[kvlonge]

Why

sqb diff only accepted compiled models, which forced users and agents to create temporary project resources for one-off warehouse comparisons. Raw query inputs should reuse the same deterministic diff engine without making fingerprints part of query execution state.

Changes

• add mutually exclusive model and raw-query input modes under sqb diff
• support inline SQL or UTF-8 files, active/explicit targets, composite --key overrides, and exact --unkeyed multiset comparison
• preserve sampling, tolerances, exclusions, schema/key preflight, bounded evidence, JSON output, and existing exit behavior
• materialize each query once in the target schema, publish immutable ownership fingerprints, clean current artifacts immediately, and reconcile only expired fingerprint-owned artifacts
• extend sqb janitor recovery while retaining untracked reserved-name relations for manual investigation
• document query diff usage and the controlled ownership boundary

Verification

• uv sync --all-extras
• make test (6,712 passed)
• targeted DuckDB CLI E2E tests for query and model diff behavior (14 passed)
• targeted unit/integration regressions (166 passed)
• make check-ci
• public-repository scanner
• bounded independent review plus focused follow-up verification