evaluate: use rayon par_sort_by for parallel LLM comparisons

Merge sort parallelizes naturally — multiple LLM comparison calls
happen concurrently. Safe because merge sort terminates correctly
even with non-deterministic comparators (unlike quicksort).

Co-Authored-By: Kent Overstreet <kent.overstreet@linux.dev>

poc-memory/src/cli/agent.rs

@@ -3,6 +3,7 @@
 use crate::store;
 use crate::store::StoreView;
 use crate::agents::llm;
+use std::sync::atomic::{AtomicUsize, Ordering};
 
 pub fn cmd_consolidate_batch(count: usize, auto: bool, agent: Option<String>) -> Result<(), String> {
     let store = store::Store::load()?;
@@ -178,17 +179,18 @@ pub fn cmd_evaluate_agents(samples_per_type: usize, model: &str) -> Result<(), S
         all_samples.len() * (all_samples.len() as f64).log2() as usize,
         model);
 
-    // Sort with LLM comparator — yes, really. Rust's sort_by with an
-    // LLM as the comparison function. Each comparison is an API call.
-    let mut comparisons = 0usize;
-    all_samples.sort_by(|a, b| {
-        comparisons += 1;
-        if comparisons % 10 == 0 {
-            eprint!("  {} comparisons...\r", comparisons);
+    // Sort with LLM comparator — yes, really. Rayon's parallel merge sort
+    // with an LLM as the comparison function. Multiple API calls in parallel.
+    let comparisons = AtomicUsize::new(0);
+    use rayon::slice::ParallelSliceMut;
+    all_samples.par_sort_by(|a, b| {
+        let n = comparisons.fetch_add(1, Ordering::Relaxed);
+        if n % 10 == 0 {
+            eprint!("  {} comparisons...\r", n);
         }
         llm_compare(a, b, model).unwrap_or(std::cmp::Ordering::Equal)
     });
-    eprintln!("  {} total comparisons", comparisons);
+    eprintln!("  {} total comparisons", comparisons.load(Ordering::Relaxed));
 
     let sorted = all_samples;