feat(power): real metered-network detection on macOS and Linux

[pmaxhogan]

What

Metered-network detection was real on Windows only (INetworkCostManager::GetCost); macOS and Linux returned a conservative unmetered stub, so skip_on_metered was inert on those platforms (issue #32). This wires both.

• Linux - reads NetworkManager's aggregate Metered property (org.freedesktop.NetworkManager) over the system bus via zbus's blocking API, dispatched through tokio::task::spawn_blocking (zbus's blocking API must not be driven from inside an async runtime) and bounded by a 5 s timeout so a wedged bus can never stall the AC/battery poll. NMMetered YES/GUESS_YES -> metered, NO/GUESS_NO -> not metered, UNKNOWN or a failed read (no NetworkManager / no system bus) -> Unknown -> not metered.
• macOS - there is no literal "metered" bit. A long-lived NWPathMonitor (Network.framework C API; its update handler is an Objective-C block via block2, delivered on a libdispatch global queue) caches nw_path_is_expensive (cellular / personal hotspot) || nw_path_is_constrained (Low Data Mode) - the documented metered proxies - into an AtomicU8 that the 30 s power poll reads cheaply.

Design

The per-OS raw read is separated from the decision logic. Three pure classify_* functions (classify_windows_cost, classify_nm_metered, classify_nw_path) plus MeteredStatus/on_metered turn each OS's raw value into the on_metered_network bool. They are #[cfg(any(test, target_os = ...))], so all three are compiled and unit-tested on every OS - the classification is CI-covered on Windows, macOS, and Linux alike; only the thin OS-call adapters (COM / D-Bus / NWPathMonitor) are compile-checked-only off their native OS.

Every read collapses ambiguity to MeteredStatus::Unknown, which maps to false (not metered) - the safe direction, since a wrong "not metered" only fails to skip a rare metered link whereas a wrong "metered" would stall ALL sync.

No stored-format or public-API changes: PowerState shape and the PowerSource trait are unchanged; this is purely additive.

Tested vs compile-only

Development was on Windows, so the non-Windows adapters are compile-verified via the 3-OS CI matrix, not locally:

• Unit-tested on all three CI OSes (run on Windows locally too): the three classify_* functions, MeteredStatus::on_metered, from_u8, reachable_hint - the full decision logic.
• Runtime-tested on the Windows CI runner (and locally): detect_metered() real COM read is total (windows_detect_metered_is_total); the existing RealPowerSource construction tests exercise the read path.
• Runtime-tested on the macOS CI runner: macos_monitor_start_and_read_is_total starts the real NWPathMonitor and reads its cache - the only runtime exercise of the Network.framework / block2 / libdispatch FFI (a wrong signature or bad link fails here).
• Compile-only (CI matrix): the macOS FFI adapter body and the Linux zbus D-Bus adapter body. There is no metered hardware in CI, and NetworkManager may be absent on the Linux runner, so the Linux read's happy path is not asserted in CI - the D-Bus/classification split keeps the classification (the part with real logic) fully covered, and the adapter is a thin, safe-defaulting wrapper.

APIs were taken from documented sources (zbus proxy/blocking docs; Apple Network.framework; objc2/block2 encoding), not guessed.

Nothing deferred

macOS isExpensive/isConstrained is the documented, feasible proxy (macOS exposes no literal metered bit), so #32 is fully addressed rather than scope-reduced.

Closes #32.

🤖 Generated with Claude Code

[github-actions]

Coverage

Area	main	this PR	delta
Rust (lib crates)	76.76%	76.74%	-0.02 (OK)
UI (vue/ts)	88.17%	88.17%	+0.00 (OK)

Gate: passed - no coverage regression (epsilon 0.1 pp).