Dbt Agent Readiness

Three checks that previously relied on LLM judgment during the deep pass are now computed deterministically by the inventory script.

• New catalogs.undefined_column_refs: per model, every SELECT scope (outer query and each CTE) is resolved against its input relations (CTEs recursively to depth 10, ref'd models through their extracted column lists); any column referenced in SELECT or GROUP BY that no input produces is flagged with confidence: 'high'. Always a Blocker candidate, ranked with broken refs. Conservative by construction: scopes are skipped when any input is unresolvable (macro-generated columns without a compiled manifest, regex-fallback extractions such as incremental-model tails, sources, subqueries). Local CTEs now correctly shadow same-named models during resolution.
• New catalogs.fan_out_joins: models joined by 2+ downstream models on a key with no unique test. Join targets are resolved directly or through grain-preserving passthrough CTEs; each row carries the join column, downstream models, a sample ON condition, and a runnable verification query. Hygiene candidate in synthesis.
• Fixed test_summary.models_with_zero_tests undercount: models with no YAML entry at all were not counted (messy-jaffle-shop reported 3, truth is 6). The summary now also enumerates the models in models_with_zero_tests_list.
• SKILL.md Step 5b and report-template.md wired for both new catalogs (Blocker collection, Hygiene verification queries, appendix tables).