fix(arm64): adm_dwt2_8_neon dropped the 4th filter tap at j=0 (ADR-1057 rediagnosed)

[lusoris]

What

Fixes the aarch64 Netflix golden failure. One character: idx < 3 → idx < 4.

The bug

adm_dwt2_8_neon special-cases the j = 0 output column because its source
indices are mirrored ({1, 0, 1, 2} rather than the regular 2j-1 .. 2j+2
stride). That special case accumulated with:

for (int idx = 0; idx < 3; idx++) {          // ← three
    accum_a += (int32_t)dwt2_db2_coeffs_lo[idx] * s_lo;

The scalar reference adm_dwt2_8 (core/src/feature/integer_adm.c:2606)
multiplies all four taps — filter_lo[0..3] against s0..s3 from
ind_x[0..3][j]. The NEON path therefore dropped ind_x[3][0] and its
coefficient -4240 from the first output column of every DWT2 row, on every
scale.

ADR-1057's diagnosis was wrong

ADR-1057 attributes this to
float-ADM NEON FMA contraction. It is not an FP problem at all — integer ADM
accumulates in int64. Proof: building the entire tree with
-ffp-contract=off on aarch64 leaves the score at 88.030322, byte-for-byte
unchanged. That is why PR #1060's fp-contract approach could not have worked and
#1063 reverted it.

Reproduced locally — no CI round-trip needed

aarch64-linux-gnu-gcc cross-build run under qemu-aarch64-static, driving the
real pytest harness via a shim at core/build/tools/vmaf:

Configuration	integer_adm2	adm3	adm_scale3	akiyo
NEON	1.116698	1.052571	1.13728	88.030322 — 3 fail
-Denable_asm=false	1.116701	1.052573	1.13729	10/10 pass
NEON + -ffp-contract=off	—	—	—	88.030322 unchanged
after this fix	—	—	—	38 passed, 1 skipped, 0 failed

The final row covers vmafexec_test.py + result_test.py on aarch64, including
every akiyo golden assertion.

x86 is unaffected — the diff touches only core/src/feature/arm64/adm_neon.c.

Two gaps this exposes

1. No NEON-vs-scalar parity test covers the ADM DWT2 kernels on any
   architecture. test_integer_adm_simd.c only exercises adm_cm_avx2. A
   three-vs-four tap error shipped because nothing compares the kernels directly.
2. The ARM golden leg is not a required check, so it stayed red for months
   rather than blocking a merge.

Both are worth closing separately; neither is in scope here.

Deep-dive deliverables

• Research digest — no digest needed: the isolation table above is the analysis and is reproduced in the commit body.
• Decision matrix — the table above: FP-contraction ruled out by direct measurement before the integer off-by-one was identified.
• AGENTS.md invariant note — no rebase-sensitive invariants: fork-local NEON kernel.
• Reproducer / smoke-test command — see below.
• CHANGELOG fragment — changelog.d/fixed/adr-1057-adm-neon-dwt2-tap.md
• Rebase note — no rebase impact: fork-local arm64 SIMD path with no upstream counterpart.

Reproducer / smoke-test command

cat > /tmp/aarch64.ini <<'EOF'
[binaries]
c = 'aarch64-linux-gnu-gcc'
cpp = 'aarch64-linux-gnu-g++'
ar = 'aarch64-linux-gnu-ar'
strip = 'aarch64-linux-gnu-strip'
exe_wrapper = '/tmp/qemu-wrap'
[host_machine]
system = 'linux'
cpu_family = 'aarch64'
cpu = 'aarch64'
endian = 'little'
EOF
printf '#!/bin/sh\nexec qemu-aarch64-static -L /usr/aarch64-linux-gnu "$@"\n' > /tmp/qemu-wrap
chmod +x /tmp/qemu-wrap

meson setup build-arm core --cross-file /tmp/aarch64.ini -Denable_cuda=false -Denable_sycl=false
ninja -C build-arm tools/vmaf

mkdir -p core/build/tools
printf '#!/bin/sh\nexec qemu-aarch64-static -L /usr/aarch64-linux-gnu "$(dirname "$0")/../../../build-arm/tools/vmaf" "$@"\n' > core/build/tools/vmaf
chmod +x core/build/tools/vmaf

PYTHONPATH=$PWD/python python3 -m pytest python/test/vmafexec_test.py \
  -k akiyo_multiply -q          # 10 passed

Bug status hygiene

• docs/state.md updated — T-NEON-FMA-FLOAT-ADM-DWT2-2026-06-06 moved to Recently closed with the corrected root cause.