Regarding the filtering, you should enable also the 2nd harmonic as it is the most prominent on your machine.
But this is not the reason for the crash. I conjecture it’s the initial tuning, which is too sloppy to recover from strong deviations. The reaction to your inputs in AltHold at 8:28 in roll and 8:58 in pitch, among other, suggests me this interpretation.
This is a larger machine than I mainly work with, but I only enter autotune if I am reasonably sure the machine can handle it, which I would not say for your case. Your machine, in my opinion, needs a bit more manual tuning, which in your case, based on in-near-hover angle deviations:
and corresponding PID contributions (roll as an example):
rather requires a proportional increase of rate P, I and D. I would probably start at a factor of 1.5 - but only after notches are fixed as noted above.
The angular acceleration measurements during the two autotune twitches also suggest increasing ATC_ACCEL_{R,P}_MAX by a factor of about 2. The measured acceleration was reported to be about 100k in the units of AC 4.6. A too low value here results in the machine very reluctant to perform sharp output changes, which might have contributed to recovery from the twitches.
After making these changes, perform a short hop (e.g. be prepared to land quickly - so maybe in Stabilize rather than AltHold) to check for oscillations. You can usually see and hear them, but after landing, it would be wise to also check PID log entries, if they exceed 0.1 in amplitude even when you are not doing much of an input, it’s not good, and PID values need to be returned back to where they were - but I conjecture it is unlikely to happen.
If no oscillations, I would test for autotune readiness as follows. Perform an AltHold test with numerous pitch, roll and yaw inputs. Inputs should preferrably be sharp (in terms of how quickly you move sticks), but of different amplitudes (how much you actually move the sticks) - this is to simulate step inputs to mimic what autotune does, but at various smaller scales to find out whether the machine can hold it. You should roughly remember your feelings (e.g. how responsive the craft felt at different stages of the test) and compare these with the logs how well the actual angle was following the desired angle (this would also calibrate your feelings). If the angles are tracked without significant overshoots, it should be much safer to run autotune.
(Sometimes autotune crashes are caused by hardware or ESC issues. I had a spectacular and terrible crash of one machine during the first ever twitch, because a gear tooth broke in one of roll servos, which resulted in constant output roll, which drove the craft into a wall and broke it into tiny pieces. But your logs don’t suggest that anything like that happened.)
