feat(tiny-dancer): add score() raw-forward inference + 0.1.22

score(modelPath, embedding) loads a trainRouter .safetensors and runs the
FastGRNN directly on the raw embedding (the inference path that matches how
trainRouter trains — Router.route's 5-feature engineering is a different model
contract). Validated easy=1.0/hard=0.0 on a freshly built binary.

Co-Authored-By: claude-flow <ruv@ruv.net>

Author: ruvnet

crates/ruvector-tiny-dancer-node/src/lib.rs

@@ -398,3 +398,29 @@ pub async fn train_router(
     .map_err(|e| Error::from_reason(format!("Task failed: {}", e)))?
     .map_err(Error::from_reason)
 }
+
+/// Score a query embedding with a trained FastGRNN model (raw forward pass).
+///
+/// Loads the `.safetensors` produced by {@link train_router} and runs the model
+/// directly on `embedding` (which must match the model's `input_dim`). Returns
+/// the sigmoid output in 0..1 — high means "the cheap model is good enough"
+/// (route to the cheaper model); low means route to a stronger model.
+///
+/// This is the inference path that matches `trainRouter` (trained on raw
+/// embeddings); it does not run `Router`'s feature engineering.
+#[napi]
+pub async fn score(model_path: String, embedding: Vec<f64>) -> Result<f64> {
+    use ruvector_tiny_dancer_core::model::FastGRNN;
+
+    tokio::task::spawn_blocking(move || -> std::result::Result<f64, String> {
+        let model = FastGRNN::load(&model_path).map_err(|e| format!("load: {e}"))?;
+        let feats: Vec<f32> = embedding.into_iter().map(|v| v as f32).collect();
+        let s = model
+            .forward(&feats, None)
+            .map_err(|e| format!("forward: {e}"))?;
+        Ok(s as f64)
+    })
+    .await
+    .map_err(|e| Error::from_reason(format!("Task failed: {}", e)))?
+    .map_err(Error::from_reason)
+}

npm/packages/tiny-dancer/index.d.ts

@@ -191,3 +191,20 @@ export function trainRouter(
   prices: Record<string, number>,
   options: TrainRouterOptions
 ): Promise<TrainRouterResult>;
+
+/**
+ * Score a query embedding with a trained FastGRNN model (raw forward pass).
+ *
+ * Loads the `.safetensors` produced by {@link trainRouter} and runs the model
+ * directly on `embedding` (which must match the model's `inputDim`). Returns the
+ * sigmoid output in 0..1 — high means "the cheap model is good enough" (route to
+ * the cheaper model); low means route to a stronger model. This is the inference
+ * path that matches `trainRouter`; it does not run `Router`'s feature engineering.
+ *
+ * @example
+ * ```javascript
+ * const s = await score('./router.safetensors', queryEmbedding);
+ * const useCheap = s >= 0.5;
+ * ```
+ */
+export function score(modelPath: string, embedding: number[]): Promise<number>;

npm/packages/tiny-dancer/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@ruvector/tiny-dancer",
-  "version": "0.1.21",
+  "version": "0.1.22",
   "description": "Neural router for AI agent orchestration - FastGRNN-based intelligent routing with circuit breaker, uncertainty estimation, and hot-reload",
   "main": "index.js",
   "types": "index.d.ts",
@@ -31,14 +31,14 @@
     "@napi-rs/cli": "^2.18.0"
   },
   "optionalDependencies": {
-    "@ruvector/tiny-dancer-linux-x64-gnu": "0.1.21",
-    "@ruvector/tiny-dancer-linux-x64-musl": "0.1.21",
-    "@ruvector/tiny-dancer-linux-arm64-gnu": "0.1.21",
-    "@ruvector/tiny-dancer-linux-arm64-musl": "0.1.21",
-    "@ruvector/tiny-dancer-darwin-x64": "0.1.21",
-    "@ruvector/tiny-dancer-darwin-arm64": "0.1.21",
-    "@ruvector/tiny-dancer-win32-x64-msvc": "0.1.21",
-    "@ruvector/tiny-dancer-win32-arm64-msvc": "0.1.21"
+    "@ruvector/tiny-dancer-linux-x64-gnu": "0.1.22",
+    "@ruvector/tiny-dancer-linux-x64-musl": "0.1.22",
+    "@ruvector/tiny-dancer-linux-arm64-gnu": "0.1.22",
+    "@ruvector/tiny-dancer-linux-arm64-musl": "0.1.22",
+    "@ruvector/tiny-dancer-darwin-x64": "0.1.22",
+    "@ruvector/tiny-dancer-darwin-arm64": "0.1.22",
+    "@ruvector/tiny-dancer-win32-x64-msvc": "0.1.22",
+    "@ruvector/tiny-dancer-win32-arm64-msvc": "0.1.22"
   },
   "publishConfig": {
     "access": "public"