fix(cuda): download every plane of device input for CPU extractors

[lusoris]

Summary

With the pictures in device memory and an extractor that runs on the CPU, libvmaf copied the luma plane into the host picture only; the chroma planes stayed as vmaf_picture_alloc() left them and the CPU extractor scored them without an error: psnr_cb came out as the 60 dB cap where the CPU gives 12.54 dB. The device-to-host copy now takes every plane the picture has.

Closes T-CUDA-DEVICE-INPUT-CHROMA-NOT-DOWNLOADED-2026-10-01 (found and closed here). Netflix/vmaf#1613 reports the same line on upstream. RC3, correctness (ADR-1421).

Cause

translate_picture_device() in core/src/libvmaf.c called vmaf_cuda_picture_download_async(pic, pic_host, 0x1). The host-to-device upload next to it takes 0x7 (0x1 for 4:0:0). A CPU extractor gets a device picture's host copy when a registered extractor has no CUDA twin or the context carries a gpumask; the CUDA twins read the device pictures directly, so the default model on CUDA is not affected (psnr_cuda y/cb/cr equal the CPU on the Netflix pair, host input, at --threads 0 and 4: 0.0 difference over 48 frames).

Output on an RTX 4090

test_cuda_device_input_host_extractors (4 frames of noise on every plane at 192x128, pictures from the DEVICE preallocation pool, gpumask = 1 so psnr runs on the CPU, every psnr_y / psnr_cb / psnr_cr compared with == against a CPU-only run on the same pictures, n_threads 0 and 4):

	Result
Before (run twice)	threads=0 frame 0 psnr_cb: cpu=12.543948908054073 device-input=60, fail
After (run twice)	pass, both thread counts

Type

• fix — bug fix

Deep-dive deliverables (ADR-0108)

• Research digest — no digest needed: a one-constant fix; the evidence is in the state row.
• Decision matrix — no alternatives: only-one-way fix (every plane the picture has).
• AGENTS.md invariant note — core/src/cuda/AGENTS.md, "Device-to-host picture download takes every plane".
• Reproducer / smoke-test command — under "Reproducer" below.
• CHANGELOG fragment — changelog.d/fixed/cuda-device-input-chroma-download.md.
• Rebase note — docs/rebase-notes.md, "translate_picture_device() downloads every plane".

Reproducer

flock ~/.cache/vmafx-locks/cuda-4090.lock timeout 120 build/test/test_cuda_device_input_host_extractors

Notes

• core/src/libvmaf.c is also touched by the ownership PR (ADR-1431) in vmaf_read_pictures() and check_ring_buffer(); this change is in translate_picture_device().
• Touches core/test/meson.build (one block). No public API, ABI or FFmpeg patch impact. No Netflix golden assertion changed.