fix(tests): make the suite hermetic against real user state

[Mokuroh54]

Why

The suite reads and writes the developer's real ~/.cache/huggingface/ in several places. On a clean CI runner that is invisible, which is exactly why it survived: every one of these is green in CI and only misbehaves on a machine that has actually been used.

1. The unprotected from-import path (the one with a hardware cost)

tmp_lerobot_home patched the path constants on makermodslab.utils.config. That reaches every consumer reading them through the module — but not a from .utils.config import FOLLOWER_CONFIG_PATH, which binds a copy of the string at import time. Rebinding the name on cfg afterwards leaves that copy pointing at the real ~/.cache.

Seven such copies exist (auto_calibrate ×3, arm_identity ×2, server ×2) and nothing in the suite redirected any of them.

The concrete hazard verified on this branch: the calibration-config DELETE endpoint (server.py) builds its os.remove() target from server.py's own copy of FOLLOWER_CONFIG_PATH. No test repoints it — tests/test_server.py repoints only LEADER_CONFIG_PATH, and says so in a comment — while the developer's real calibration/robots/so_follower/home.json sits at the other end. Recovering that file costs a physical recalibration.

2. Real pins / hidden lists leaking into listing assertions

tests/test_models.py fails 7 tests on an unmodified main on a populated machine: the model listing folds in the real saved_custom_models.json, so a listing asserted to be ["imported_policy"] also carries the developer's real pins (lerobot/smolvla_base, …). The dataset listing leaks the same way in both directions — real pins added, real hidden_datasets.json entries silently subtracting a row a test just seeded. tmp_lerobot_home never covered these four files, and listing tests that don't request it got no redirection at all.

3. Real HF identity

cached_whoami() calls get_token(), so an unmocked test resolves the developer's real username and orgs — which decide default repo owners and which Hub rows survive a listing.

What changed

Guardrails

• _FROM_IMPORTED_PATH_CONSTANTS — the seven from-imported copies, each repointed on its own module inside tmp_lerobot_home.
• The fixture's guard widened into a REFUSING TO RUN probe over every config constant and every from-imported copy, so an unpatched path fails at fixture time instead of being discovered by a write.
• A session-scoped autouse calibration canary fingerprinting the real calibration/** tree (mtime+size+sha1, .DS_Store ignored, absent-dir safe for CI) that fails the session loudly if anything changed.
• One consolidated autouse isolation floor (_isolate_real_user_state) covering all six redirects: the four SAVED_CUSTOM_*/SAVED_HIDDEN_* files, HF_LEROBOT_HOME, and huggingface_hub.constants.HF_HUB_CACHE. Plain monkeypatches, so tmp_lerobot_home and per-test overrides still win — a floor, not a wall.
• _stub_cached_whoami — same from-import disease, same per-module cure: the five copies (server, models, datasets, jobs, runners/hf_cloud) stubbed to None, the "not logged in" shape existing per-test mocks already use. hf_auth.cached_whoami is left real so tests/test_utils_hf_auth.py keeps testing genuine caching behaviour; per-test mocks still override.

Cleanup

• Corrected the tmp_lerobot_home docstring, which claimed from-imports were covered by patching cfg.
• Removed _reset_model_cache from tests/test_models.py: conftest's _reset_hub_listing_caches already calls invalidate_model_listing_cache(), so it was redundant and its docstring ("but not this one") was false.
• The isolation fixture was drafted twice, once per test file; it is consolidated into tests/conftest.py rather than duplicated.

Verification

Check	Result
Full suite (real HOME)	1265 passed
Populated-HOME + HF_HUB_OFFLINE=1, before	8 failed (7 × test_models, 1 × test_list_all_datasets_merges_hub_and_local)
Populated-HOME + HF_HUB_OFFLINE=1, after	267 passed, 0 failed
Real ~/.cache/huggingface/lerobot sha256 manifest (1053 files)	byte-identical before/after every run
ruff check / ruff format --check	clean
pre-commit run --all-files	all hooks passed

The populated-HOME simulation seeds an adversarial HOME-shaped tree — pins, hidden lists naming ids the tests seed themselves, local dataset dirs, and a hub-cache entry — to prove the redirects hold rather than merely that CI is empty.

Deferred / follow-ups

• Root cure: a call-time-accessor refactor of utils/config.py. After it, _FROM_IMPORTED_PATH_CONSTANTS empties and the guard becomes a pure regression detector. The list is a stopgap that must be kept in sync with grep -rn "from .utils.config import" makermodslab/.
• Rig-side, intentionally not carried here: three stray-cleanup tests in tests/test_auto_calibrate.py (added by 5e95ea9, which is not on main) write through auto_calibrate's copy of CALIBRATION_BASE_PATH_ROBOTS into the real follower calibration. Their fix is commit b95680a, which must land on rig before the P0 safety train carve — this PR targets main, where neither those tests nor the feature they cover exist. The conftest protections here are derived from b95680a but authored fresh against main's actual fixture.
• Discovered, not fixed: 6 tests/test_jobs.py cloud-resume tests fail under a token-less fake HOME (test_local_parent_resumed_on_the_cloud_*). Identical failures at baseline, so not a regression from this PR, but they indicate those tests depend on ambient HF auth — the same family of leak, worth a separate pass.

🤖 Generated with Claude Code