fix(tests): exclude blue-named .pth files from green skip-download property

[MrAnkleManic]

Problem

tests/test_pbt_auto_download.py::TestExistingCheckpointSkipsDownload::test_existing_pth_skips_download fails once Hypothesis generates a .pth filename containing the blue token, e.g. blue.pth (after which the falsifying example is cached in .hypothesis/ and the failure becomes deterministic):

OSError: [Errno 9] Bad file descriptor
Falsifying example: test_existing_pth_skips_download(..., pth_name='blue.pth')

Root cause

Two layers:

1. The property is stale since feat: dedicated blue-screen model (CorridorKeyBlue) with auto-detection #241. _discover_checkpoint now classifies any checkpoint whose basename contains blue (case-insensitive) as a blue checkpoint. The test calls _discover_checkpoint(TORCH_EXT) with the default screen_color="green", so a generated name like blue.pth is (correctly) filtered out of green discovery and the auto-download path runs — which the test forbids. The property "any existing .pth skips download" stopped being true when the filename convention landed.

2. Why it surfaces as EBADF instead of a clean assertion: the test patches hf_hub_download with a bare mock.patch (no return_value), so the "downloaded path" is a MagicMock. That flows into shutil.copy2() → open(src, 'rb'), and since MagicMock.__index__() returns 1, FileIO treats the mock as file descriptor 1 — opening pytest's captured stdout for reading and closing it on exit, which raises OSError: [Errno 9] Bad file descriptor and poisons subsequent examples in the run.

Fix (test-only)

• Restrict Property 2's strategy to green-classified names by filtering out the blue token (imported as BLUE_FILENAME_TOKEN, matched case-insensitively exactly as the backend does).
• Add a companion property: a blue-named .pth skips the download when screen_color="blue".
• Give the must-not-download mocks (Properties 2 and 3) side_effect=AssertionError(...) so any future regression fails with a clear message instead of fd corruption.

Verification

All 5 tests in the file pass, including replaying the previously cached blue.pth falsifying example from .hypothesis/. ruff check and ruff format --check are clean. uv.lock untouched.

🤖 Generated with Claude Code