nvidia_smi.gpu_temperature could stay empty on newer NVIDIA driver XML despite valid <gpu_temp> values. Driver 580 adds extra temperature siblings and mixed N/A/whitespace values; current parsing and power-selection logic could drop otherwise valid metrics.

• Parser hardening (value normalization)

  • Normalize incoming XML text before validation/parsing (TrimSpace).
  • Make unit stripping robust to leading/trailing/multi-space formatting, not just single-space tokenization.

• Power metric fallback correctness

  • Reworked power_draw selection to skip invalid candidates (N/A, empty) and continue to instant_power_draw / average_power_draw when present.
  • Prevents losing power metrics when primary field exists but is not usable.

• Regression coverage for Blackwell/580 shape

  • Added a focused fixture for RTX PRO 6000 + driver 580 temperature/power XML layout.
  • Extended collector tests to assert temperature and fallback power extraction from this payload.

for _, value := range []*string{
    gpuPowerDraw(gpu.PowerReadings),
    gpuPowerDraw(gpu.GPUPowerReadings),
    gpuInstantPowerDraw(gpu.GPUPowerReadings),
    gpuAveragePowerDraw(gpu.GPUPowerReadings),
} {
    if value == nil || !isValidValue(*value) {
        continue
    }
    addMetric(mx, px+"power_draw", *value, 0)
    return
}