Add NEUI_API_IOS, an iOS-only host extension

[defiantnerd]

UIKit exposes a lot that has no portable equivalent, and a client had no way to reach any of it: keep the screen awake through a set, stop a thumb near the bezel from swiping the app away mid-song, ask whether the device dropped into Low Power Mode, fire a haptic tick.

What this adds

NEUI_API_IOS (include/neui/d/ios.h) — a 22-method vtable returned only by neui.host.ios and the xpl host on platform_ios.mm. Every other host returns NULL from get_interface, the same feature-detect contract NEUI_API_EMBED and NEUI_API_METRICS already use. Four groups: idle timer / screen-edge gesture deferral / home indicator / brightness; status bar / orientation / forced appearance; Dynamic Type category and the accessibility switches; device / battery / thermal / Low Power / haptics / keyboard inset.

Unlike the xpl host's inert NEUI_API_EMBED on iOS, it is never handed out empty: a host that returns it implements every method, and a call it cannot answer says so in its return value. Changes arrive as one NEUI_EVENT_IOS_ENVIRONMENT_CHANGED carrying a bitmask of what moved, in its own event category — Dynamic Type stays on METRICS_CHANGED, where clients already handle it.

Safe-area insets are deliberately not here. They are portable, they already exist in NEUI_API_METRICS, and an Android host can implement the same seam from WindowInsets. Entries that do have an Android counterpart are marked inline for whoever writes that host; three are marked Android: none.

Seams are registries now, including the metrics ones

neui_init() registers the native iOS host and then xpl, whose platform_init() runs last, so an assigned single-slot seam always ended up holding xpl's — and for a native-host client it failed silently.

This was already true of NEUI_API_METRICS. safe_area_insets measured zeros on every edge in a binary linking both hosts, masked because get_client_rect computes its top inset directly rather than through the seam. Confirmed by temporarily restoring last-wins and re-running examples/ios:

old (last-wins):  safe_area_insets: l=0 t=0   r=0 b=0
fixed:            safe_area_insets: l=0 t=106 r=0 b=34

Both seam families now add rather than assign; the frame lookups try each until one claims the frame. The safe-area seam signature became bool so "not my frame" is distinguishable from "mine, and the insets are zero". The example prints its insets permanently, so a regression shows up on the next run.

Verification

	result
iOS 26.5 simulator	351 cases / 2107 checks, 0 failed
macOS	343 / 2046, 0 failed
Both, -Werror	clean

tests/test_metrics.cpp is new and covers the seam rule with two fake hosts — metrics.h had no test at all before. tests/test_ios_api.cpp pins the enum and bit-flag ABI on all four CI jobs; tests/test_ios_api_device.mm is iOS-only and runs on the simulator in CI.

Not verified: battery, thermal, Low Power, brightness and whether haptics actually fire need hardware — the simulator can only confirm they answer in range.

Docs

docs/host-ios.md is new and entered through the CLAUDE.md subsystem index, which is the path an agent actually follows. Also corrected what would otherwise mislead: CLAUDE.md's platform list and host table, TODO.md listing iOS as unported, the neui.ios.* namespace missing from attrs.h, and the "phase-2 stubs" claims in hosts/ios/host.h and platform_ios.mm for widgets that are long since implemented.