[kvlonge]

Why

Raw-query diffs compared data exactly, but their evidence and automation surface were harder to consume than the comparison itself. Inputs were generically named, long values could overwhelm terminal output, JSON was file-only and model-shaped, unrelated project extensions could be imported, and automation could not distinguish incomplete evidence from execution failure.

Changes

• add optional left/right query labels and query-oriented comparison names
• load only project configuration and the selected adapter for raw-query mode
• report elapsed connection, reconciliation, inspection, materialization, comparison, cleanup, and total phases
• render bounded values around the first differing character with explicit length/truncation metadata
• add value suppression and explicit full-value opt-ins while retaining identifiable keys
• add JSON stdout with progress on stderr, sparse column mismatches, query provenance, timings, and strict JSON-safe evidence
• distinguish pass, findings, incomplete evidence, and execution failure with exit codes 0–3
• preserve model JSON fields and exact comparison, count, cleanup, and ownership behavior
• document and test the complete CLI contract

Verification

• make test — 6,717 passed
• focused DuckDB CLI, integration, and unit regressions passed
• uv sync --all-extras
• make check-ci
• independent review plus focused follow-up found both reported machine-output issues resolved
• public-repository scanner passed