your third paragraph does illuminate the situation somewhat.LEM and DA/SA add more "ifs". If you forget to decock, you must also forget to place the thumb to the hammer because if you place the thumb to the hammer, you will feel a cocked hammer, which will then likely prompt you to decock. Even if it doesn't, and you holster a cocked pistol, there has to be more coincidences. Also, you don't really have to "feel" the hammer, you push the pistol into the holster via the hammer, the whole gun will 'jam' and not go into the holster if something tries to pull the trigger but is blocked because the trigger is being pushed forward into the holster by the hammer, which the trigger is trying to pull back if it is fouled.If the operator forgets to decock (that does happen even during non-stressful encounters) and tries to reholster, it is even less safe than a SFA
Striker:
#1, there has to be something to foul the trigger. (including, say, your finger)
#2, you have to fail to notice it (if there's nothing there and you fail to notice, no bad result).
So, striker has a double coincidence to net a failure.
LEM:
#1, there has to be something to foul the trigger.
#2, you have to fail to notice it
#3, you have to fail to control the hammer (the first two both have to met in order for failing at #3 to cause an incident).
DA/SA:
#1, foul the blahblah
#2, miss said blahblah
#3, you have to fail to decock the pistol
#4, you have to fail to control the hammer
3 and 4 have some interaction here as well.
you have to multiply the individual chance of each failure, the more layers you stack on the less chance there is for them to all happen together. I heard it referred to elsewhere on this board as the swiss cheese model, each layer has holes but the more layers you have the less likely they all are to line up.
As you alluded to, the gunsite data really can't say anything without knowing the sample before the accidents were selected out.