feat(format): add transpile_meta opt-in for dialect-aware MODEL headers

[albertosuman-1k5]

Summary

• Adds opt-in format.transpile_meta so sqlmesh format can render MODEL/AUDIT/METRIC headers with the model dialect again.
• Default remains dialect-agnostic headers (fix for TSQL dialect formatting corrupts boolean true in MODEL header to (1=1) #5773 / fix(format): keep MODEL/AUDIT/METRIC header dialect-agnostic (#5773) #5864): SQLMesh boolean properties stay as TRUE/FALSE instead of dialect rewrites like T-SQL (1 = 1).
• Projects that author headers in warehouse dialect (column types, audits, labels, macros, etc.) can set format.transpile_meta: true to preserve those values.

Context

#5864 stopped transpiling meta expressions to avoid corrupting SQLMesh DDL booleans. That also rewrote dialect-specific header values (for example DATETIME2(6) → TIMESTAMP(6)). A surgical per-property approach (see #5926 for columns) would need similar handling for audits, labels, macros, and other header properties; this flag restores the previous whole-header behavior as an explicit opt-in instead.

Trade-off: enabling the flag reopens the T-SQL boolean rewrite from #5773 for SQLMesh properties in the header. Documented in docs/reference/configuration.md.

Test plan

• Unit tests for format_model_expressions(..., transpile_meta=True) (multi-expression and single-meta paths)
• Config test that transpile_meta is excluded from generator_options (does not leak into SQLGlot Generator)
• End-to-end Context.format() with FormatConfig(transpile_meta=True) preserves DATETIME2(6)
• Manual: project with T-SQL headers containing columns / audits / macros — format with and without the flag

Made with Cursor

[fresioAS]

@albertosuman-1k5
I'm not sure I understand the rationale here? We changed all our T-SQL models over to dialect agnostic headers - the main reason being that (1=1) was not respected and we got into issues with everything from disable_restatement (1=0) to formatting (1=0) etc. In addition we had fields like description being [] instead of ''.

Based on this I don't see who format.transpile_meta: true would be useful for?

On the other side - it is a pain now when the formatter wants to change our columns to something that does not exist - hence the proposed PR #5926

[mday-io]

@albertosuman-1k5 thank you very much for proposing this. I do worry that this all-or-nothing approach will continue to cause people headaches. I believe #5926 by @fresioAS (with a few additions) might get us a quick-fix and set a foundation for a more scaleable solution.