* refactor: fix temporary value drop order for Rust 2024 compatibility

Bind async results to explicit let variables before match expressions
to ensure temporaries are dropped before MutexGuard and other locals
with custom Drop impls, preventing drop order changes in Rust 2024.

* refactor: rename gen variables for Rust 2024 keyword reservation

Rename local variables named 'gen' to 'pcie_gen_value', 'generation',
or 'generator' since 'gen' becomes a reserved keyword in Rust 2024.
Use r#gen for external crate field access (amdgpu link.gen).

* refactor: apply Rust 2024 edition auto-fixable changes

- Change macro expr fragment specifiers to expr_2021 for forward compat
- Convert if-let-else chains to match expressions (Rust 2024 requirement)
- Remove unnecessary ref in if-let patterns
- Add unsafe block inside unsafe fn body (Rust 2024 requirement)

* chore: remove accidentally committed tmp.pb

* refactor: switch Cargo.toml edition to 2024

Update edition from 2021 to 2024 and apply all resulting changes:
- Collapse nested if-let blocks using Rust 2024 let-chain syntax
- Reorder imports per Rust 2024 formatting rules
- Fix indentation in let-chain expressions

* refactor: fix ref mut in implicit borrowing patterns for Rust 2024

Remove explicit ref mut binding modifiers in patterns that already
implicitly borrow through &mut references (Rust 2024 requirement).
Also add tmp.pb to .gitignore.

* feat: add DGX Spark (GB10) unified memory architecture support

Detect UMA on Blackwell-class GPUs where NVML memory_info() returns
unavailable, and fall back to system memory from /proc/meminfo.
Suppress misleading PCIe metrics for UMA devices and annotate device
details with memory type and interconnect information.

* fix: correct swapped total/used memory destructuring for UMA devices

get_system_memory_for_uma() returns (total, used) but the NVML code
path destructured it as (used, total), causing memory values to be
swapped on GB10 systems.

* fix: correct swapped total/used memory for discrete GPUs in NVML path

The previous fix (af86da0) swapped the destructuring order to fix UMA
devices but inadvertently broke discrete GPUs by binding m.used to
total_memory and m.total to used_memory. Align the tuple elements with
the (total_memory, used_memory) destructuring.

* refactor: reduce redundant NVML calls for UMA detection

Call memory_info() once per device per tick instead of 3 times.
Cache the result and pass memory_total to is_uma_device_with_mem().
Remove redundant Memory Type insertion from hot path (already cached
in device detail).

* test: expand UMA coverage and extract is_uma_device_name helper

Extract a pure `is_uma_device_name` function from the duplicated
name-matching logic in `is_uma_device_with_mem` and the nvidia-smi
fallback path, eliminating duplication and making the predicate
directly testable.

Add 10 new unit tests covering:
- is_uma_device_name: gb10, dgx spark (case-insensitive), non-UMA GPUs,
  empty string
- read_meminfo_memory: successful parse from a temp file, missing file
- parse_meminfo_content: available > total saturating_sub, malformed lines

…#148)

* feat: enrich chassis info with DMI data, thermal zones, and GPU power

- Read DMI fields (product name, vendor, board, version, BIOS) from
  /sys/class/dmi/id/ on Linux (no sudo required)
- Read ACPI thermal zones for inlet/outlet board temperatures
- Wire up chassis reader in API server data collection loop
- Aggregate GPU power into chassis total_power_watts
- Add all_smi_chassis_info Prometheus metric with DMI labels
- Add inject_gpu_power helper for both TUI and API paths

Closes #147

* fix: improve Prometheus label escaping and cache DMI data

- Escape backslash and newline in Prometheus label values per spec
  (previously only double-quote was escaped)
- Cache static DMI fields at GenericChassisReader construction time
  instead of re-reading /sys/class/dmi/id/ on every collection cycle

* test: add tests for Prometheus label escaping and chassis DMI metrics

Add unit tests for MetricBuilder label escaping (backslash, double-quote,
newline) and ChassisMetricExporter covering the all_smi_chassis_info DMI
labels metric, inlet/outlet temperature, no-DMI guard, and
MetricPresenceFlags::all_present.

…s macOS warnings

DMI reading, thermal zone functions, and their tests are only
applicable on Linux. Add #[cfg(target_os = "linux")] to prevent
dead_code and unused_mut warnings on macOS and Windows builds.

Two changes to reduce terminal output bandwidth:

1. Only enable fast animation ticks (200ms) when there are actual
   scroll animations running, not unconditionally in remote mode.
   Previously !is_local_mode always activated fast ticks, causing
   200ms redraws even when no text was scrolling.

2. Increase SCROLL_UPDATE_FREQUENCY from 1 to 5, so marquee text
   updates every 500ms instead of every 100ms. This reduces the
   number of changed lines per frame by ~5x for multi-node views.

The root cause of keyboard lag over SSH: stdout.flush() is a blocking
syscall that stalls the entire async event loop when the terminal pipe
is congested. On slow SSH connections, flush can block for hundreds of
milliseconds, during which keyboard events cannot be processed.

Changes:
- DifferentialRenderer now returns a byte buffer instead of writing
  directly to stdout, separating diff computation from I/O
- Terminal write+flush is offloaded to spawn_blocking (fire-and-forget)
  so the event loop immediately returns to processing keyboard input
- If a previous flush is still in progress, the current frame is
  skipped to prevent unbounded backpressure on slow pipes
- Added flush_in_progress AtomicBool flag for lock-free coordination

Reverts the spawn_blocking fire-and-forget approach which caused frame
drops and unresponsive UI. Instead, keeps synchronous flush but adds
event draining after each render cycle.

The root cause of key buffering over SSH: flush blocks for hundreds of
ms, during which key events accumulate in the channel. Previously,
each queued event triggered its own render+flush cycle, creating a
cascade where a single keypress during a slow flush led to multiple
sequential re-renders.

Now after each render, all pending terminal events are drained from
the channel and batch-processed, followed by a single re-render.
This eliminates the per-event render cascade.

…:detect()

Root cause found: AppState::new() called RuntimeEnvironment::detect()
which runs `systemd-detect-virt` (external process fork+exec) on every
invocation. Since as_app_state() calls AppState::new() once per render
frame, this added ~1 second of blocking per frame.

- Add Default impl for RuntimeEnvironment (no detection, instant)
- Change AppState::new() to use RuntimeEnvironment::default() instead
  of RuntimeEnvironment::detect()
- RuntimeEnvironment::detect() is still called once at startup in
  main.rs and propagated via state; the per-frame as_app_state()
  copies the already-detected value from the snapshot

Also fixes divide-by-zero panic in dashboard sparkline when
available_width is 0.

SCROLL_UPDATE_FREQUENCY 5 → 2 (scroll every 200ms instead of 500ms).
The previous slowdown was a workaround for the RuntimeEnvironment::detect()
stall which is now resolved.

Audit found two additional hot-path performance issues:

1. cpu_renderer.rs: /.dockerenv existence check + up to 3 cgroup file
   reads executed on EVERY frame when per-core CPU view is enabled.
   Now cached in OnceLock (read once, reused forever).

2. ui/tabs.rs test helper: RuntimeEnvironment::detect() called in
   test fixture, adding ~1s per test run from systemd-detect-virt.
   Changed to RuntimeEnvironment::default().

Test suite speedup: lib 1.02s→0.28s, integration 1.02s→0.10s.

- Use fully qualified std::sync::OnceLock in #[cfg(target_os = "linux")]
  static to avoid unused import on macOS
- Add #[allow(unused_mut)] to dmi_detail in GenericChassisReader::new()
  since mutation is cfg-gated to Linux only