Summary

Closes #216. The `Freq:` field in the `view` TUI's GPU List row was flickering on near-idle Apple Silicon Macs, vanishing for a refresh cycle and shifting every field after it (`Pwr:` etc.) left. Two compounding bugs were at play; this PR fixes both.

Bug A — data layer (Apple Silicon)

`src/device/macos_native/ioreport.rs`. `calc_gpu_freq_with_table` and `calc_freq_from_residencies` were returning `(0, 0.0)` whenever the GPU/cluster was present (`total_residency > 0`) but spent the entire 100ms sampling window in `IDLE`/`OFF`/`DOWN` states — the normal condition for an idle Apple GPU. That 0 propagated all the way to `GpuInfo.frequency`. An idle Apple GPU is not running at 0 Hz; it is parked at its lowest P-state (~350 MHz on M1).

Fix: when active residency is zero but the cluster is present, fall back to the parked clock instead of 0.

• `calc_gpu_freq_with_table` uses `freq_table[0]` — the lowest entry of the IOKit pmgr table, which is sorted ascending.
• `calc_freq_from_residencies` tracks the minimum parseable active-state frequency during its scan and uses it as the idle fallback. This covers both the GPU fallback path (no pmgr table available) and CPU clusters via `process_cpu_channel`.
• When neither signal is available (`freq_table` empty AND no parseable active states), `(0, 0.0)` is still returned and the renderer substitutes N/A.

OpenMetrics side effect: The `all_smi_gpu_frequency_mhz` gauge for idle Apple Silicon will now export the parked P-state clock (e.g. ~350 MHz on M1) instead of 0, eliminating misleading zero-gaps in Prometheus/Grafana dashboards during idle periods.

Bug B — render layer (cross-platform)

`src/ui/renderers/gpu_renderer.rs`. The entire `Freq:` label+value block was gated on `info.frequency > 0`. Any refresh where Bug A produced 0 made the field — and therefore every field after it on the row — vanish.

Fix: mirror the existing `Util:` / `VRAM:` / `Temp:` patterns in the same function. Always render the `Freq:` label, substitute `N/A` (right-aligned at width 7, matching `Temp:`'s N/A width) when the value is 0.

This change is intentionally cross-platform. Readers that statically report `frequency: 0` (Rebellions, Intel Gaudi, AMD via WMI) will now show ` Freq: N/A` instead of hiding the field, making row layout consistent across all device types. NVIDIA / Jetson rarely hit 0 (NVML reports a stable idle graphics clock; sysfs `cur_freq` is always readable when accessible), so they are unaffected in practice — no spurious N/A.

Out of scope

Per the issue's "What NOT to change" section, this PR deliberately does NOT change `GpuInfo.frequency` from `u32` to `Option`. After Fix A, the only way `frequency == 0` reaches the renderer is genuine "no data", which the renderer now handles via N/A. Changing the type would have touched every reader, the JSON API, and the OpenMetrics exporter — out of scope here.

Test plan

• `cargo fmt --check` — clean
• `cargo clippy --all-targets -- -D warnings` — no warnings
• `cargo test` — all existing tests pass; existing `test_calc_freq_from_residencies` and `test_calc_gpu_freq_with_table` (active path) still pass unchanged
• `cargo build --release` — succeeds on Linux (CI host); the Apple-specific code path remains gated behind `#[cfg(target_os = "macos")]`
• Four new unit tests cover the new idle-state behaviour:
  • `test_calc_freq_from_residencies_all_idle` — all-idle, no parseable states → `(0, 0.0)`
  • `test_calc_freq_from_residencies_idle_with_known_states` — idle plus known states at zero residency → returns the min parsed frequency at 0% active residency
  • `test_calc_gpu_freq_with_table_idle` — fully-idle GPU with non-empty freq table → returns `freq_table[0]`
  • `test_calc_gpu_freq_with_table_empty_table_idle` — fully-idle GPU with empty freq table → `(0, 0.0)`
• Runtime check on an idle Apple Silicon Mac: `sudo ./target/release/all-smi view`, watch the GPU row over 20-30s. Expected: `Freq:` stays visible at a stable idle clock (e.g. 350-400 MHz on M1) across refresh cycles; no field-shift on idle. Under load, the field still reports the weighted active frequency.
• Runtime check on NVIDIA / Jetson: no spurious N/A in the `Freq:` column.

Files touched

• `src/ui/renderers/gpu_renderer.rs` (renderer: always render Freq, show N/A on 0)
• `src/device/macos_native/ioreport.rs` (data layer: idle Apple GPU/cluster falls back to parked clock; four new unit tests)