IIRC (it's been a while), the industrial_robot_simulator (IRS) does not perform the smoothing you seem to be describing. The code is here, it's not too complicated.

You write "recorded path of the real robot", but then you mention the IRS, which is still not an actual real robot.

If you are seeing smoothing with actual real hw, then it would most likely be the controller executing your trajectory (ie: the industrial controller) that smooths it. As you already noticed in #q334069, trajectories do not always exhibit the correct level of continuity, so (industrial) robot controllers need to do some post-processing themselves (or pre-processing, depending on how you look at it) before executing them.

If you can provide some more details on your actual hw, I may be able to point you to some literature.

Oh, and:

Is the trajectory time-parameterized before it gets passed into the robot simulator?

yes. As one of the last steps in MoveIt's planning pipeline, the trajectory (which is really more a path up till that point) is time parameterised. See here for all the default planning adapters applied by the planning pipeline.

I simulate robot movement by launching my simulation setup with

moveit_planning_execution.launch sim:=true

A reproducable, extreme example would be to plan a cartesian triangular shaped trajectory with three waypoints in a plane. While the trajectory points at corners make an instantaneous 90° direction change turn, the recorded and with FK transformed /joint_state messages show a curved corner interpolation. I assumed this to be a result of the industrial_robot_simulator?

How did you "plan a cartesian triangular shaped trajectory with three waypoints in a plane"? Using MoveIt's computerCartesianPath(..), or stitching together three OMPL motion requests? The former should result in straight paths (provided sufficient resolution and bar any interpolation by the controller (in real hw)). The latter will most likely not.

I used computeCartesianPath() to ensure I get a 90° kink in the trajectory message. I then assigned this trajectory to a plan and executed the plan under the sim:=true launch file as described earlier. For this reason, since I am not using real hw, I can only see the /joint_states that the industrial_robot_simulator publishes. Using FK and plotting these points over the trajectory line, I see that the "real" robot branches out of the straight line before the kink to impose a radius, rather than an instantaneous switch in direction. I tested this with eef_step = 0.005 and 0.01.