test(ios): runner coverage gaps from the A7 pass — TvRemote has zero tests, ScreenRecorder.append monotonicity untested #1897

Two coverage gaps surfaced by the #1781 A7 confirmation pass (the independent re-derivation of #1861's classification — full table in #1781 (comment)). Both are gaps in what exists, not misclassifications; neither blocks anything.

1. RunnerTests+TvRemote.swift has zero test methods

The file exists in apple/runner/AgentDeviceRunner/AgentDeviceRunnerUITests/ but contains no func test… — the tv-remote command path has no runner-side test at all, in any lane. The only tvOS evidence anywhere is the transcript-fake test/integration/provider-scenarios/tvos-remote.test.ts, which exercises the daemon side, not the Swift dispatch. A decision-level test (key-name → event mapping, invalid-key rejection) would be host-runnable under the #1861 classification — no simulator needed.

2. ScreenRecorder.append(image:) frame-timestamp monotonicity is untested

The sole recording test in the suite (testRecordStopIsIdempotentAfterNativeRecorderAlreadyStopped, RunnerTests+RecordingTests.swift) keeps activeRecording nil throughout, so the frame-append path — including the monotonicity guard on frame timestamps — never executes under test. A regression here produces unplayable/glitched MP4s, the same failure family as the Android moov bug (duration-0 validator, patched in place). A host-runnable test feeding fabricated timestamps through append covers it.

Lane accounting (per the #1781 rule)

• Catches: silent tv-remote dispatch regressions; non-monotonic frame timestamps producing corrupt recordings.
• Evidence: gap found by the A7 classification pass; the recording family has a real prior on Android.
• Cost: two host-runnable tests in the existing Swift Runner Host XCTests lane (~0 added CI time).
• Kill criterion: if tv-remote is dropped as a command, delete with it; the recorder test lives as long as ScreenRecorder does.