Okay, so it's not quite what I thought.

These pages are due to prerenders, and I've confirmed that they're not recording FCP and the like as the observing clients are cleared even though the tracker lives on.

Yeah, that's right, prerenders should be ignored here:

https://cs.chromium.org/chromium/src/chrome/browser/page_load_metrics/page_load_tracker.cc?sq=package:chromium&dr=C&rcl=1481717748&l=308

There's a change in flight to change the prerender policy slightly:
https://codereview.chromium.org/2423383002

If you do see any PageLoad.* histograms (other than perhaps PageLoad.Internal) logged for prerenders or omnibox results currently, that's definitely a significant bug & I'll fix it right away.

RE: omnibox requests, I don't know that we currently explicitly filter those out. We should make sure they don't get included. We have some logic to ignore NTP requests that may make sense to augment to ignore omnibox requests as well.

More thought here:
omnibox requests that do cause a new page load, such as a SRP load, should be included.

Omnibox requests that incrementally update the SRP as a user types additional text into the omnibox should not be included.

Closing this as WontFix as I don't think there is any bug here.

I'll note though that most of my page loads are prerenders, so the navigations we are recording are quite biased to those that I visit less frequently, and are less cached. 

Er, Bryan, did you want to leave this bug open for the SRP pages?

After debugging this a bit more I think everything is working & we can close it out.