[kvlonge]

Why

CI on branch pushes runs every unit test and scenario, even when a change can only affect a few. --since <ref> runs only what the change can affect. When in doubt it runs more, never less.

Changes

• sqb test --since <ref> and sqb scenario test --since <ref>. The base is compiled from git objects at the merge base, in a separate interpreter. It is cached under target/since/ (newest few kept) and shared by both commands. @base reads [git] base_ref.
• Each node gets a definition fingerprint, so formatting and description edits select nothing. The fingerprint covers:
  • model SQL, resolved config, columns, audits and meta
  • @@var and @@ENV: values
  • macros with the macros and project modules they import
  • enums, constants, audits, seeds, sources, functions and Python nodes
• A test or scenario is selected when one of these changed:
  • its own file
  • anything it executes, including every case of a parameterized test
  • the declared columns of a relation it mocks
  • a project-wide input: settings, connection and targets, adapter modules, or Python lock files
• migrate_from renames compare with the old definition. Deleted nodes are listed and select nothing.
• --explain and --inspect show the reasons. --json adds a since section.
• Clear errors (C480–C485) with no tracebacks. A target missing at the base falls back to the default target.
• Docs, including known limits, and the docs skill are updated.

Verification

• E2E test_since.py has 39 cases, including:
  • variables, defaults, macro imports, parameterized cases, mocked column shapes, connection settings, and tracked vs untracked lock files
  • scenarios, JSON output, renames, shallow clones, a broken base, and a target missing at the base
• Rust tests cover change detection, import closures and dependency files.
• Existing test, scenario and custom-rule tests pass.
• make check-ci passes.