3.3.0

Changed

• Snapshots render a third of the device's resolution rather than rendering it full size and
  shrinking the bitmap afterwards. Screen dimensions, dpi and density scale together, so every dp
  is unchanged and the layout is identical -- there are simply fewer pixels in it. Measured over
  7176 images, verification runs 27% faster and the goldens take 45% less disk, and text comes out
  sharper because nothing is resampled after rendering.
  uk.co.fractalmotion.mugshot.downscale changes it; 1 is the device's own resolution and the
  most detail available, and values below that are rejected. Reach for 1 in a module shipping
  density-qualified drawables, since the qualifier resolves from the scaled density and can select
  a different asset than the real device would. Every golden changes, so re-record.
• Pixel 10 replaces Nexus 5 as the default device, for generated preview tests as well as a
  hand-written rule.
• A module that generates preview tests gets mugshotTest<Variant>, which runs the generated test
  and nothing else. verifyMugshot used to depend on the module's whole unit test suite, so it ran
  every test the module had, and layoutlib shared a JVM with all of them. Set
  uk.co.fractalmotion.mugshot.isolateTests=false for the old wiring.

Removed

• useDeviceResolution. It only ever meant "render at full resolution", which is downscale = 1;
  the two produce byte-identical images.

Fixed

• A module using Kotlin test fixtures builds again. debugTestFixtures is not named like a test
  source set, so the preview catalogue was generated into the fixtures compilation, whose classpath
  never carries the annotations -- and the failure named that compilation rather than Mugshot.
• Adding Mugshot to a module no longer breaks that module's Robolectric tests. Layoutlib and
  Robolectric both instrument the android.* classes and cannot share a JVM; the generated
  screenshot test now runs in its own.

Performance

• The Android framework's resources are resolved once per configuration rather than once per
  snapshot. It is the most expensive thing building a session does, and the answer depends only on
  the configuration -- a matrix of previews asks for the same two dozen over and over. Worth about
  10% wherever verification is CPU-bound, and nothing where cores are spare.