deterministic-expecteds

Pin one shared reference clock for all date formatting, fake every nondeterministic input such as network, randomness, locale, and viewport, and load real fonts in the render harness. Use fixed-duration pumps rather than waiting for UI to settle.

Feed fake case data to the real render function in a jsdom document seeded from the real HTML, fold CSS in as inline styles, rasterize with satori and resvg, then compare with pixelmatch at zero tolerated diff. This is deterministic and dependency-light.

Yes, Flutter widget tests support golden files via matchesGoldenFile, pumping the real app shell against a scripted fake world. Use --update-goldens to refresh, and load fonts from the FontManifest including icon fonts.

Test environments default to a glyph-less stub font, so text renders as boxes. Load the product's bundled font families and icon font in the harness, and pin the family explicitly on styles that do not inherit it, such as button text.

No, comparison should be pixel-exact with zero tolerated diff ratio. A tolerance invites unreviewed drift; if a platform renders unstably, fix the determinism through fonts, clock, and fakes rather than loosening the threshold.