Give every pwsh call in the suite the private startup cache - #10900

What this is, after a correction

I opened this against d49efcfe7 with a measured flake: tests/studio/install/test_installed_release_backend_line.py called pwsh through subprocess.run rather than through tests/_shared/unsloth_pwsh_runner.py, so it never got the per-worker startup cache, and at -n 16 it failed 1 run in 3, 14 cases at a time, returncode=-6 / Stack overflow.

That file has since been fixed on main by an equivalent change. Merging main in produced a conflict where upstream's side already routes it through run_pwsh; I took upstream's side. So the numbers this PR opened with describe a defect main has already closed, and they are not evidence for what is left here.

Re-measured on current main, same command, three runs:

2 failed, 7595 passed, 128 skipped   (234.68s)
2 failed, 7595 passed, 128 skipped   (231.61s)
2 failed, 7595 passed, 128 skipped   (233.77s)

No pwsh deaths, no host-fault skips. The two failures are the pre-existing tests/python/test_torchcodec_torch_compat.py pair on this box. So there is no failing test left for this PR to fix, and it should be judged as prevention plus a residue, not as a bug fix.

What is left, and why it is still worth having

Five call sites in five files still start PowerShell outside the private cache. They are in the same race, at lower volume than the file that was fixed: two subprocess.Popen holders in tests/python/test_windows_installer_concurrency_guard.py plus one subprocess.run there, one in tests/studio/test_installer_av_shapes.py, one in tests/studio/test_install_phase_timing.py, one in tests/python/test_pwsh_runner_encoding.py.

Three of those are skipif(os.name != "nt"), so they cannot show up on Linux at all, and one of the Linux ones — test_installer_av_shapes.py — converts an interpreter crash into a skip through its _PWSH_HOST_FAULT banner, which names System.IO.FileLoadException outright. That is the honest reason there is nothing to measure: that file has been quietly skipping past this bug rather than reporting it.

pwsh_env(env = None) in the runner returns env (default os.environ) with XDG_CACHE_HOME pointed at the existing per-worker cache dir; run_pwsh now calls it instead of doing the same six lines inline. Pure extract-method, behaviour unchanged, its own test file 27/27. It exists because three of those sites cannot become run_pwsh callers without changing what they test: a Popen holder written to over stdin while a second shell races it for a named mutex, a call site whose policy needs the crashed CompletedProcess back rather than an exception, and a deliberate control that must invoke pwsh the pre-fix way to show the fix changes something. Rewriting those around run_pwsh would bend the tests; handing them the cache directory removes them from the race and leaves their control flow alone.

tests/studio/test_pwsh_calls_use_the_shared_runner.py is the part that lasts. It walks the AST of every *.py under tests/ — not just test_*.py, so a conftest or a _shared helper cannot hide — resolves subprocess aliases and from-imports across run / Popen / call / check_call / check_output, and resolves the interpreter as a literal, as an .exe path, or through a module constant bound to one. A call passes if it goes through run_pwsh or passes env = pwsh_env(...), matched on the AST, so the exemption is a checkable property of the call and not a comment.

_ALLOWED_DIRECT_PWSH_CALLS mirrors _EXPECTED_CI_SKIPS in test_ci_shell_suite_coverage.py. One file is allowlisted, tests/test_windows_amd_gpu_scan_fallback.py (3 sites), because it hands its child a hermetic {"PATH": ..., "HOME": tmp_path} so XDG_CACHE_HOME already resolves inside tmp_path — which is why the runner's own header names it as the one pwsh-heavy file with zero failures in the CI run that prompted that module.

Without the guard, the fix that just landed on main is one subprocess.run away from being undone, in a file nobody would think to check.

Checks

• Guard: 5 tests, 5 passed, including self-tests for five detection shapes (literal, .exe path, from-import, Popen([PWSH, ...]), args= kwarg) and seven that must not flag.
• Not vacuous: inserting subprocess.run(["pwsh", "-NoProfile", "-Command", "Write-Output hi"]) into another test file fails it with tests/studio/test_ci_shell_suite_coverage.py:31: subprocess.run(...), naming file and line. Reverted after.
• The touched files re-run after formatting: 605 passed, 19 skipped.
• ruff check --select E9,F63,F7,F82 clean on all of them; the repo's full ruff check passes; scripts/run_ruff_format.py is a fixed point by md5 across two passes.

Not verified here: the three Windows sites are skipif(os.name != "nt") and never execute on Linux. Their change is reasoned and lint-clean, and only Windows CI can exercise it.