Fix two stale ROCm assertions; skip Apple fixtures the build cannot run

[gstoner]

Follow-up to #619, which fixed the ROCm lld invocation failed root cause but left 59 failures in a full sweep on the Strix Halo box. I checked all 59 rather than assuming they shared a cause: none are ROCm lane defects — all are test-side. This closes 42 and narrows the rest to one file.

ROCm — two stale assertions

test_autodiff_rocm_matmul_composed compared last_backward_execution with == over the whole dict. E2E-REAL-6 added six tracer-authority fields (family, frontend_authority, four *_consumer) and changed implementation to family_plugin_composition, so it failed on additive metadata while the numerics passed. Seven sibling tests already assert the new shape — this was the only one left on the old one. Now asserts the fields the proof depends on by key: additive-tolerant, still fail-closed (a dropped key raises KeyError).

test_rocm_activation_rope_codegen grepped for kind must be gelu, silu, or relu, a pass-body string. Rejection has since moved to the ODS attribute constraint — Decision #21a, EnumAttr over unvalidated StrAttr — so the op is refused before the generator ever runs. Correct, but the new diagnostic named only "ROCm activation kind" and dropped the legal set, which Decision #21 wants a diagnostic to carry.

Rather than weaken the test to match a worse message, ROCM_ActivationKindAttr now names its values — matching ROCM_ReductionAttr two declarations above it in the same file:

error: 'tessera_rocm.activation' op attribute 'kind' failed to satisfy
constraint: ROCm activation kind: 'gelu', 'silu', or 'relu'

Apple — 57 fixtures failed where they should have skipped

tessera-opt registers the Apple pipelines only under -DTESSERA_BUILD_APPLE_BACKEND=ON. The ROCm/x86 boxes build it OFF, so these fixtures died on Unknown command line argument '-tessera-lower-to-apple_cpu-full' (or the --pass-pipeline parser's equivalent).

The existing guard asks whether the binary exists; it never asked whether that binary has the Apple backend in it. A capability the build lacks must skip, not fail — 57 red that say nothing about Apple crowd out real signal, which is exactly how a total ROCm serializer outage went unnoticed until #619.

skip_if_apple_pipeline_unregistered() is called at the point of invocation and is narrow in three ways so it cannot hide a defect:

• fires only on the two "pipeline is not registered" signatures;
• only when the Apple backend is genuinely absent — if it IS built, an unregistered pipeline is a registration regression and stays loud;
• ignores any other nonzero exit, so the negative fixtures still assert their own diagnostics.

A file-level skipif was tried first and rejected: it also disabled 54 fixtures in those files that pass fine without the Apple backend.

Evidence

Strix Halo (gfx1151, Ubuntu 24.04/WSL2) — a host with no Apple backend.

Check	Result
Full sweep	59 failed → 17 failed / 16344 passed
-k "rocm or gfx" -m "not slow"	2828 passed, 0 crashes (unchanged)
ninja -C build check-tessera-rocm	63/63
ruff (python/tessera/ + changed test files) / mypy	clean / 0

Probe validated both directions, so a False is real signal and not a stuck probe:

Probe	Result
Apple pipeline registered	False
tessera-lower-to-rocm (control)	True
CMakeCache TESSERA_BUILD_APPLE_BACKEND (independent of --help)	OFF — agrees

The property that matters: with the probe forced True to simulate the Mac, the fixtures FAIL rather than skip. That is what keeps this from silently disabling the Apple suite on the box that can actually run it.

Not fixed here, deliberately

The remaining 17 (all test_apple_value_target_ir.py) fail through the Python front door (KeyError: 'compiler_path', target_ir_artifact vs value_target_ir) and a --help introspection check — not through a pipeline invocation. Guarding those needs an Apple build to confirm they still run, and this host cannot build the Apple backend; a guard written blind there could mask a real Apple regression. They belong on the Mac.

Five fixtures that previously "passed" here now skip. They asserted a nonzero exit and were getting one from the missing CLI flag rather than from the rejection under test — false passes, not lost coverage.

Claim integrity

All ROCm/gfx1151 evidence was produced on the box with that device. No Apple claim is made: this host cannot build or run the Apple backend, so the Apple change is verified only as "these fixtures skip here, and provably do not skip when the backend is present." Whether they still pass on the Mac is unverified here and needs a Mac run before that is asserted.

🤖 Generated with Claude Code

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 2c3beecdcb

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".