[kvlonge]

Why

sqb janitor judged a relation stale by project membership and age, not by whether anything still reads it, so archiving could silently break readers. Separately, custom Rules dominated cold compiles with no per-Rule cost breakdown, --no-cache still reused Rule results, and SQBRPROJECT107 needed one entry per qualifier.

Changes

• Janitor usage check: adapter relation_usage returns latest read, readers, query count and the views a relation was read through, or "not checked" with a reason.
  • Snowflake ACCESS_HISTORY (direct and base objects) plus OBJECT_DEPENDENCIES view chains; BigQuery INFORMATION_SCHEMA.JOBS; Databricks system.access.table_lineage; ClickHouse system.query_log.
  • Missing access means "not checked" naming the grant; other errors fail.
• Preview In use section with via and a frozen flag; in-use count in the prompt; keep_in_use / --keep-in-use keep them as Kept (in use); in_use_window defaults to "7d"; additive --json fields.
• compile_timings adds per-Rule invocations, CPU and cache counts, plus monitor and guard CPU; existing keys unchanged.
• compile --no-cache turns the Rules cache off.
• The module-state monitor checks once per model (project Rules still after each invocation), so split-host findings match a single process.
• SQBRPROJECT107 entries accept a qualifier glob and a paths form.
• Docs.

Verification

• Custom Rules: synthetic 1,000-model project, 20 Rules, true cold compile: 100.8 G -> 72.6 G instructions (-28%).
• Janitor: per-adapter unit tests (SQL, parameters, chunking, names, errors, clone false positives); DuckDB E2E via a test usage adapter; ClickHouse E2E (testcontainers) with a view chain read by another user. Not run against live Snowflake, BigQuery or Databricks.
• Rules: E2E for timings, --no-cache, persistent vs undone module state across 1 and 2 hosts, globs and paths.
• Focused E2E and integration files on the combined branch, make check-ci.