fix(ios): use the resolved display for screenshots and visual verification

[thymikee]

Part of #2725. Investigation evidence and reproducible probe: #2727 (comment).

Resolved design / dependencies

Use the resolved app window's screen, not XCUIScreen.main, for iOS app-screen capture and visual verification. Resolve app.windows.firstMatch.frame before obtaining that window's private screen; screenshot the resulting XCUIScreen. Open, half-open, closed, a single-panel control, and a native fixture alert were measured successfully. Do not choose by a hardcoded display number or the first screen in XCUIScreen.screens.

Prerequisite: PR #2724; implement on its branch with an explicit PR base, or after it merges. No unresolved design decision remains. Consumer-level checks below remain required implementation validation; the spike proves the capture primitive, not the complete migration.

Small owning interface

Introduce one iOS capture helper, conceptually captureResolvedAppScreen(app:) -> CapturedAppScreen, returning an upright image and actual display ID, encoded pixel width/height, and pixels-per-point. Keep the private window/screen lookup exception-safe in the native XCTest bridge. Reuse the window-resolution order established by #2724; do not maintain independently drifting display resolvers or introduce a session-wide display cache.

Normalize the image by drawing it in its oriented logical size with its actual scale before PNG encoding. Measured open Duo: window 951x669, UIImage logical size about 950.667x668.667, scale 3, raw CGImage/PNG 2006x2852, UIImage orientation right; upright drawing yields 2852x2006. Do not infer orientation from raw CGImage dimensions, rely on PNG metadata being respected by every consumer, or stretch the one-pixel XCTest discrepancy to CoreDevice's nominal 2853x2007 size.

Wire and failure contracts

Carry screenshot-only metadata with the runner screenshot result (e.g. screenshotMetadata: {displayID, pixelWidth, pixelHeight, pixelsPerPoint}), alongside the existing file/base64 payload. Declare it at the owning cross-language contract, not by per-consumer casts. Host normalization uses those actual source facts, never SIMULATOR_MAINSCREEN_SCALE for an image from another panel.

An unresolved/invalid app window or screen fails required screenshot capture with a typed reason, without selecting main. Optional visual checks return their existing unavailable/unknown result; no sample is not evidence of stability or navigation change. Do not make an observation activate an app merely to obtain a screen; use the caller's intended app context.

Bounded migration

• Route the iOS screenshot case, captureRunnerFrame, and navigation fallback visual capture through this helper. Keyboard settling already consumes frame capture and keeps its separate keyboard visibility check.
• Update packages/platform-apple/src/core/screenshot.ts runner fallback normalization to consume source metadata.
• Preserve macOS/tvOS behavior. Do not redesign gestures, text entry, host display inventory, or recording touch overlays (Recording touch overlay exports at ~220x480 or all-black whenever touches are drawn #2707).
• Follow the repository split-before-behavior rule for oversized sources, with mechanical structural changes separated from behavior.

Acceptance and validation

• Force simctl screenshot failure: runner fallback captures the visible inner content upright at correct density.
• Navigation change/no-change controls and keyboard settling observe the lit display. A failed sample cannot satisfy a positive visual assertion.
• Test open, half-open, closed, and a single-panel control; preserve a visible native overlay in the full-screen capture. Record pose, display, image orientation, actual pixel dimensions, scale, and app state.
• Repeat the pose matrix on the implementation PR head. During the spike, moving Device Hub from the second host display to the main display cleared device-hub-window-missing; document this troubleshooting detail rather than treating it as a missing simulator.
• Add focused contract/encoding tests, native builds and XCTest selection checks, affected gates, and local semantic evidence. Update ADR 0025 only to claims proved by those checks.

Ready for implementation with the explicit #2724 branch/merge prerequisite. Follow AGENTS.md; keep one task per PR and clean up manual sessions.