fix(arm64): make adm_dwt2_8_neon bit-exact with the scalar kernel

[lusoris]

Summary

ADR-1057 was a dropped filter tap in
adm_dwt2_8_neon that reached master and drifted the ARM Netflix golden. It
survived every test because no unit test covered adm_dwt2 on any
architecture — test_integer_adm_simd.c reaches adm_cm, not adm_dwt2.

This adds that test. It immediately found two more divergences from the scalar
kernel.

1. Stale intermediate columns for widths ≡ 8 (mod 16)

integer_adm.c dispatches this kernel on !(w % 8):

if (flags & VMAF_ARM_CPU_FLAG_NEON) {
    if (!(w % 8))
        s->dwt2_8 = adm_dwt2_8_neon;

but its vertical pass advances 16 columns per iteration and stops at
w - 15, with no scalar tail:

for (int j = 0; j < w - 15; j += 16, ...)

For a width such as 584 or 776 the final 8 columns of tmplo / tmphi were
never written in that iteration, so the horizontal pass consumed whatever the
previous row had left in the shared buf->tmp_ref scratch.

2. Unmirrored final output column, at every width

The horizontal pass walks tmplo / tmphi with plain pointer arithmetic and
never consults ind_x, so for the last column it read tmplo[2j + 2] ==
tmplo[w] — one past the lo half, i.e. tmphi[0], since tmphi == tmplo + w —
where the scalar mirrors the index back to w - 1.

Score impact: none

I want to be precise about this rather than oversell it. I built a synthetic
584-wide clip specifically to trigger defect 1, and measured:

build	VMAF
ARM NEON, pre-fix	80.322171
ARM NEON, post-fix	80.322171
x86 scalar	80.322171

ADM crops the affected border columns before they reach a feature value. And the
Netflix golden fixtures are 576, 1280 and 1920 wide — all multiples of 16 —
so defect 1 could never have surfaced there either.

This is a correctness and cross-backend-parity fix, not a scoring fix. It is
still worth making: defect 1 reads scratch entries that were never written in
that iteration, so the kernel's output depends on the previous row's residue.

Type

• fix — bug fix
• test — adds the missing coverage
• simd — backend-specific

Checklist

• Commits follow Conventional Commits.
• make lint clean; clang-format and assertion-density pass on both touched files.
• Unit tests pass: aarch64 98/98 under qemu-aarch64-static, x86 105/105.
• SIMD path touched — bit-exactness against the scalar reference is exactly what the new test asserts, across six geometries. Worst divergence after the fix: 0.
• No feature extractor added; the twin situation is unchanged.
• New .c carries the license header.
• Not a breaking change.
• No ADR — bug fix, no architectural choice another engineer could reasonably have made differently.

Bug-status hygiene

• docs/state.md updated — T-ADM-DWT2-NEON-PARITY-2026-08-30 under Recently closed.

Netflix golden-data gate

• I did not modify any assertAlmostEqual(...) score. Verified score-neutral on the 576-wide golden pair (76.667831 pre- and post-fix) and on the synthetic 584-wide clip.

Deep-dive deliverables

• Research digest — no digest needed: the investigation is the test itself, and its findings are recorded in docs/state.md and the commit message.
• Decision matrix — no alternatives: only-one-way fix. A SIMD kernel either matches its scalar reference or does not.
• AGENTS.md invariant note — captured in docs/rebase-notes.md instead: the invariant is about the relationship between the dispatch granularity (w % 8) and the vector width (16), which is not a package-scoped concern.
• Reproducer / smoke-test command — below.
• CHANGELOG fragment — changelog.d/fixed/adm-dwt2-neon-scalar-parity.md.
• Rebase note — docs/rebase-notes.md, "fix/adm-dwt2-neon-parity".

Reproducer

meson setup build-arm64 core --cross-file build-aux/aarch64-linux-gnu.ini \
  -Denable_cuda=false -Denable_sycl=false -Denable_dnn=disabled
ninja -C build-arm64
meson test -C build-arm64 --suite=fast        # 98/98, incl. test_adm_dwt2_neon

Show the defects on the pre-fix kernel — revert adm_neon.c and the test reports
per-geometry mismatch counts:

16x16: 32 mismatches (32 last-column)     <- defect 2
24x16: 160 mismatches (32 last-column)    <- defect 2 + 128 stale columns (defect 1)
40x17: 180 mismatches (36 last-column)

128 = 4 columns × 8 rows × 4 subbands, exactly the region the missing vertical
tail leaves unwritten at w=24.

Known follow-ups

• The same "no test covers this kernel" gap likely exists for other arm64/
  kernels. adm_cm has coverage via test_integer_adm_simd.c; the rest of
  adm_neon.c (decouple, csf, sum) does not.