EKF based SLAM, marginalization and key-frames

I have a couple of tightly related questions regarding EKF-based visual SLAM. It is common in the Kalman based VSLAM litterature to marginalize 3d points and past poses. Condsidering the EKF as a bayesian filter, one can write its prediction step

$\tag{1} P(x_t|z_1,...,z_{t-1})=\int_{x_{t-1}} P(x_t,x_{t-1}|z_1,...,z_{t-1})dx_{t-1}$

with $x_i$ and $z_i$ respectively denoting states and measurements.

The right side of (1) clearly marginalizes the past state. So, here are the questions:

• I understand that in EKF based SLAM, the current state needs to contain new features (e.g. newly triangulated points) but does not contain those that are not viewed anymore. So, $x_t$ and $x_{t-1}$ are not exactly the states of the same system. However, I don't see any other way of marginalizing older feature points than using (1). Is this the manner in which older points/features are magrinalized, or are there other intermediary steps?
• In Key-frame based Bundle Adjustment approaches, one discards redundant information by dropping key-frames (although nowadays, more and more sytems take the filtering approach here and marginalize instead of discarding). I was wondering, since I couldn't find anything relevant, if there exists Key-frame EKFs? Is there an inherent property of the EKF that makes it incompatible with Key-frame selection?