feat(topology): LLM-judge scorer + async Scorer trait

Make Scorer async and add JudgeScorer (behind the `provider` feature): asks a
cm-llm model to rate a run's output 0-100 vs the task and normalizes to [0,1],
giving the comparison harness real quality numbers. Robust integer parsing
(handles "Score: 92/100", clamps >100); provider errors score 0.0.

11 tests with --features provider (judge incl. parse + scripted-provider score);
core stays 7. Clippy clean.

Co-Authored-By: Claude Opus 4.8 <[email protected]>

crates/cm-orchestrator/src/harness.rs

@@ -13,10 +13,12 @@ use cm_topology::{TopologyGraph, TopologyKind};
 use crate::{execute, OrchestratorError, RunRecord, TurnExecutor};
 
 /// Scores the quality of a run's final output in `[0,1]`. Real deployments use
-/// an LLM judge; tests/benchmarks can use a deterministic scorer.
+/// an LLM judge ([`crate::JudgeScorer`]); tests/benchmarks can use a
+/// deterministic scorer.
+#[allow(async_fn_in_trait)]
 pub trait Scorer {
     /// Quality of `record` for `task`, in `[0,1]` (higher is better).
-    fn score(&self, task: &str, record: &RunRecord) -> f64;
+    async fn score(&self, task: &str, record: &RunRecord) -> f64;
 }
 
 /// One topology's result within a comparison.
@@ -72,7 +74,7 @@ pub async fn compare<E: TurnExecutor, S: Scorer>(
     let mut results: Vec<TopologyResult> = Vec::with_capacity(graphs.len());
     for g in graphs {
         let rec = execute(g, task, executor).await?;
-        let quality = scorer.score(task, &rec).clamp(0.0, 1.0);
+        let quality = scorer.score(task, &rec).await.clamp(0.0, 1.0);
         results.push(TopologyResult {
             kind: rec.kind,
             quality,
@@ -152,7 +154,7 @@ mod tests {
     /// Deterministic quality proxy: longer (richer) output scores higher.
     struct LengthScorer;
     impl Scorer for LengthScorer {
-        fn score(&self, _task: &str, record: &RunRecord) -> f64 {
+        async fn score(&self, _task: &str, record: &RunRecord) -> f64 {
             (record.final_output.len() as f64 / 200.0).min(1.0)
         }
     }