Infrastructure: Qualify Windows/Win32 capture fidelity and provenance #4348

Goal

Qualify native Windows Fabric/Win32 capture and add per-artifact provenance above the implementation delivered by #4260. Parent: #4275.

Status - 2026-10-01

Open; the capture backend is implemented. PR #4306 delivered window-scoped Windows Graphics Capture, D3D11 readback, WIC PNG encoding, element crops, scale factors, and explicit minimized/unavailable-capture errors.

Do not repeat the obsolete WinAppDriver-versus-desktop-fallback experiment: the supported driver is Appium-independent. Current documentation still requires interactive capture and live-DPI qualification; a native self-test or green semantic smoke job does not provide it.

Remaining scope and acceptance criteria

• Capture representative Composition content on unlocked interactive Windows Fabric and Win32 Paper endpoints, with retained original PNGs and native bounds.
• Qualify window and element crops at live 100%, 150%, and 200% DPI on both endpoints, including edges/focus rings/overflow, scale conversions, occlusion, and secondary windows.
• Prove recognizable nonblank content and correct dimensions/crops; provider/transport success alone is not capture fidelity.
• Record per-artifact backend/source, endpoint/renderer, window/element identity, logical bounds, pixel dimensions, scale/DPI, helper/manifest/runtime identities, and capture outcome. Do not expose raw native handles as the public identity.
• Perform explicit locked/noninteractive/session-authority preflight and return distinct capture-unavailable diagnostics before attempting an invalid capture. Minimized windows remain unsupported unless a reviewed restore policy is added.
• Measure the selected hosted/interactive runner's capture reliability and keep unqualified or unavailable coverage visible.
• Verify artifact confinement, privacy/redaction, bounded output, failure behavior, and ownership-safe cleanup against real captures.
• Update driver/Storybook reproduction and artifact documentation, with relevant automated checks and evidence links.
• Screenshots remain optional visual evidence. Capture failures are explicit infrastructure/coverage results, and pixel differences never become a deterministic gate under this issue.

Evidence and coordination

• Windows provider
• Remaining qualification
• Capture implementation
• Delivered milestone: Infrastructure: Implement native Windows Composition screenshot capture #4260. Infrastructure: Add native Storybook visual comparison reports for Windows, Win32, and macOS #4328 owns baseline/candidate/diff comparison and reviewer reports, not this backend. Coordinate runner/retention policy with Infrastructure: Qualify and require native Storybook CI coverage #4349 and release evidence with Infrastructure: Complete Desktop Driver production release qualification #4345.