Interesting discussion. I have minimal contributions but I would like to add the following:
Google Play Music: Can only cast music that are uploaded to the Google account and cannot cast local files. Files that are uploaded to the Google account and are set to download for offline use can be cast. These offline files can only be played using the Google Play Music app. All use of Chromecast Audio with Google Play Music require an internet connection, and offline storage is bypassed. i.e. the CA hooks the steam from the Google internet based server and not the local offline file. It's just like Pandora so to say. This explains why local files cannot be cast. The only method that will force CA to use the local file is too fire up the Chromecast app (now rebranded as Home) and cast from there. i.e. forces the device to render the media locally, which implies transcoding, and stream using unknown encoding and unknown bitrate. Using this method will allow any local file to be cast and does not require an internet connection.
BubbleUPnP: Local files can be cast. Internet connection is not required. The file is transmitted to the CA, in that you can disconnect the CA from the wireless network and it will continue to play. This explains why seek works so well. Therefore the CA must employ a small amount of storage. The file gets transferred to the CA. I would suggest there is 128MB storage or more as I have audio podcast files that are 68MB. Using playlists, I believe that when one track finishes, the next track is transmitted to the CA and stored locally. The ability to have gapless playback would require that two tracks be transmitted and then the CA would be able to playback gapless. So it would always need the current track and next track on this local temporary storage that I am speculating about. Of course, if you use the Chromecast (Home) app to cast, instead of the cast feature of Bubble, which as mentioned has to employ transcoding, then gapless is achieved. In this scenario, I highly doubt local storage is used, nor buffering and it's probably UDP if I was to take a guess.
All of the above statements are based on testing using CA connected to an Android phone with hotspot enabled, where the same Android phone was controlling the CA.
I have not performed a packet capture to observe these behaviors at the packet level, but I have tested and measured data usage extensively in order to formulate these conclusions.
