test(vision_mixer): register the cudadownload stand-in once, never remove it

[srperens]

Description

vision_mixer_cuda_input_test (from #948) crashed in about one run in four: 11 of 40 runs locally, and it took down CI on #949. The failure was an ElementFactory instance_of assertion or a segfault. The bug is in the test, not in the input front.

Cause: each test registered a stand-in cudadownload and removed it from the registry again, so that the missing-adapter test could find it absent. Registry::remove_feature frees the factory while the element class keeps a plain, non-refcounted pointer to it. The next registration or lookup under that name then read freed memory. The lock between the tests did not prevent this.

Fix:

• The stand-in is registered once per process and never removed.
• The missing-adapter test moves to its own binary, vision_mixer_cuda_missing_test, so nothing is registered in its process.
• The shared Media Player → Vision Mixer harness moves to backend/tests/vision_mixer_cuda/.
• The test bodies and assertions are unchanged.

Type of Change

• Bug fix (test only)

Testing

• Before: 11 of 40 runs of vision_mixer_cuda_input_test failed locally (macOS).
• After: 40 of 40 runs of vision_mixer_cuda_input_test (2 tests) and 40 of 40 of vision_mixer_cuda_missing_test (1 test, no skip) pass.
• cargo clippy --all-targets -- -D warnings and cargo fmt are clean.
• No revert check: the guards' assertions are unchanged, and this PR changes only how the stand-in is registered.

🤖 Generated with Claude Code