router/semantic: tests for eval scoring and shadow harness (slice 12)

internal/router/semantic/eval_test.go

@@ -0,0 +1,96 @@
+package semantic
+
+import (
+	"context"
+	"testing"
+
+	"github.com/kami/maven/internal/router"
+)
+
+// SemanticRouterFunc adapts a bare function to SemanticRouter.
+type SemanticRouterFunc func(ctx context.Context, text string) (SemanticRouteDecision, error)
+
+func (f SemanticRouterFunc) Route(ctx context.Context, text string) (SemanticRouteDecision, error) {
+	return f(ctx, text)
+}
+
+func TestEvalScoring(t *testing.T) {
+	model := SemanticRouterFunc(func(ctx context.Context, text string) (SemanticRouteDecision, error) {
+		return SemanticRouteDecision{Route: RouteKnowledge, Confidence: 0.9}, nil
+	})
+
+	evalSet := []EvalCase{
+		{ID: "a1", Text: "привет", ExpectedRoute: RouteConversation},
+		{ID: "a2", Text: "сколько воды", ExpectedRoute: RouteKnowledge},
+		{ID: "a3", Text: "выключи свет", ExpectedRoute: RouteAction},
+		{ID: "a4", Text: "запиши заметку", ExpectedRoute: RouteMemoryWrite},
+	}
+
+	rep := ScoreEval(model, evalSet)
+	if rep.Total != 4 {
+		t.Errorf("Total = %d, want 4", rep.Total)
+	}
+	if rep.Passed != 1 {
+		t.Errorf("Passed = %d, want 1 (only knowledge)", rep.Passed)
+	}
+	if rep.FalseAction != 0 {
+		t.Errorf("FalseAction = %d, want 0", rep.FalseAction)
+	}
+	km := rep.ByRoute[RouteKnowledge]
+	if km.Precision != 0.25 {
+		t.Errorf("knowledge precision = %.3f, want 0.250", km.Precision)
+	}
+	if km.Recall != 1.0 {
+		t.Errorf("knowledge recall = %.3f, want 1.000", km.Recall)
+	}
+	t.Logf("eval report:\n%s", rep.String())
+}
+
+func TestShadowHarness(t *testing.T) {
+	model := SemanticRouterFunc(func(ctx context.Context, text string) (SemanticRouteDecision, error) {
+		if text == "выключи свет" {
+			return SemanticRouteDecision{Route: RouteAction, Confidence: 0.9}, nil
+		}
+		return SemanticRouteDecision{Route: RouteConversation, Confidence: 0.7}, nil
+	})
+
+	h := NewShadowHarness(model)
+
+	h.Observe(context.Background(), "выключи свет",
+		router.Decision{Intent: router.IntentAct}, false)
+	h.Observe(context.Background(), "привет",
+		router.Decision{Intent: router.IntentChat}, false)
+	h.Observe(context.Background(), "перезапусти докер",
+		router.Decision{Intent: router.IntentAct}, true)
+
+	rep := h.Summarize()
+	if rep.Total != 3 {
+		t.Errorf("Total = %d, want 3", rep.Total)
+	}
+	if rep.Agree != 2 {
+		t.Errorf("Agree = %d, want 2", rep.Agree)
+	}
+	if rep.Disagree != 1 {
+		t.Errorf("Disagree = %d, want 1", rep.Disagree)
+	}
+	if rep.FastPathTotal != 1 {
+		t.Errorf("FastPathTotal = %d, want 1", rep.FastPathTotal)
+	}
+	if rep.ResidualTotal != 2 {
+		t.Errorf("ResidualTotal = %d, want 2", rep.ResidualTotal)
+	}
+	t.Logf("shadow report:\n%s", rep.String())
+}
+
+func TestShadowHarnessNilModel(t *testing.T) {
+	h := NewShadowHarness(nil)
+	h.Observe(context.Background(), "привет",
+		router.Decision{Intent: router.IntentChat}, false)
+	rep := h.Summarize()
+	if rep.Total != 1 {
+		t.Errorf("Total = %d, want 1", rep.Total)
+	}
+	if rep.Agree != 0 {
+		t.Errorf("Agree = %d, want 0 (nil model → uncertain)", rep.Agree)
+	}
+}