Hardening: make arrow-precedence preprocessing robust to literals/comments

[ndenev]

Summary

fix_arrow_precedence performs regex-based rewriting over raw SQL text. This works for common cases but is structurally brittle because it is not token/AST-aware.

Affected code

• src/datafusion_integration/preprocess.rs (fix_arrow_precedence, LEFT_ARROW_PATTERN, RIGHT_ARROW_PATTERN)

Current status

• Existing tests cover many happy/idempotency paths.
• Quick live checks did not reproduce literal/comment corruption in simple probes.
• This is filed as hardening (low priority), not a confirmed semantic bug.

Why track this

Regex-over-raw-SQL transforms can accidentally match inside contexts that should be inert (string literals/comments) or miss edge syntax forms. Even if currently okay, this is fragile and benefits from explicit guardrails.

Proposed improvements

1. Add explicit regression tests for:
   • arrow-like text in single-quoted literals
   • arrow-like text in -- and /* ... */ comments
   • unusual whitespace/operator layouts
2. If edge breakage appears, migrate to tokenizer/AST-aware rewriting for precedence wrapping.
3. Keep idempotency guarantees.

Related issue

Related selector-side robustness problem (validated in live run) is tracked here:

• Validate selector representability before label/field pushdown #72

That issue is about selector representability/validation; this issue is about SQL preprocessing robustness.