[houlianpi]

Changes

• Add a managed Android execution reference with platform/context/element identity, input/persistence and failure-specific recovery guidance.
• Share control-first/visual-fallback observation rules across Android/UiAutomator2, iOS/XCUITest and native macOS/Mac2.
• Require changing-state evidence and retain usable sessions while verification is incomplete.
• Add synthetic visual fixtures with actual image-delivery, pixel-identity, pointer-hit, ordering and cleanup verification.
• Normalize actual Copilot image-delivery fields without treating source-only images, artifact links or display metadata as delivered observations.
• Update managed-reference integrity coverage, experiment/HANDOFF/changelog and bounded evidence records. Runtime, setup and pinned official inputs are unchanged.

Verification

• Focused regressions: 193 passed.
• Lock consistency, Ruff formatting/lint and Mypy passed.
• Two independent policy/correctness and safety/image-delivery reviews: no remaining reportable findings.
• One fresh installed-wheel native Agent run: all six control/visual scenarios passed; control flows used zero screenshots, visual flows delivered matching pre/post images and hit the target. All sessions were deleted, supplied servers preserved and isolation unchanged.
• Evidence: evidence/appium-skill/observation-policy-20261105.json; prior failures remain retained, not rescored.

Synthetic W3C endpoint behavior only: no real device/driver qualification or Android World score improvement is claimed.

Related: #98. This does not close the broader release/qualification work item.

[houlianpi]

Clean-commit integration verification: detached worktree at 8509bf9, full pytest suite 797 passed in 100.82s. This checkout excludes the pre-existing staged evidence deletions and local .gitignore change. The earlier installed-wheel native Agent proof remains 6/6 synthetic scenarios passed; no physical platform or Android World qualification is claimed.