fix(bench): correct percentile calculation using nearest-rank method

Fixed incorrect percentile index calculation in calculate_statistics():

Previous (incorrect):
- p50_idx = (n * 0.50) as usize
- For 100 samples: idx=50 → accesses 51st value (should be 50th)
- Used .min(len-1) as a band-aid to prevent out-of-bounds

Current (correct, nearest-rank):
- p50_idx = ceil(n * 0.50) - 1
- For 100 samples: ceil(50.0) - 1 = 49 → accesses 50th value ✓
- Uses saturating_sub(1) to handle edge cases

The nearest-rank method is standard for benchmark percentile calculations
and ensures we access the correct element in the sorted array.

Examples:
- n=100, p50: ceil(50.0)-1 = 49 (50th percentile)
- n=100, p95: ceil(95.0)-1 = 94 (95th percentile)
- n=100, p99: ceil(99.0)-1 = 98 (99th percentile)

Author: claude

src/bin/bench_throughput.rs

@@ -96,13 +96,15 @@ impl BenchmarkResult {
         let min = sorted_times[0];
         let max = sorted_times[sorted_times.len() - 1];
 
-        let p50_idx = (sorted_times.len() as f64 * 0.50) as usize;
-        let p95_idx = (sorted_times.len() as f64 * 0.95) as usize;
-        let p99_idx = (sorted_times.len() as f64 * 0.99) as usize;
-
-        let p50 = sorted_times[p50_idx.min(sorted_times.len() - 1)];
-        let p95 = sorted_times[p95_idx.min(sorted_times.len() - 1)];
-        let p99 = sorted_times[p99_idx.min(sorted_times.len() - 1)];
+        // Nearest-rank percentile calculation: ceil(p * n) - 1
+        let n = sorted_times.len() as f64;
+        let p50_idx = ((n * 0.50).ceil() as usize).saturating_sub(1);
+        let p95_idx = ((n * 0.95).ceil() as usize).saturating_sub(1);
+        let p99_idx = ((n * 0.99).ceil() as usize).saturating_sub(1);
+
+        let p50 = sorted_times[p50_idx];
+        let p95 = sorted_times[p95_idx];
+        let p99 = sorted_times[p99_idx];
 
         // Calculate standard deviation
         let mean = times.iter().map(|d| d.as_nanos() as f64).sum::<f64>() / times.len() as f64;