First-party reference app with visual and accessibility evidence

[turinglambdaai]

Goal

Make Rivet the best place to learn how a production-quality Racket GUI application is structured without turning the ten-minute starter into a kitchen sink.

The repository needs one maintained reference application whose shared application logic is Racket and whose Windows, macOS, and Linux interfaces are independently first-party native. It should teach the architecture by being executable evidence, not by presenting pseudocode.

Shape

• keep raco rivet new intentionally small and fast
• add a separate reference application with the same project layout generated users receive
• use one realistic list/detail workflow rather than a component gallery
• keep product/domain logic in Racket and UI/accessibility/lifecycle policy in WinUI 3, SwiftUI/AppKit, and GTK4

Contract coverage

The reference backend should exercise named Records and Enums, typed RPC, Event, State, request cancellation, bounded background progress, resources, and one representative system-service capability. Native clients must demonstrate UI-thread dispatch, loading/error/empty states, cancellation, and reconnect/startup failure handling.

Evidence

• guided English and Chinese walkthrough from clone to first change
• equivalent information architecture on Windows, macOS, and Linux without shared UI code
• stable accessibility labels/identifiers and keyboard navigation
• screenshot baselines plus accessibility-tree/interaction smoke where platform tooling permits
• clean-runner build, package, and verify on every advertised desktop architecture
• schema compatibility baseline and backend behavior tests
• explicit performance budget for startup and a representative 1,000-row view
• source-package smoke proves the example and tutorial ship in the public package

Non-goals

• no cross-platform widget DSL
• no replacement for platform design guidance
• no production credentials in the repository
• no claim of visual automation on a platform until CI executes it

This becomes the canonical deep-learning example; the generated counter remains the quickest proof that the toolchain works.

[turinglambdaai]

Current implementation status after the Taskboard work:

Completed evidence now in-tree:

• maintained Racket list/detail backend with separate WinUI 3, SwiftUI, and GTK4 interfaces
• English and Chinese guided walkthroughs
• Records, Enums, typed RPC, Events, State, cancellation/progress, resources, system-service integration, backend behavior tests, and a checked schema baseline
• stable WinUI automation IDs, SwiftUI accessibility identifiers, GTK accessible labels/widget names, and documented keyboard navigation
• explicit startup and 1,000-row budgets in PERFORMANCE.md / PERFORMANCE.zh-CN.md
• source-package smoke asserts the example, both walkthroughs, performance guide, schema, and all three UI sources ship and install
• clean-runner build/package/verify of Taskboard on the primary Windows, macOS, and Linux runners

Remaining before this issue should close:

1. checked screenshot baselines and real accessibility-tree/interaction smoke where runner tooling supports them;
2. if “every advertised desktop architecture” is read literally for the reference app itself, extend the Windows ARM64, macOS Intel, and Linux ARM64 gates from the generated smoke app to Taskboard too (those architecture gates currently prove Rivet's generated app path, while Taskboard is proven by the primary platform runners).

Keeping this open for those evidence gaps rather than treating source-level accessibility identifiers as an automated accessibility test.