* Also fix bug where you couldn't picxk a different server if
you were already registered as a guest (because it still sent
the access token which the new server rejected)
* Propagate errors from UI auth back to registration so it goes
back to the registration screen
- Shift to the left _before_ adding an avatar so that there are always `MAX_READ_AVATARS` visible, instead of there being `MAX_READ_AVATARS + 1` avatars displayed following the first "collapse".
- Use `right` instead of `left` so that double-digit remainders don't get overlapped.
- Use avatar initial instead of "R" or "?"
- Use Fuse.js to do case-insensitive fuzzy search. This allows for better sorting of results as well as search based on weighted keys (so userId has a high weight when the input starts with "@").
- Added debounce of 200ms to prevent analysis on every key stroke. Fuse seems to degrade performance vs. simple, non-fuzzy, unsorted matching, but the debounce should prevent too much computation.
- Move the selection to the top when the query is changed. There's no point in staying mid-way through the items at that point.
For E2E rooms, display "Send an encrypted message…" otherwise display "Send a plaintext message…" as the placeholder for the input box in [old] message composer.
This adds a 5 minute auth cache to speed up the process of deleting
old devices. It has the following nastinesses (mainly due to being
written on a flight whilst juggling kids):
* the auth cache is done as context attached to MatrixChat.
one could argue that it should be per-client instead, but we don't
yet have multiple clients.
* the auth cache is only maintained currently in DevicesPanelEntry
(i.e. set & invalidated). One could argue that it might be better
maintained in InteractiveAuth.js or a dedicated cache object
abstraction, but given the only use I can think of is when managing
devices, perhaps this is good enough for now.
To fix https://github.com/vector-im/riot-web/issues/2916, force the checking of scroll position by calling _onWidgetLoad (might need renaming...) when a MELS is expanded/contracted.
Also use an keying scheme for MELS that doesn't depend on the events contained, but rather a simple incrementing index based on the order of the MELS as it appears amongst all MELS.
