I took a look and it seems nothing wrong with PlzNavigate in this case.

For render-initiated navigation, we will run decidePolicyForNavigation twice:
- first, when the navigation is initiated, it is called and decides to send the request to the browser.
- second, the browser replies with the message FrameMsg_CommitNavigation, which starts a load and eventually calls into decidePolicyForNavigation again.

The behavior seems reasonable. It feels hacky to try to do some special handling to skip calling decidePolicyForNavigation a second time.

Camille: Is there a standard way to flag these tests as "skipped as intended"? We could update the expectation when we remove the non-PlzNavigate path, I think.

Unfortunately we don't. Since this is an issue in other tests as well, I'll start a conversation with the blink team to see what can be done.

Thanks, Camille!

I didn't know that it also broke tests other than
  [virtual/mojo-loading/]http/tests/misc/policy-delegate-called-twice.html

The following revision refers to this bug:
  https://chromium.googlesource.com/chromium/src.git/+/a6bd9edba94b179253c09f99ced143a0cf195e74

commit a6bd9edba94b179253c09f99ced143a0cf195e74
Author: yzshen <yzshen@chromium.org>
Date: Tue Feb 07 06:02:25 2017

PlzNavigate: avoid checking navigation policy twice for render-init requests.

Some layout tests expect that navigation policy is checked only once, although
we should probably change the expectation and avoid this hack once PlzNavigate
fully takes over.

BUG= 673765 

Review-Url: https://codereview.chromium.org/2680453004
Cr-Commit-Position: refs/heads/master@{#448559}

[modify] https://crrev.com/a6bd9edba94b179253c09f99ced143a0cf195e74/components/test_runner/web_frame_test_client.cc
[modify] https://crrev.com/a6bd9edba94b179253c09f99ced143a0cf195e74/third_party/WebKit/LayoutTests/FlagExpectations/enable-browser-side-navigation