Cassette Build Report 030 — The Skip Came After the Import
S12's Apple Silicon boundary was real, but its test module imported MLX before it could skip, so the portability fix had to move earlier.
Scope note — This report covers S12’s test-harness portability boundary and the repair that moved the skip before platform-bound imports. It does not claim that S12’s MLX behavior runs outside the Apple Silicon Metal environment that owns its acceptance proof.
S12 was the first step where Cassette’s mathematical description became executable tensor operations. The new certificate kept metrics, compatibility witnesses, atom cover, distortion, error, risk, observation adequacy, and physical resources in separate bounded fields. pager.py validated those records before MLX could allocate, then sent admitted operations through generated dispatch.
On the Mac, the work was sound. Ten golden rows covered matrix multiplication, four-bit quantized multiplication, RMS normalization, rotary position encoding, attention, convolution, embedding, keyed categorical sampling, autograd, and SGD. Wrong shapes and undeclared operators returned typed errors. Eight disposable guard removals made their fixtures fail. Twenty-eight tests passed.
The review found a problem outside the operator behavior. tests/test_s12_pager.py imported mlx.core and pager at module load. The platform decision came later. On a machine without MLX, collection failed before pytest could skip. On a Linux machine with MLX, the fixture ran far enough to hit the Apple Silicon Metal guard. Both paths meant a later env step could not run the complete repository suite.
Opus 5 Extra got part of the description wrong. It called MLX Apple-only because its own Linux environment could not import it, although the lockfile contained Linux wheels. It also proposed copying a pytestmark pattern from earlier files. A marker placed after a failing import cannot protect the import.
That ordering mattered beyond Linux: an environment that cannot import MLX still needs to run the repository’s portable checks, which means the test must decide whether it is eligible before it asks Python to load the dependency that may not exist.
I accepted the finding and ignored the overreach. Then I reproduced both failures. In a synthetic Linux environment with an import blocker, the module reached mlx before pytest could do anything. With MLX available under the same synthetic platform, the fixture ran and stopped at the Metal capability check. The defect was one of order, with the platform decision happening after platform-bound imports.
The correction is small enough to read as a contract.
if platform.system() != "Darwin" or platform.machine() != "arm64":
pytest.skip(
"S12 requires arm64 macOS with MLX Metal; skip before importing the runtime",
allow_module_level=True,
)
import mlx.core as mx
import pager
The S12 module gate now decides eligibility before either runtime-bound import runs. That establishes a truthful skip for an ineligible environment. It does not turn Linux into Apple Silicon evidence, and it does not weaken the Mac proof.
The mutation made the boundary testable. Removing only the early guard caused the synthetic Linux probe to reach the forbidden MLX import. Restoring it left the original Mac behavior unchanged: all three S12 fixtures passed on arm64 macOS, and the complete suite passed twenty-eight tests with no skips.
The exchange also exposed a difference between Opus’s review and the implementation agent’s response. Opus found the right defect while making two claims that did not survive inspection and offering a remedy that arrived too late. The implementation agent did not need to accept or reject the review as a package. It separated the valid observation from the invalid explanation, reproduced the failure, and applied the smallest repair that could be falsified.
Drew’s request, “Please remediate,” changed the role of the turn. The task was no longer to discuss whether the finding sounded plausible. It was to reopen S12, make the repository fail for the old reason, repair it, remove the repair in a disposable copy, and rerun the declared platform proof.
S12 now distinguishes its two environments without confusing them, because on Apple Silicon the certificate and generated operators execute while elsewhere the test module exits before touching MLX or pager, allowing portable work to run the rest of the suite and state exactly what it did not run. The boundary is part of the test.
