fix(hip): let the scaffold posture report -ENOSYS so its tests skip instead of failing

[lusoris]

Summary

On an ordinary -Denable_hip=true build, four HIP parity tests failed where they were meant to skip. master (551d35a63) on ROCm 7.2.4 / gfx1036: 173 ok, 4 fail.

enable_hipcc defaults to false, so that build compiles the HIP host code but no device kernels. core/meson_options.txt documents the posture — extractors "fall back to -ENOSYS" — and every HIP parity test turns -ENOSYS into [skip: …] and passes. Two independent things broke that.

The extractor side

float_vif_hip.c and integer_psnr_hvs_hip.c wrote their scaffold path as:

int err = vmaf_hip_kernel_submit_pre_launch(&s->lc, s->ctx, NULL, 0, 0);
if (err != 0)
    return err;
return -ENOSYS;      /* unreachable */

The call passes rb == NULL, and rejecting a NULL rb is that helper's first statement, so it always returned -EINVAL and the -ENOSYS below was dead code. The extractor reported "invalid argument" where the contract promises "not implemented", the tests' skip branch never matched, and they failed. The call did nothing else — it returned before touching anything — so it is gone.

The test side

speed_temporal_hip returns -ENOSYS correctly, but from extract(), which surfaces through vmaf_read_pictures(): registration succeeds because only extract() knows there is no kernel. test_hip_speed_singular_parity.c checked -ENOSYS only at the vmaf_use_feature() site. Its own file header already claimed "the same skip contract as test_hip_speed_temporal_parity.c", and that sibling has always checked both sites — this one just never implemented the second.

How it was found

Not by inspection. AMD_LOG_LEVEL=1 showed no HIP API error, and the extractor's own six -EINVAL guards all log at ERROR and printed nothing at VMAF_LOG_LEVEL_DEBUG. Tagging every return -E* in the read path and the HIP runtime with its __LINE__ named the source in one run: core/src/hip/kernel_template.c, the lc == NULL || rb == NULL guard. The speed test's own error printed as -38, i.e. -ENOSYS — which is what pointed at the test rather than the extractor.

Verification

Both configurations, because a fix that makes tests skip has to be shown not to be hiding a real failure.

Build	Master	This branch
-Denable_hip=true (default, no kernels)	173 ok, 4 fail	177 ok, 0 fail
-Denable_hip=true -Denable_hipcc=true (real gfx1036 kernels)	—	183 ok, 0 fail (2 expected fail, 1 unexpected pass)

The four now print explicitly what they did:

test_hip_psnr_hvs_parity         [skip: HIP extractor is a scaffold (-ENOSYS)]
test_hip_float_vif_parity        [skip: HIP scaffold ENOSYS on feed]
test_hip_speed_singular_parity   [skip: HIP scaffold ENOSYS on submit]

and with enable_hipcc=true the same four run against real kernels and pass, so the skip is a genuine "not built", not a silenced failure. Both edits are inside #ifndef HAVE_HIPCC / an err == -ENOSYS branch, so they are inert once kernels exist — the real-kernel run confirms that rather than asserting it.

The remaining test_hip_adm_parity unexpected pass and two expected fails in the real-kernel run are the stale should_fail markers PR #1493 removes, not this change.

praetorctl audit passes and the baseline tightens 1433 → 1415: deleting the dead call removed 18 findings. Re-recorded from a clean clone, branch verified. mkdocs build --strict exits 0; check-state-md-rows OK.

Type

• fix — bug fix

Checklist

• Commits follow Conventional Commits.
• make format is green; pre-push hooks pass.
• Unit tests pass — see the table, in both HIP configurations.
• I did not modify any assertAlmostEqual(...) score in the Netflix golden Python tests.
• docs/state.md: row T-HIP-SCAFFOLD-ENOSYS-MASKED-2026-09-19 added as closed.

Deep-dive deliverables (ADR-0108)

• Research digest — no digest needed: trivial. The diagnosis is three measurements, recorded in this description, ADR-1264 and the state row.
• Decision matrix — ADR-1264 ## Alternatives considered, four options including relaxing the helper's NULL guard and marking the tests should_fail.
• AGENTS.md invariant note — core/src/feature/hip/AGENTS.md, plus docs/rebase-notes.md.
• Reproducer / smoke-test command — below.
• CHANGELOG fragment — changelog.d/fixed/hip-scaffold-enosys.md.
• Rebase note — docs/rebase-notes.md.

Reproducer

# The default posture: 4 failures on master, 0 here.
meson setup build core -Denable_cuda=false -Denable_sycl=false -Denable_hip=true
ninja -C build && meson test -C build --suite=fast

# With real kernels, to show the skip is not hiding anything (needs ROCm + hipcc):
meson setup build-hipcc core -Denable_cuda=false -Denable_sycl=false \
    -Denable_hip=true -Denable_hipcc=true
ninja -C build-hipcc && meson test -C build-hipcc --suite=fast

Rebase hazard

Both shapes come back easily, and docs/rebase-notes.md plus core/src/feature/hip/AGENTS.md name them: a scaffold path returns -ENOSYS directly and never calls a kernel-submit helper with placeholder arguments, and a HIP parity test checks -ENOSYS at both vmaf_use_feature() and vmaf_read_pictures(). Do not relax the helper's NULL rb guard to accommodate the old call — roughly twenty real callers depend on it.

[lusoris]

Absorbed into the integration train #1506, folded unchanged except where that PR's description says otherwise. This PR was green on its own; it and its three siblings kept knocking each other into conflict on docs/state.md and the generated ADR indexes every time one of them merged, so they land together.