After a good deal of experimentation, I was able to get SteamVR to run.
My laptop has an integrated intel GPU as well as an nvidia discrete GPU.
The difficulty in making this work has to do with exactly what vulkaninfo reports, as well as how the vrcompositor handles the detection of direct mode vs extended mode.

If vulkaninfo reports GPU 0 as the intel GPU, the error described above occurs.
By removing the intel json files from my vulkan/icd.d directory, I coerced vulkan to only report the NV gpu. This resolved the segfault, but vrcompositor still hangs, due to trying to enable direct mode on a screen that is not attached to the NV gpu.

In the end, here is what is needed to make steamVR work:
First, make sure the HMD is enabled as a second desktop screen.
You may need to do something like:
xrandr --output HDMI-1-2 --set non-desktop 0

When vrcompositor starts up, it needs to see zero vulkan-enabled GPUs.
This causes it to skip the direct mode attempts, and immediately fall back to extended mode (HMD is on desktop).
As soon as it DOES fall back to extended mode, the nv GPU needs to be vulkan-enabled so that the compositor can draw to it.

I was able to accomplish this with the following bash script:

#!/bin/bash
####
##  WARNING! Don't use this script.
## In all likelihood it will damage your vulkan capabilities, and may do far worse things.
## Be your own sysadmin, and fully understand any script before executing it.
####

if [ ! -d /usr/share/vulkan/icd.disabled ]
then
  echo "Disabling intel GPUs"
  mv /usr/share/vulkan/icd.d /usr/share/vulkan/icd.disabled
  mkdir /usr/share/vulkan/icd.d
fi
if [ ! -d /etc/vulkan/icd.disabled ]
then
  echo "Disabling nvidia GPUs"
  mv /etc/vulkan/icd.d /etc/vulkan/icd.disabled
  mkdir /etc/vulkan/icd.d
fi
echo "Removing old files if necessary"
rm /etc/vulkan/icd.d/*
echo "Creating pipe"
mkfifo /etc/vulkan/icd.d/pipe.json
for x in {1..2}
do
  echo $(echo $x > /etc/vulkan/icd.d/pipe.json)
  echo "Read $x times"
done
echo "enabling nv"
cp /etc/vulkan/icd.disabled/nvidia_icd.json /etc/vulkan/icd.d/
echo "Removing pipe"
cat /etc/vulkan/icd.disabled/nvidia_icd.json > /etc/vulkan/icd.d/pipe.json
rm /etc/vulkan/icd.d/pipe.json

Run the script as root immediately before launching steamVR.
The script works by moving around the icd json files to enable and disable GPUs for vulkan, and handles timing by using a fifo to catch attempts to read icd json files. Experimentation found that if all GPUs were disabled for the first two reads, and the nv GPU was enabled for the third and subsequent reads, it allowed steamvr to launch using extended mode, rendering with the discrete gpu.
The compositor window did initially appear on the wrong screen, and I needed to manually send it to the HMD screen.

Final thoughts:
This setup is of course not supported by Valve.
It could be, however, with two fairly small configuration options.
First, allow a configuration option to completely disable attempts at using direct mode.
Second, when enumerating vulkan-enabled GPUs, either intelligently choose the nv GPU over the intel GPU, or allow a configuration override to select which GPU to use.

Thanks,
John Hanely