Fix iOS 27 reader defects and speed up CI

[onevcat]

Summary

Fixes two iOS 27 defects in the iOS mirror reader, and makes CI faster and cheaper: the mirror clients run only when they change, CI runs no iOS UI tests, and the Mac app job gets a working compilation cache fallback.

iOS 27: "Load Earlier 200 Lines" does not respond

On iOS 27, a Button in the same stack as .textSelection(.enabled) text rows gets no taps. The history view put the button as the first child of the LazyVStack that holds the selectable lines (a plain VStack also fails; removing the text selection makes it work). The button now sits outside the stack of selectable rows. Text selection is unchanged.

iOS 27: live reading position moves after a pane switch

The reading view restored an absolute ScrollPosition(y:), but LazyVStack estimates the heights of rows above the visible area that it has not laid out. On iOS 27 the same offset then shows different text (17–31 pt off; exact on iOS 26.5).

The live reader now stores an anchor: a document row and the distance from the top of that row to the top of the visible area.

• MirrorDocumentView is one flat LazyVStack of rows: a code or table block, a short text block, or one 32-line / 4096-scalar chunk of a long text block. Each row reports its top in the scroll content space.
• On appear, the reader scrolls the anchor row to the top, then applies the offset in the row when the row reports its laid-out top. If the row no longer exists, it goes to the bottom.
• MirrorSession.liveReadingAnchor replaces liveReadingOffset; it is @ObservationIgnored because the reader writes it on every scroll.
• Rendering stays lazy. The 10,000-line first-viewport performance test is not slower (0.323 s against 0.349 s on main).

The history view still restores an absolute offset (single-line rows, so the estimates are exact for the fixture).

iOS mirror tests

• iPhone-only layout: the only phone-specific UI test checked the userInterfaceIdiom == .phone branch of the initial split-view layout, which cannot run on iPad. That decision is now ContentView.initialLayout(for:) with a unit test, and the UI test is removed. The other "iPhone" UI tests pass on iPad, so all UI tests run on one iPad simulator.
• testRemoteScrollShowsLoadingAndKeepsHistoryIndependent was flaky: the fixture Host confirmed the scroll after 2 s (and the app gives up after 5 s), and on a slow runner each XCUITest query took about 3 s. The fixture now holds each scroll until the test posts a Darwin notification (--mirror-ui-scroll-hold-fixture), and its confirmation timeout is longer. MirrorSession takes the timeout as a parameter (default 5 s).
• make test-mirror-ios runs the unit and UI tests on iPad; make test-mirror-ios-unit runs only the unit tests. Both print a timestamped heartbeat and keep the raw log next to the result bundle.
• select_ios_simulator.py works with the Python 3.9 that macOS includes.
• Two compiler warnings are fixed (resignFirstResponder() result; a main-actor UIKey init in a test UIPress override).

CI

Workflow	Runs
test (Mac app)	Every change except changes only to Mirror/iOS/, Mirror/Android/ or their workflows
mirror-ios	Changes to Mirror/iOS/, Mirror/Shared/ or its workflow, and each published release. Unit tests only
mirror-android	Changes to Mirror/Android/ or its workflow, and each published release

The UI suite takes minutes on a hosted runner (each test launches the app; the runner is 3–5x slower than a Mac mini), so it runs locally with make test-mirror-ios.

Mac app job:

• The Xcode compilation cache (CAS) step cached ~/Library/Developer/Xcode/DerivedData/CompilationCache.noindex. Because the build uses -derivedDataPath, Xcode keeps the CAS in build/ci-derived-data, so that path never existed and the step never restored or saved anything. Each miss of the strict App incremental cache (changed Makefile, workflow, file roster, …) therefore compiled everything without a cache: about 10 minutes in the previous run of this PR. The step is removed. When the incremental cache misses, the job now restores the newest App build state for the same Xcode, keeps only its CompilationCache.noindex, and builds clean against that warm CAS. This uses existing cache entries and needs no more cache space.
• Caches are restored with actions/cache/restore and saved at the end of the job by a new save-macos-caches action. The large SwiftPM, CLI and App entries are saved only from main. Each pull request run used to save about 1.6 GB, which pushed the repository over its 10 GB quota and evicted the entries of main that all pull requests restore. It also removes about 50 s of cache upload from pull request runs. Mise and GhosttyKit entries are small and slow to rebuild, so pull requests still save them.
• make lint and make check-localization run in the parallel step beside the build (new checks task with the script tests).
• make test-app exports each pass's test action log; the CI summary shows when the test runner launch and the first test suite start. (The test-progress heartbeat looked like a 2-minute wait before the first suite; the action log shows it is xcodebuild output buffering.)
• docs-ai/016-dev-build-and-ci-workflow/008-cas-fallback-and-cache-scope.md records the cache findings and measurements, and corrects the CAS note in 016.007.

Measured build job on this pull request (both runs missed the strict App incremental key, because this pull request changes the Makefile and the workflows):

Run	Compile	"Build app and run tests" step	build job
Before (no CAS)	about 10 min	782 s	16 min 28 s
After, run 1 (CAS from main, 1.3 GB)	about 2 min 20 s	322 s	8 min 46 s
After, run 2 (516 Swift cache hits, 30 misses)	96 s	267 s	7 min 33 s

Verification

• Unit tests (iOS mirror): new MirrorDocumentTests row test, MirrorReadingAnchorTests (3), MirrorInitialLayoutTests (2); MirrorSessionTests checks that a connection change clears the anchor.
• iOS 27 iPad: pane switch restores the same marker at 0 pt difference at 4 scroll depths (before: −17 / −31 pt); History → Live Output 0 pt; a temporary 600-line plain-text fixture 0 pt.
• The six former iPhone UI tests pass on iPad with iOS 26.5 and iOS 27.
• Full local suite (make test-mirror-ios, iPad, iOS 27): 86 passed, 0 failed (unit and UI). make test-mirror-ios-unit: 67 passed in 28 s.
• Mac with the CI toolchain (Xcode 26.6 / iOS 26.5): the iPad suite takes 238 s on this branch and 233 s on main, with the same per-test durations.
• make check, make test-scripts, actionlint on the changed workflows.
• CI: build, mirror-ios and mirror-android pass.