Golden Files and Device-Class Buckets

Golden baselines multiply quickly once you respect font scaling and locale. We bucket devices into “compact ja”, “compact en”, and “tablet en” only—adding a bucket requires deleting another to keep storage flat.

CI uploads failures as triptychs: expected, actual, and diff mask tinted for daltonic reviewers. The triptych naming matches the branch slug so artifact browsers stay searchable.

Flaky renders often trace to shader warmup. We pre-roll a warm-up frame in test harnesses before capturing goldens, documented in the track workbook so students do not mistake it for cheating.

When a baseline changes legitimately, the PR must include a one-paragraph rationale referencing the widget diff—not only the image swap—so future auditors understand intent.