoted form. Identifiers that normalize to themselves (for example billno or the _sync_row_hash pseudo column) and names the adapter already returned quoted are written unchanged, so existing projects' generated files are not churned.

Quoting on the write side is lossless: _columns_validator in sqlmesh/core/model/meta.py passes the quoted key through normalize_identifiers, which preserves the quoted form, so the round trip restores the original case.

Tests:

• New test_create_external_models_quotes_case_sensitive_columns covers uppercase columns (ID, ORGANID, CDATE) emitted quoted, a lowercase column and the _sync_row_hash pseudo column left unquoted, and asserts the quoted keys round-trip through create_external_model with their original case.
• The full tests/core/test_schema_loader.py suite passes (9 tests).

Fixes #6058

[devtechedge]

Hi - polite check-in on the external-models quoting fix whenever you have bandwidth.

sqlmesh create_external_models was writing case-sensitive column names from information_schema as bare keys, so on postgres a column created as "ID" became unquoted ID, then normalize_identifiers folded it to id and plan/lint failed with ambiguousorinvalidcolumn (Fixes #6058).

This adds _serialize_column_name in sqlmesh/core/schema_loader.py so a column is quoted only when the dialect's identifier normalization would rewrite its unquoted form; names that already normalize to themselves, and names the adapter returned quoted, stay unchanged to avoid churning existing external_models.yaml files.

Quoting on the write side stays lossless because _columns_validator in sqlmesh/core/model/meta.py passes the quoted key through normalize_identifiers, which preserves the quoted form on the round trip.

test_create_external_models_quotes_case_sensitive_columns in tests/core/test_schema_loader.py covers uppercase columns emitted quoted, a lowercase column and _sync_row_hash left unquoted, and the round trip through create_external_model; the full schema_loader suite (9 tests) passed locally.

Happy to tweak the helper, adjust which dialects quote, or add another fixture if that would help review.

No rush at all - still ready from my end whenever it fits.