quit-mnemonic reftest sometimes fails with an extra row of pixels rendered, dependent on installed fonts
Steps to reproduce
Build GTK on Debian unstable (with packages from experimental where required, e.g. Pango during the current pre-release freeze) and run reftests with --setup=x11. Do this both with and without the additional packages that are pulled in by building the gtk-doc documentation.
For context, our official autobuilders do one build with documentation disabled for each supported architecture, plus an additional build on an unspecified architecture - currently always x86_64 - to build architecture-independent packages like the API documentation. This was a test failure on the x86_64 build that was not also seen on the documentation build.
Current behavior
Reference: 
(32x31 px)
Actual output: 
(32x32 px)
Difference: 
(one row was added at the bottom)
Expected outcome
Identical rendering.
Version information
Debian unstable rolling release, Pango 1.48.0-1, Cairo 1.16.0-5, Cantarell font 0.111-3.
The bad build had the following additional packages compared with the good build:
- fonts-liberation_1:1.07.4-11
The good build had the following additional packages compared with the bad build:
- docbook_4.5-6
- docbook-to-man_1:2.0.0-45
- gtk-doc-tools_1.33.2-1
- libcairo2-doc_1.16.0-5
- libcmark-gfm-extensions0_0.29.0.gfm.0-6
- libcmark-gfm0_0.29.0.gfm.0-6
- libglib2.0-doc_2.66.7-1
- libosp5_1.5.2-13+b2
- libpango1.0-doc_1.46.2-3
- opensp_1.5.2-13+b2
- pandoc_2.9.2.1-1+b1
- pandoc-data_2.9.2.1-1
- python3-lxml_4.6.2-1
- python3-pygments_2.7.1+dfsg-1
- ttf-bitstream-vera_1.10-8.1
All other packages were at identical versions in the two builds.
I wonder whether this was triggered by having fonts-liberation vs. ttf-bitstream-vera, or whether that's just coincidence. Am I right in thinking that Adwaita consistently uses Cantarell, and should not be rendering differently because other fonts are installed?