Remove skip gate that blocks real-world crash test execution

The module-level skipif(not CLONE_DIR.exists()) causes this test to skip on a fresh checkout, so _ensure_repo() never runs and the repo is never cloned despite the test/docs saying it will. In practice this means CI and most local runs silently skip the regression baseline, so crash-rate regressions are not detected unless the directory is pre-created manually.

Useful? React with 👍 / 👎.

Add client mixin path to PR workflow trigger

The new workflow only runs on pull_request when files under the listed paths change, but src/fastmcp/client/mixins/tools.py is included for push and omitted for pull_request. That means a PR that changes this OpenAPI-related client path can merge without running the crash baseline, and regressions are only caught after merge when push runs on main. Include the same path in the PR filter so this gate works pre-merge.

Useful? React with 👍 / 👎.

Guard removal of arbitrary existing clone path

_ensure_repo() recursively deletes CLONE_DIR whenever it exists but is not already at the pinned commit, without validating that the path is actually an openapi-directory checkout. Since CLONE_DIR comes from OPENAPI_DIRECTORY_PATH, a misconfigured value (for example, a parent directory like /tmp) would cause unrelated files to be deleted during a test run. Please fail safely (or require explicit opt-in) instead of unconditional rmtree on any existing path.

Useful? React with 👍 / 👎.

Preserve pytest timeout alarm when resetting per-schema alarm

The per-schema cleanup calls signal.alarm(0), which cancels any existing process alarm, including the pytest.mark.timeout(1200) guard on POSIX runs that use SIGALRM. Because the code only restores the handler (not the prior timer), the global test timeout can be silently disabled after the first schema, so a later non-interruptible hang may run until the workflow/job timeout instead of failing at 1200s.

Useful? React with 👍 / 👎.

Defer provider parametrization until repo is ensured

The provider list is computed at collection time via @pytest.mark.parametrize("provider", _collect_providers()), but _ensure_repo() (which clones/checks out the pinned commit) only runs inside the test body. This means the parametrization can be built from the wrong checkout (or from no checkout at all), so in RUN_REAL_WORLD_SCHEMA_TEST=1 fresh-checkout runs pytest collects an empty provider set ([NOTSET]), and with a non-pinned local clone it can miss providers that exist in the pinned dataset. As a result, the per-provider timeout guard is silently skipped or incomplete even though the test is intended to give per-provider coverage/progress.

Useful? React with 👍 / 👎.

Skip local run when clone path exists but is not repo

_skip_locally only checks CLONE_DIR.exists(), so any existing path at /tmp/openapi-directory (including an unrelated directory or file) disables the skip and runs the test, after which _ensure_repo() raises RuntimeError because the path is not an openapi-directory clone. This turns a test documented as “skipped by default unless cloned” into a hard failure in normal local pytest runs when that path happens to exist, which is a common /tmp collision case.

Useful? React with 👍 / 👎.

Make aggregate crash baseline xdist-safe

This module aggregates provider results in the process-local _results dict, so when the test is run with pytest -n auto (xdist), each worker builds a separate partial view and test_z_aggregate_crash_rate only sees whatever ran on that single worker. In practice this can either skip the aggregate check (_results empty on that worker) or fail the >200_000 assertion with partial totals, so the regression baseline is not reliable under parallel execution.

Useful? React with 👍 / 👎.

Normalize required entries when coercing YAML keys

Converting every dict key to str here changes property names (e.g. YAML 200: becomes '200') but list scalars are left as-is, so required: [200] remains an int. In _create_dataclass, required detection uses prop_name in required, so '200' no longer matches 200 and a required field is silently treated as optional. This weakens validation for YAML-loaded schemas with unquoted numeric keys instead of preserving their JSON semantics.

Useful? React with 👍 / 👎.