First, additional supply bypass capacitance values were used. Smaller values were added in parallel and larger values switched in. These steps had minimal effect on the impulses.
Second, the QFN pad ground contact was revisited by trying to flow solder into the via under the device. This is a tricky action as it is difficult to get good thermal distribution from the bottom with such a small via. The solder mask is tight around it on the bottom which disallows the initial formation of a solder blob/bubble. In addition, I do not have a good way of warming the part itself from the other side properly. The best technique I have found is to stick a small resistor lead wire into the via, tilt it to make contact with the via sides, allow it to transfer heat down the via and onto the pad, then apply solder and let it wick along the lead. I'm not entirely sure about the success, however, the via did fill with solder and should have made good contact with some of the QFN pad. Unfortunately, this makes no difference in the attenuator impulse noise.
Third, in an effort to understand if there is some kind of interaction between the attenuator and the switch (recall none of the other channels produce such impulse noise) various permutations of input port selection with output port selection were attempted with the input and output ports monitored (i.e. set the input RF mux to enter the attenuator but configure the output mux to select channel 0, set the input RF mux to select channel 0 while the output mux selects the attenuator, ...). In the end, even with no mux ports configured to the attenuation channel, the impulse noise is only present when the attenuator powered on.
Finally, on review of the datasheets for follow on parts to the one used, recommendations on not powering up in parallel mode were noted. This is in an effort to establish proper internal voltage values with ~400uS required from power up until latch enable (LE) for the parallel attenuation value. The design for the R board used the parallel interface for simplicity and provides a power enable to quiet each of the 4 channels while not in use. Unfortunately, as layed out LE is not individually controllable. To investigate this, the board was modified by cutting the LE trace to the attenuator, connecting it to one of the spare GPIO lines (W0 to the unused channel 0), and corresponding software updates made. (trying to get a white wire on a pulled out pad which is 0.5mm pitch from others is an exercise in patience). Those changes, with the accompanying attenuator channel results are shown below with a 400MHz input.