0.41.1 - a fix that was there and did nothing
The Waveshare ESP32-P4-WIFI6 pulls its six SDIO lines up through 51k where Espressif ask for 10k. One of those lines is how the WiFi co-processor says it has data, and a weak pull-up is enough to lose that. The ESP32-P4 has pull-ups of its own inside, so 0.40.1 switched them on for that board.
It did nothing. I had written the function and the setting, and never called the function. The person who reported the stall flashed it, saw no change, went and read the code, found the missing line, and sent the patch. With the call in place: resets, reassociations, ten disconnect cycles, a hundred requests in a row, no errors, at the full bus speed. So 0.41.1 is his line, plus one of mine - the components I maintain here now fail to build if they contain a function nothing calls. The warning had been sitting in the build log all along.
Then, while checking that the release changed nothing else, the device rebooted on me. Closing a tab that was watching the H.264 stream could take the whole device down: the task pushing video was still writing a frame into a TLS session the web server had already freed. The crash landed later, in whatever asked for memory next, which is why the same bug had looked like three different ones over the past week - a dead web interface, a failed update, a stream that stopped. A build with heap poisoning on named the freed object at the moment of the bad write. Sends and session teardown now take the same lock: twenty-four forced disconnects in a row, no crash, where the same build without the fix died on the second.