Maintenance 2026 - #85

- previously they would never interrupt CI, even with a failure

- also fixes local env detection in cli when run from repo

- allows the test suite to run in CI

- it was using map() which started being lazy in python3 and was failing to invoke the callback

- comparing to a tuple was always a non-match

- add missing dunder methods and drop otiose ones from MagicNumber (the base class for Dimension & Unit)
- add unit tests exercising numeric comparison ops
- fixes #67

- previously the vector would be rasterized at the width/height of the PDF or SVG, then that bitmap mask would be scaled
- now the vector image is scaled to the size it will occupy on the canvas *before* being rasterized, so the masking is still pixel-based, but will no longer be super-low res
- partially addresses #83 (though it doesn't provide the vector-based mask the user exptected)

- initial migration uses some runloop trickery in the EditorView.source getter to stay compatible with the synchronous file-saving api
- web inspector is now only enabled by default on dev builds (though it can be added to dist builds with a defaults write to `WebKitDeveloperExtras`)

- removed the need for the runloop trickery by keeping a cached copy of the document in _last_source (updated via the undo-manager callback)

- source file is app/art/PlotDevice-app.icon
- Assets.car can be rebuilt with `make.py icon`

set popup background via fixed selectors, not inline styles

- accept fill()-style gradient args in stroke()
- fill(), stroke(), and background() inherit color()'s mode, range, and space
- unknown color keywords raise instead of being dropped, as do extra arguments after an image

- the Colors pop-ups and Proof Colors include macOS's RGB spaces
- improve errors for unknown/invalid colors= values

- includes protanopia, deuteranopia, tritanopia, and their anomalous forms
- hovering the menu item previews its profile
- RGB profiles now keep transparency rather than a white matte

- rendered in P3 and then converted to requested primaries and matrix by AVFoundation

- move culling and layer-sizing state into FrameBounds, with each grob measuring its own _render_box()
- scaled-up text with effects no longer clips its glyphs' overhangs
- Text.flow() properly keeps the original size unit even if the canvas's changes before flowing

- FrameLayers collects the canvas's CA layers and tiles
- ViewTracker monitors the viewport's scale and position changes
- with export() only renders on a separate thread if the GIL is off

- ViewTracker applies zoom on its own, so it just needs the dpr

- if a nested render call occurred (e.g., for an embedded image/canvas), keep the lock until that completes

- each grob/frob now makes its own clone (isolated from ctx state for the rendering pass)
- if there aren't any variables referencing the grob, the copy is skipped

- stop using NSImage.lockFocus in favor of thread-friendly CGImage

- make sure ProRes frames are drawn in the movie's color space
- pass codec & colors as names rather than indices

- also switch to using NSApp() accessor rather than the global

- don't wildcard all of Foundation just to confirm PyObjC is present
- defer building the font family list until first use
- don't import all of CoreServices just to get the image type UTI strings
- use html.escape rather than pulling in xml.sax.saxutils (and transatively urllib.request)
- reduces `import plotdevice` from 950ms to 210ms

- now checks for exact family matches before switching to fuzzy (and avoids sending weight names to the family matcher)
- only construct the full family name list on the first fuzzy match request

- an image() can now be passed to stroke(), fill(), or background() and will be used as a repeating pattern
- the x/y/width/height values passed to image scale the tile and position it relative to the canvas origin
- svgs and canvas-backed patterns now remain vectors (though text fills still rasterize)
- requires pdf and eps exports to hold the render lock while CG reenters to draw the tiles

- previously the script's canvas reference could still affect a committed Frame's clip/mask shapes

- otherwise a canvas's default P3 space could be used for a pattern in a CMYK export