Enhance benchmark: server power via IPMI, efficiency metrics, FP64, power limit check

- Sample server power (IPMI dcmi) during baseline+steady phases in parallel;
  compute delta vs GPU-reported sum; flag ratio < 0.75 as unreliable reporting
- Collect base_graphics_clock_mhz, multiprocessor_count, default_power_limit_w
  from nvidia-smi alongside existing GPU info
- Add tops_per_sm_per_ghz efficiency metric (model-agnostic silicon quality signal)
- Flag when enforced power limit is below default TDP by >5%
- Add fp64 profile to bee-gpu-burn worker (CUDA_R_64F, CUBLAS_COMPUTE_64F, min cc 8.0)
- Improve Executive Summary: overall pass count, FAILED GPU finding
- Throttle counters now shown as % of steady window instead of raw microseconds
- bible-local: clock calibration research, H100/H200 spec, real-world GEMM baselines

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

audit/internal/platform/benchmark_report.go

@@ -56,6 +56,9 @@ func renderBenchmarkReportWithCharts(result NvidiaBenchmarkResult, charts []benc
 		fmt.Fprintf(&b, "  Status: %s\n", gpu.Status)
 		fmt.Fprintf(&b, "  Composite score: %.2f\n", gpu.Scores.CompositeScore)
 		fmt.Fprintf(&b, "  Compute score: %.2f\n", gpu.Scores.ComputeScore)
+		if gpu.Scores.TOPSPerSMPerGHz > 0 {
+			fmt.Fprintf(&b, "  Compute efficiency: %.3f TOPS/SM/GHz\n", gpu.Scores.TOPSPerSMPerGHz)
+		}
 		fmt.Fprintf(&b, "  Power sustain: %.1f\n", gpu.Scores.PowerSustainScore)
 		fmt.Fprintf(&b, "  Thermal sustain: %.1f\n", gpu.Scores.ThermalSustainScore)
 		fmt.Fprintf(&b, "  Stability: %.1f\n", gpu.Scores.StabilityScore)
@@ -77,13 +80,7 @@ func renderBenchmarkReportWithCharts(result NvidiaBenchmarkResult, charts []benc
 				}
 			}
 		}
-		fmt.Fprintf(&b, "  Throttle counters (us): sw_power=%d sw_thermal=%d sync_boost=%d hw_thermal=%d hw_power_brake=%d\n",
-			gpu.Throttle.SWPowerCapUS,
-			gpu.Throttle.SWThermalSlowdownUS,
-			gpu.Throttle.SyncBoostUS,
-			gpu.Throttle.HWThermalSlowdownUS,
-			gpu.Throttle.HWPowerBrakeSlowdownUS,
-		)
+		fmt.Fprintf(&b, "  Throttle: %s\n", formatThrottleLine(gpu.Throttle, gpu.Steady.DurationSec))
 		if len(gpu.Notes) > 0 {
 			fmt.Fprintf(&b, "  Notes:\n")
 			for _, note := range gpu.Notes {
@@ -121,6 +118,26 @@ func renderBenchmarkReportWithCharts(result NvidiaBenchmarkResult, charts []benc
 		}
 	}
 
+	if sp := result.ServerPower; sp != nil {
+		fmt.Fprintf(&b, "Server Power (IPMI)\n")
+		fmt.Fprintf(&b, "-------------------\n")
+		if !sp.Available {
+			fmt.Fprintf(&b, "Unavailable\n")
+		} else {
+			fmt.Fprintf(&b, "  Server idle:         %.0f W\n", sp.IdleW)
+			fmt.Fprintf(&b, "  Server under load:   %.0f W\n", sp.LoadedW)
+			fmt.Fprintf(&b, "  Server delta:        %.0f W\n", sp.DeltaW)
+			fmt.Fprintf(&b, "  GPU reported (sum):  %.0f W\n", sp.GPUReportedSumW)
+			if sp.ReportingRatio > 0 {
+				fmt.Fprintf(&b, "  Reporting ratio:     %.2f  (1.0 = accurate, <0.75 = GPU over-reports)\n", sp.ReportingRatio)
+			}
+		}
+		for _, note := range sp.Notes {
+			fmt.Fprintf(&b, "  Note: %s\n", note)
+		}
+		b.WriteString("\n")
+	}
+
 	fmt.Fprintf(&b, "Methodology\n")
 	fmt.Fprintf(&b, "-----------\n")
 	fmt.Fprintf(&b, "- Profile %s uses standardized baseline, warmup, steady-state, interconnect, and cooldown phases.\n", result.BenchmarkProfile)
@@ -175,6 +192,42 @@ func stripANSIEscapeSequences(raw string) string {
 	return ansiEscapePattern.ReplaceAllString(raw, "")
 }
 
+// formatThrottleLine renders throttle counters as human-readable percentages of
+// the steady-state window.  Only non-zero counters are shown.  When the steady
+// duration is unknown (0), raw seconds are shown instead.
+func formatThrottleLine(t BenchmarkThrottleCounters, steadyDurationSec float64) string {
+	type counter struct {
+		label string
+		us    uint64
+	}
+	counters := []counter{
+		{"sw_power", t.SWPowerCapUS},
+		{"sw_thermal", t.SWThermalSlowdownUS},
+		{"sync_boost", t.SyncBoostUS},
+		{"hw_thermal", t.HWThermalSlowdownUS},
+		{"hw_power_brake", t.HWPowerBrakeSlowdownUS},
+	}
+	var parts []string
+	for _, c := range counters {
+		if c.us == 0 {
+			continue
+		}
+		sec := float64(c.us) / 1e6
+		if steadyDurationSec > 0 {
+			pct := sec / steadyDurationSec * 100
+			parts = append(parts, fmt.Sprintf("%s=%.1f%% (%.0fs)", c.label, pct, sec))
+		} else if sec < 1 {
+			parts = append(parts, fmt.Sprintf("%s=%.0fms", c.label, sec*1000))
+		} else {
+			parts = append(parts, fmt.Sprintf("%s=%.1fs", c.label, sec))
+		}
+	}
+	if len(parts) == 0 {
+		return "none"
+	}
+	return strings.Join(parts, "  ")
+}
+
 func renderBenchmarkSummary(result NvidiaBenchmarkResult) string {
 	var b strings.Builder
 	fmt.Fprintf(&b, "run_at_utc=%s\n", result.GeneratedAt.Format(time.RFC3339))