0.31.0 - updates that take on the first try
Some people updated, watched the device come back, and found the old version running. It looked like the new firmware had been rejected and rolled back. It had not.
The chip could fault part-way through its own restart. The boot that followed read the new firmware wrong and got nowhere; a few seconds later the watchdog reset the board properly, the old version came up, and from outside that is indistinguishable from a rollback. The fault is in ESP-IDF's restart path on chips with external RAM - it never needed an update to happen, a plain restart could do it - and it is gone in ESP-IDF 6.1, which the releases are built with now.
Updating to this version is enough, and no cable is needed. But the restart that installs it is still done by the old firmware, so if the device comes back on the old version, run the update once more. From this version on it stops happening. Found from a serial log by @petrn in #22.
Uploads that died at a few percent. Over HTTPS a quiet moment on the socket arrives as a raw "nothing yet" from the TLS library, not the timeout the code was watching for, so it was read as a broken connection and the whole upload was thrown away. Live video can starve a socket for half a minute, so updating with the console open could fail every time. A silence is now treated as a silence.
The rescue hotspot also learned to show a join code on the mono OLED, at whatever size the glass can hold.