[kvlonge]

Why

SQBRDECLARATION102 previously authorized a compiled numeric comparison whenever the same numeric token appeared anywhere in the full authored model file. Header metadata, comments, casts, output values, and macro-expanded SQL could therefore create both false positives and false negatives.

Fixed output bucket projections such as CASE WHEN offset_seconds = 30 ... AS amount_30 are structural schema definitions rather than unnamed policy thresholds.

Changes

• pass the authored query body separately from the complete model file to native Rules
• match complete authored comparison expressions instead of model-wide numeric tokens
• ignore comments while matching authored comparisons
• inspect direct comparison operands so nested arithmetic and cast precision do not authorize a finding
• recognize equality-based fixed output buckets when the literal is encoded in the output name
• retain strict findings for threshold comparisons, including aliases that happen to end in the threshold
• add native regressions for constants, comments, casts, nested comparisons, and fixed buckets

Verification

• native Rust test suite
• focused Python Rule evaluation and cache tests
• Rust clippy with warnings denied
• uv sync --all-extras
• make check-ci
• downstream cache-free static compile across 974 models with the SQBRDECLARATION102 inventory reduced to zero after genuine thresholds were declared as constants
• public repository hygiene scan