--recode-video on an audio file with embedded cover art still routes into the video transcode path

[crippledgeek]

Found in code review of #637. Adjacent to what that issue fixed, and invisible to its test matrix.

Problem

#637 routes a source with no video stream to an audio-only conversion, gated on SourceVideo::Absent, which comes from MediaInfo::has_video.

But probe sets has_video = true for any stream of type video — including an attached_pic cover-art still (crates/rdlp-ffmpeg/src/ffmpeg/probe.rs:201-203). So an .m4a/.mp3/.flac that already carries embedded artwork is not SourceVideo::Absent. It classifies as a video source, can_remux_video answers about the mjpeg still as though it were the video track, and a transcode re-encodes a single cover image as a video stream.

Why #637's matrix didn't catch it

The fixture is built with ffmpeg -f lavfi -i sine ... -c:a aac — no cover art. Every container in that matrix is genuinely audio-only, so the Absent branch is always taken.

Why it isn't reachable through rdlp's own pipeline (today)

ThumbnailStage (index 7) runs after RecodeStage (index 4) — see crates/rdlp-api/src/orchestrator/mod.rs. So rdlp never embeds art before recoding.

It is reachable when the source file already has cover art, which is common:

• an audio file downloaded from a site that ships embedded artwork
• a local file passed to process_local_file
• any second rdlp run over an already-thumbnailed output

Expected

An attached-picture stream should not count as video for routing purposes. AV_DISPOSITION_ATTACHED_PIC is the discriminator — MediaInfo would need to either exclude such streams from has_video or expose the disposition so SourceVideo::from_probe can.

Care needed: has_video is consumed elsewhere (thumbnail embedding genuinely cares that an attached pic exists), so this is likely a new field rather than a change to has_video's meaning — changing it in place risks breaking the embed paths that rely on it.

Acceptance

• --recode-video on an audio file with embedded cover art takes the audio-only route.
• The artwork is preserved or dropped deliberately, not re-encoded as a video track.
• has_video's existing consumers (thumbnail embed) are unaffected.
• Test fixture carries real embedded cover art — the --recode-video on an audio-only source fails for 12 of 16 containers (transcode path requires a video stream) #637 matrix cannot see this class.

Refs #637