The ffmpeg command line for a conversion: pure, so what is run can be read and tested without running it.

Video copied as it is; audio copied when it is AAC already, else AAC stereo (a browser plays stereo regardless, and TrueHD 7.1 cannot be put in AAC as it is). Fragments at keyframes, at least two seconds long, in the MP4 muxer's own form. One audio track goes into media.mp4 with the video; several each get their own audioN.mp4, so the player can switch between them as HLS renditions. Text subtitles come out as WebVTT files.

HEVC goes through MPEG-TS on the way: a release often keeps the parameter sets inside the video rather than in the container, and an MP4 written straight from it has an empty hvcC that nothing decodes. Read back from TS, ffmpeg's parser puts them where hvc1 needs them.

delay_moov holds the header back until the first fragment: the AAC decoder configuration comes out of the TS stream with its first packet, and a header written before that has none, which Safari refuses to decode (ffmpeg plays it anyway, so a local check does not show it).

A part begun later in the file (-ss, which lands on the keyframe before) keeps the file's own clock: -copyts leaves the times as they are and frag_discont makes the first fragment carry them, so the video, each audio file and the subtitles all agree on where they are — without it the MP4 muxer counts each file from its own first packet, and audio that begins a second after the keyframe comes out a second early. The TS stage would add its usual 1.4 s to everything; -muxdelay 0 stops it.