feat(test): treat model file paths as test selectors

[tripleaceme]

Description

Closes #6023.

sqlmesh test already took multiple positional arguments, but they were only ever looked up in the test-file index. A model file matched nothing and the command exited 0, so a pre-commit hook passing staged model files silently ran no tests.

Each positional argument now resolves to one of:

• a test file → every test in it
• file.yaml::test_name → that one test
• a model file → the tests targeting that model

The results are unioned. Per your note on the issue, they are also deduplicated on fully_qualified_test_name, so overlapping selectors don't cause duplicate runs — sqlmesh test models/a.sql tests/test_a.yaml, where the YAML holds a's tests, runs each of them once, and tests/a.yaml::test_x tests/a.yaml collapses to one run too.

An argument resolving to neither a known test nor a known model file is now an error rather than a silent skip. That part is opt-in through raise_on_unknown_paths and only the CLI turns it on, because LSPContext.get_document_tests calls select_tests with whatever document happens to be open and depends on getting an empty list back — raising unconditionally would make the extension throw on every non-test file. A known model that simply has no tests still selects nothing and is not an error, since failing a hook because a model lacks a unit test isn't the intent here.

--select-model is unchanged: it still narrows the selection rather than adding to it.

Two notes on cost, since path matching is on the hot path for the pre-commit use case:

• Selectors are matched exactly as given before any normalization, so the existing absolute-path callers pay nothing new, and normalization is os.path.abspath rather than resolve() — no stat or symlink syscalls per model.
• The model-path index is built at most once per call, and only if some selector turns out not to be a test file. Passing 50 staged model files doesn't rebuild it 50 times.

Test Plan

Nine tests in tests/core/test_test.py:

• test_model_path_selects_its_tests / test_model_path_without_tests_selects_nothing
• test_model_and_test_paths_are_unioned_without_duplicates — the dedup case from your comment
• test_overlapping_yaml_and_named_test_are_deduplicated
• test_relative_paths_select_tests — pre-commit passes repo-relative paths
• test_unknown_path_is_ignored_by_default — guards the LSP's behavior
• test_unknown_path_errors_when_requested / test_unknown_test_name_errors_when_requested
• test_select_model_still_filters_path_selection

Two in tests/cli/test_cli.py covering the CLI wiring and the non-zero exit on an unknown path.

pytest tests/core/test_test.py tests/cli/test_cli.py tests/lsp
222 passed, 1 skipped

The 4 remaining failures in those files (test_pyspark_python_model, three test_dlt_*) are pre-existing on main in my environment — missing pyspark and dlt.

Checklist

• I have run make style and fixed any issues
• I have added tests for my changes (if applicable)
• All existing tests pass (make fast-test)
• My commits are signed off (git commit -s) per the DCO

[tripleaceme]

@cmgoffena13 — closes #6023, and implements the dedupe you described on the issue.

Selectors are unioned and then deduplicated on fully_qualified_test_name, so passing a model file and the test file holding its tests runs each test once rather than twice.

One decision worth your eye, described more fully in the PR body: making an unresolvable path a hard error is opt-in, because LSPContext.get_document_tests calls select_tests with whatever document is open and relies on an empty result. Only the CLI turns it on.

Zero checks here too, pending workflow approval.

[cmgoffena13]

Hey @tripleaceme -- would you please resolve the file conflicts. Just merged the --local flag PR, will tackle this one next.

[tripleaceme]

Done in ab4a951 — rebuilt on top of main now that #6060 is in, and it shows MERGEABLE again.

Both conflicts were additive rather than contradictory, so both sides are kept:

• tests/cli/test_cli.py — each branch appended its own tests to the end of the file.
• docs/concepts/tests.md — path selection and --local document different things. Selection comes first, then the --local section.

sqlmesh/cli/main.py and docs/reference/cli.md merged cleanly, but I checked them rather than trusting the auto-merge: the --local flag and its group-callback gating from #6060 are both still there, alongside raise_on_unknown_paths=True and the TESTS docstring from this branch. I also diffed the test section of the CLI reference page against real sqlmesh test --help output to confirm the two entries still mirror it.

One thing I added while resolving, since neither branch could cover it alone: test_test_local_with_model_paths pins that the two features compose.

$ sqlmesh test --local models/full_model.sql
.**Successfully Ran `1` Tests Against `duckdb`**

$ sqlmesh test --local models/nope.sql
Error: '.../models/nope.sql' is not a known model or test file.

That's the pre-commit shape from #6020 — --local for no state, a model path for selection, and a stale path still failing loudly rather than silently running nothing. Happy to drop that test from the merge commit if you'd rather keep the resolution purely mechanical.

pytest tests/cli/test_cli.py tests/core/test_test.py tests/lsp
230 passed

The test_dlt_* and test_pyspark_python_model failures are pre-existing on main in my environment (ModuleNotFoundError).

Unrelated and not touched here, but while diffing that page I noticed two entries in the test block still drift from the real --help: -v, --verbose is missing "Use -vv for very verbose output", and --select-model TEXT carries a "Can be specified multiple times" line the command doesn't print. Both predate these PRs. I left them out to keep this commit to the conflict resolution — happy to fix them separately if useful.