test: remove tests requiring private TensorAuto/*-untrained repos (#272)

[WilliamYue37]

What this does

Fixes #272 (🐛 Bug). The nightly Run Pytest on GPU job has been failing with 8 fixture-setup ERRORs because the CI HF_TOKEN can't see three private repos under TensorAuto/*:

• TensorAuto/pi06-untrained
• TensorAuto/pi07-high-untrained
• TensorAuto/pi07-low-untrained

This PR removes the affected tests and their now-unused fixtures, taking option (3) from the issue's suggested fixes (skip-on-unreachable, but as a removal rather than a runtime probe — the tests can be added back behind a token-aware fixture if/when the repos are made public or the token is rotated).

Removed:

• tests/policies/test_pi07_untrained.py (entire file — every test in it loaded one of the private pi07 repos)
• The "Published TensorAuto/pi06-untrained checkpoint" section at the bottom of tests/policies/test_pi06.py (the pi06_untrained_policy / gemma3_4b_pt_aligned_state_dict fixtures, the _diff_state_dicts_against_reference helper, and the three tests that consumed them — including the already-@pytest.mark.skip'd VLM-match test, since its fixtures were the only remaining users of the helper).

The CPU-only tests in test_pi06.py (block-causal mask semantics, RoPE θ selection, SDPA equivalence, gradient-checkpointing equivalence, etc.) are untouched.

How it was tested

pytest tests/policies/test_pi06.py -m "not gpu" -n auto
# 37 passed in 9.53s

pre-commit run --files tests/policies/test_pi06.py
# all hooks passed

A grep for TensorAuto/pi0[67]*-untrained shows only one remaining hit, in a docstring at tests/configs/test_default.py:246 — that's just historical context, not a code path.

How to checkout & try? (for the reviewer)

git checkout fix/272-remove-private-hf-tests
pytest tests/policies/test_pi06.py -m "not gpu" -n auto
pytest tests/policies/test_pi07_untrained.py 2>&1 | head -3   # file no longer exists

On the GPU box, the previously-failing collection should now be empty:

ssh 5090
cd ~/OpenTau
pytest -m "gpu" -n 0 -v tests/policies/test_pi06.py tests/policies/test_pi07_untrained.py

Checklist

• I have added Google-style docstrings to important functions and ensured function parameters are typed.
• My PR includes policy-related changes.
  • If the above is checked: I have run the GPU pytests (pytest -m "gpu") and regression tests.

Note: Before submitting this PR, please read the contributor guideline.

[claude]

Reviewed deletions only — verified no remaining references to the removed fixtures/helpers (pi06_untrained_policy, pi07_*_untrained_policy, _diff_state_dicts_against_reference, _assert_no_missing_or_unexpected, gemma3_4b_pt_aligned_state_dict) and no orphaned imports in test_pi06.py. CPU subset of test_pi06.py is intact. One minor stale doc reference noted inline.

[claude]

[claude-review] summary for commit 6db70fd

No blocking issues found. The two nits from the prior review pass were addressed in 6db70fd as docstring/comment-only edits — verified the new comment in tests/utils/test_vision_utils.py:53-56 redirects to bilinear_resample_pos_embed's docstring, which already explains the byte-for-byte bootstrap motivation (src/opentau/utils/vision_utils.py:45-48), and the rewrites at tests/configs/test_default.py:208-209,245-246 preserve the bug context without naming the private repos. Repo-wide grep finds zero remaining references to pi06-untrained / pi07-{high,low}-untrained / the removed fixtures and helpers (pi06_untrained_policy, pi07_*_untrained_policy, gemma3_4b_pt_aligned_state_dict, _diff_state_dicts_against_reference, _assert_no_missing_or_unexpected, test_pi06_untrained_siglip_matches_gemma3_4b_pt). CPU-only test_pi06.py tests are intact and its imports are all still in use.

[WilliamYue37]

@claude fix the nits

[claude]

[claude-fix] both nits addressed in 6db70fd.

• tests/utils/test_vision_utils.py:54 — reworded the test_deterministic comment to drop the reference to the removed test_pi06_untrained_siglip_matches_gemma3_4b_pt symbol; it now points readers at the bilinear_resample_pos_embed docstring (which already explains the byte-equality bootstrap motivation), matching the reviewer's suggestion to redirect to a stable reference.
• tests/configs/test_default.py:209,246 — even though the reviewer flagged this only "for completeness," I went ahead and rephrased both docstrings to drop the stale TensorAuto/pi07-{high,low}-untrained repo names while preserving the historical bug context (codec round-trip needed for a π0.7 planner checkpoint bootstrap from public Gemma 3 weights). Tests there are unchanged.

Tests skipped: docstring/comment-only edits in two test files, no imports / function bodies / control flow / runtime-read configs touched.