The CLAP encoder's CUDA bicubic backward has no deterministic kernel and
Adam amplifies that noise into large parameter swings between runs. CPU is
bit-exact and no slower at fixture size, so the test now runs the CLI in a
subprocess with CUDA hidden and compares floats with a small relative
tolerance, naming the JSON path on mismatch.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>