Thanks for suggestions, I didn't know that marginal quality external I2C modules can cause FC reboots... Anyway, I'll do a bench test.
I'm aware that Attitide mode is not safe enough when using INS13. I also had similar experiences. Setting the dynamic model of GPS to 4G worked for me to avoid rapid jumps in GPS position. The tradeoff is more uncertain "Position Hold" - the position wanders. For me that's OK, I use this quad to fly in RATE mode and the GPS Assisted modes are the "lifebelt" - failsafe is set to RTB/Land if something wrong happens with radio control or video link
Reboot during flight was what I first suspected, although I thought it might be a BEC failure. But after examination of video records from HDCam and OSD, and GCS telemetry log, it's evident that there wasn't a reboot situation or some 5V power outage. All the data flowed continuously and no evidences for reboot present, based on my understandings.
I2C wires are short enough, no more than 20cm. I even plan to shorten them more, because it turned out that the frame is built good enough to allow GPS/MAG to be mounted directly on the upper plate, without any additional stand (twisted all power wires, most of them sandwiched between lower two plates, that I separated with additional spacers to 8mm gap). I made numerous tests and autonomous features work flawlessly in this configuration, no "Toilet Bows" at all, good Position Hold, etc. Telemetry sometimes shows yellow MAG alarms during flight but it seems EKF handles these deviations good enough. So I can shorten the GPS/MAG wires even more.
Already updated to RC1, that changed nothing - last crash was with FC rinning 16.09-RC1.
I suspect the MPU6000 chip to be the culprit so I ordered a new one from Drotek (ebay) and will try to replace it.
Meanthile I replaced the Revo on the quad with a new one, and I'm waiting the weather to permit some extensive testing.