fix(tests): use sys.executable in doc-accuracy subprocess calls (unblocks the whole improve/goal pipeline)

[ProtocolWarden]

Summary

Root-caused via the failure-diagnostics capture added in #345. Every improve/goal task was failing N of N stages failed because the executor's baseline-validation gate (.venv/bin/pytest -q, run on the fresh workspace clone) exited 1 — and the single failing test was tests/unit/test_documentation_accuracy.py::test_marker_configuration_in_pytest.

Root cause

That test (and 6 siblings) shells out with bare ["python", "-m", "pytest", ...]. In the workspace environment, python on PATH has no pytest, so stdout is empty → assert 'integration' in ''. CI passes only because there bare python is the pytest-equipped interpreter. That one failure exited pytest 1 → baseline validation failed → the executor refused to proceed on every task.

Fix

import sys + replace all 7 bare-python subprocess calls with sys.executable (the interpreter actually running the test, which always has pytest).

Verification

• The file: 48/48 pass (was 1 failing).
• Full tree: 9377 passed / 0 failed → baseline validation now passes, unblocking the pipeline.
• ruff/ty clean; pre-push audit 0 findings.

Must land in main (baseline validation clones main). Closes the recurring #264/#265 failures.