test(metal): add CPU-vs-Metal parity tests for remaining 4 kernels (round 3)

[lusoris]

Summary

Round 3 of Metal kernel test coverage. With this PR all 8 registered Metal extractors have a real per-kernel CPU-vs-Metal score parity gate:

• PR test(gpu): GPU backend kernel parity coverage gap-fill (CUDA/HIP/SYCL/Metal) #351 closed the registration audit (all 8 discoverable).
• PR test(metal): add CPU-vs-Metal parity tests for 4 kernels #379 added parity tests for the 4 highest-priority kernels (motion_v2, integer_psnr, float_psnr, float_ssim).
• This PR adds parity tests for the remaining 4: integer_motion, float_motion, float_moment, float_ms_ssim.

New tests (all core/test/, all suite : ['fast', 'gpu']):

Kernel	CPU twin	Score key(s)	Tolerance	ADR
integer_motion_metal	motion	VMAF_integer_feature_motion2_score	1e-4	ADR-0214
float_motion_metal	float_motion	VMAF_feature_motion2_score	1e-4	ADR-0214 / PR #1018
float_moment_metal	float_moment	4 keys (ref1st, dis1st, ref2nd, dis2nd)	1e-4	ADR-0214
float_ms_ssim_metal	float_ms_ssim	float_ms_ssim	1e-3	ADR-0589 (SSIM family)

Skip path. Each test calls vmaf_metal_state_init. On Linux / Windows / Intel Mac that returns -ENODEV; the test emits [skip: no Metal device] and passes cleanly. Apple-Family-7+ macOS CI lanes run the live kernels and assert the per-kernel tolerance. Pattern mirrors PR #379 and core/test/test_sycl_motion3_parity.c.

Wiring. 4 new executable() + test() entries appended inside the existing enable_metal guard in core/test/meson.build, immediately after the round-2 block (or after test_metal_install_header pre-PR-#379-merge) — no boundary edits, so upstream syncs cannot conflict here.

PR overlap (avoided)

• PR fix(metal): dispatch table — canonical score names (silent CPU fallback) #294 (Metal dispatch) — covers the dispatch table; this PR consumes it.
• PR fix(core): emit -ENOSYS stubs for libvmaf_hip.h / libvmaf_metal.h when backend OFF #308 (Metal stubs) — covers -ENOSYS contracts; this PR exercises the real runtime.
• PR test(gpu): GPU backend kernel parity coverage gap-fill (CUDA/HIP/SYCL/Metal) #351 (registration audit) — only asserted vmaf_get_feature_extractor_by_name discoverability; this PR runs the actual extractor pipeline for 4 more kernels.
• PR test(metal): add CPU-vs-Metal parity tests for 4 kernels #379 (round 2: motion_v2 / integer_psnr / float_psnr / float_ssim parity) — sibling; this PR is the second-half companion. If PR test(metal): add CPU-vs-Metal parity tests for 4 kernels #379 lands first, the new meson block slots after the round-2 block; if this PR lands first, PR test(metal): add CPU-vs-Metal parity tests for 4 kernels #379 trivially re-anchors after test_metal_install_header.

Per-kernel coverage count (post-merge)

Round	PR	Kernels covered
1 (registration)	#351	8 (discoverability only)
2 (parity, batch A)	#379	4 (motion_v2, integer_psnr, float_psnr, float_ssim)
3 (parity, batch B)	this PR	4 (integer_motion, float_motion, float_moment x4 keys, float_ms_ssim)
Total real-parity		8 / 8

Test plan

• CPU-only build green: cd core && meson setup build-cpu -Denable_cuda=false -Denable_sycl=false -Denable_metal=disabled succeeds (new tests gated off on non-Metal).
• pre-commit run --files on every touched file -> all hooks pass (clang-format, copyright header, semgrep, conventional-commit, etc.).
• bash scripts/ci/check-copyright.sh core/test/test_metal_*_parity.c -> silent (clean).
• bash scripts/ci/assertion-density.sh core/test/test_metal_*_parity.c -> PASS (every fork-added function ≥20 lines has ≥1 assert).
• clang-format --dry-run --Werror core/test/test_metal_*_parity.c -> no diff.
• Apple Silicon CI lane (macOS Metal): meson test -C build --suite=fast test_metal_integer_motion_parity test_metal_float_motion_parity test_metal_float_moment_parity test_metal_float_ms_ssim_parity exercises the live kernels and asserts parity. (Will run automatically when the macOS Metal CI lane picks up this PR — Apple-Family-7+ device required.)

ADR-0108 deliverables

• Research digest: no digest needed: trivial test-only addition (follow-up to PR test(gpu): GPU backend kernel parity coverage gap-fill (CUDA/HIP/SYCL/Metal) #351 / PR test(metal): add CPU-vs-Metal parity tests for 4 kernels #379 documented gap; same pattern, different kernels).
• Decision matrix: no alternatives: only-one-way fix (the synthetic-fixture + skip-on-ENODEV pattern is the established test_sycl_motion3_parity.c + PR test(metal): add CPU-vs-Metal parity tests for 4 kernels #379 precedent; tolerance choices inherit from existing ADR-0214 / ADR-0589).
• AGENTS.md invariant note: no rebase-sensitive invariants in this PR. PR test(metal): add CPU-vs-Metal parity tests for 4 kernels #379 already extends core/test/AGENTS.md "GPU tests must skip gracefully" rule to mention vmaf_metal_state_init and Intel-Mac -ENODEV; this PR consumes that contract without re-stating it.
• Reproducer / smoke-test command:

  cd core && meson setup build-cpu -Denable_cuda=false -Denable_sycl=false -Denable_metal=disabled
  ninja -C build-cpu
  # On Apple Silicon:
  cd core && meson setup build -Denable_metal=enabled
  meson test -C build --suite=fast test_metal_integer_motion_parity test_metal_float_motion_parity test_metal_float_moment_parity test_metal_float_ms_ssim_parity

• CHANGELOG fragment: changelog.d/added/metal-kernel-coverage-round3.md.
• Rebase note: docs/rebase-notes.md -> "Metal kernel parity tests round 3 (2026-05-31)" entry added; rebase impact NONE (all new file paths are fork-local Metal-only, the meson edit lives inside the fork-added enable_metal block).

docs/state.md updated with the T-METAL-KERNEL-PARITY-ROUND3-2026-05-31 row under "Recently closed" (per CLAUDE §12 r13).

Cross-references

• ADR-0214 (cross-backend parity gate, places=4 default)
• ADR-0361 (Metal backend scaffold)
• ADR-0421 (first Metal kernel = motion_v2)
• ADR-0589 (Metal SSIM family 1e-3 bound)
• PR test(gpu): GPU backend kernel parity coverage gap-fill (CUDA/HIP/SYCL/Metal) #351 (registration audit, round 1)
• PR test(metal): add CPU-vs-Metal parity tests for 4 kernels #379 (parity tests, round 2)
• PR chore(deps): Update dependency mcp to >=1.28.0 #1018 (float_motion_metal parity with float_motion_cuda)

🤖 Generated with Claude Code

[lusoris]

parity sweep audit: clean, no symmetric bug.

Verified all 4 new test files use the correct Metal-suffixed registry names:

• test_metal_integer_motion_parity.c -> integer_motion_metal (registered at core/src/feature/metal/integer_motion_metal.mm:324)
• test_metal_float_motion_parity.c -> float_motion_metal (core/src/feature/metal/float_motion_metal.mm:335)
• test_metal_float_moment_parity.c -> float_moment_metal (core/src/feature/metal/float_moment_metal.mm:266)
• test_metal_float_ms_ssim_parity.c -> float_ms_ssim_metal (core/src/feature/metal/float_ms_ssim_metal.mm:553)

CPU baselines (motion, float_motion, float_moment, float_ms_ssim) all resolve correctly in the CPU registry. The integer_motion_metal test correctly pairs against CPU 'motion' (integer path in core/src/feature/integer_motion.c). No fix needed.