This one is gonna take some explaining:
When in resist fingerprinting mode, Firefox exhibits some funny behavior: when we ask for the the list of media devices, it gives us fake device IDs. But when the js-sdk requests a stream for any of those devices, Firefox associates the stream with the real device ID.
Now, in order to get the names of devices included in their metadata when you query the device list, you need to be holding a stream. For this reason, useMediaHandler was set up to reload the device list whenever matrix-js-sdk got a new local stream. But because of the inconsistency in device IDs, it would enter an infinite cycle telling matrix-js-sdk to request a stream for the fake device ID, but with matrix-js-sdk always responding with the real device ID.
I already wasn't happy with useMediaHandler's use of @ts-ignore comments to inspect private js-sdk fields, and in the meantime we've come up with a simpler function for requesting device names, so I decided to refactor useMediaHandler to use it instead. Importantly, it doesn't break in resist fingerprinting mode.
This created a new UX issue though: now, when on the lobby screen, useMediaHandler would request microphone access so it could get device names, followed immediately by a *second* pop-up for the lobby screen to request camera access. That's 1 pop-up too many, so I changed useMediaHandler to only request device names when a component is mounted that actually wants to show them. Currently, the settings modal is the only such component, and users normally only open it *after* granting full audio/video access, so this solution works out quite nicely.
Previously we were showing a combination of audio and video status icons on people's name badges, which meant there was no way to tell whether someone who had their video off was muted or not. The designs call for only microphone icons to be shown here.
My dev env suddenly, with no apparent prompt, went into a mode where
it wouldn't display nay video tiles which was because they were 0x0
in the top left corner, which in turn was because the ResizeObserver
was never returning the actual bounds of the video tile container.
As per comment, this uses the native impl in preference to the ponyfill,
although in practice it looks like all our target browsers should support
it, so perhaps we could just remove the ponyfill entirely.
