[kvlonge]

Why

sqb format could not back a pre-commit fixer plus a --check CI gate:

• An uppercase SCHEMA (type ...) column inside MODEL(... columns (...)) was detected as a second SCHEMA() header. Format spliced in a duplicated header and the file stopped compiling.
• Snowflake FROM VALUES (...) was rewritten to FROM (VALUES ...), which SQBRSQL039 rejects, so compile failed.
• --exclude without --select dropped SQL tests from the format scope, and path:tests could not be combined with --exclude.
• A missing description made format and --check exit 1 even when nothing needed formatting.
• Some typed-null test fixtures needed two sqb format runs to settle, so format followed by --check failed.

Changes

• Header scanning ignores candidates inside an accepted header. Authored and rewritten headers are parsed before any write, and a file whose header fails to parse is left untouched with a header-parse fault.
• The native formatter keeps each VALUES relation's authored form. Wrappers are matched per occurrence.
• Format selection: the default scope is every formatter-owned SQL file minus exclusions. Path and model selectors combine, and --select <model> still formats only that model.
• In format, description-present is a warning. Exit status depends on faults and on pending changes for --check. Lint and Rules are unchanged.
• Formatter declines (unsupported SQL, comment attachments) are silent body skips again, and header formatting still applies.
• Fixture simplification also runs on the canonical formatted body, so one sqb format reaches a fixed point.

Verification

• Real-CLI integration tests for each case, with idempotency checks.
• Rust tests for mixed and nested VALUES, strings and comments.
• make test (6,940 passed) and make check-ci.
• Validated privately on a large Snowflake project: one format, then --check exits 0, compile passes, and SQL test results are unchanged.