Merge discrete mono audio tracks into the rendition instead of dropping them

[bjay-wk]

What

create_rendition built its ffmpeg command with .codec_audio("aac") and no -map arguments, relying on ffmpeg's default stream selection — which writes exactly one audio stream to the output.

Avid/broadcast MXF masters conventionally carry the stereo program as two discrete mono tracks (Track 3 / Track 4), so half the program audio was silently lost in every rendition made from one.

This mirrors the backend fix in video_processing_worker's downscale_video (tellers-backend #1454).

Repro

Ran the real create_rendition on a 1080p25 DNxHD MXF with two pcm_s24le mono streams (440 Hz / 880 Hz):

	audio stream in rendition
before	aac, 1ch, mono — half the program lost
after	aac, 2ch, stereo

Channel-splitting the fixed output and running a DFT confirms L = 440 Hz (track 3), R = 880 Hz (track 4) — both tracks present, in the right positions.

How

New build_mono_merge_filter(probe) returns a -filter_complex graph only when there is more than one audio stream and every one is mono:

-map 0:v:0
-filter_complex "[0:a:0][0:a:1]amerge=inputs=2,aformat=channel_layouts=stereo[aout]"
-map [aout]

Two implementation notes:

• -map 0:v:0 is required. A complex filtergraph disables ffmpeg's automatic output stream selection, so the video stream must be mapped explicitly once -filter_complex is present.
• aformat=channel_layouts=stereo rather than -ac 2. -ac on a filtergraph output is unreliable, and the in-graph form also downmixes correctly with more than two mono tracks (capped at MAX_MERGED_AUDIO_STREAMS = 8).

The TUI shows an info line ("Source carries discrete mono audio tracks; merging them to stereo") when the merge kicks in.

Deliberately narrow

• Merge only when every stream is mono. A master with a stereo mix plus an alternate-language stereo pair would otherwise get two programs blended together. ffmpeg's default already picks the stream with the most channels there, which is correct — left alone.
• Single-stream sources are untouched.
• Streams missing a channels field are not merged — without it we can't distinguish discrete mono tracks from alternate mixes, so we keep the default.
• Speed-corrected (half-rate MXF) path skips the probe, since extract_video_essence writes -an and there's no audio to merge.

Testing

• 8 new unit tests: the merge, input cap, single-track, multi-stereo, mixed-channel, audioless, no-streams, and missing-channels cases. All 15 tests pass.
• Verified end-to-end against real ffmpeg (single stereo aac copy-through path and the two-mono merge both checked).
• cargo fmt/clippy findings in this diff's file are all pre-existing drift outside the changed hunks.

🤖 Generated with Claude Code