feat(wing): backtest metric (coverage/precision at a threshold) and shared TopWing (#205 S1)

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>

ingestion/internal/wing/backtest_test.go

@@ -0,0 +1,38 @@
+package wing_test
+
+import (
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+
+	"github.com/mathiasbq/hyperguild/ingestion/internal/wing"
+)
+
+func TestEvaluate_PolicyAtAThreshold(t *testing.T) {
+	samples := []wing.Sample{
+		{Truth: "a", Suggested: "a", Confidence: 0.9, ScorerTop: "a"},  // auto, correct
+		{Truth: "a", Suggested: "b", Confidence: 0.9, ScorerTop: "b"},  // auto, WRONG (both models agree on a wrong wing)
+		{Truth: "a", Suggested: "a", Confidence: 0.5, ScorerTop: "a"},  // below threshold -> exception
+		{Truth: "a", Suggested: "a", Confidence: 0.95, ScorerTop: "b"}, // scorer disagrees -> exception
+		{Truth: "a", Suggested: "", Confidence: 0.99, ScorerTop: "a"},  // abstain -> exception
+	}
+	r := wing.Evaluate(samples, 0.8)
+	assert.Equal(t, 5, r.N)
+	assert.Equal(t, 2, r.Auto)
+	assert.Equal(t, 1, r.Correct)
+	assert.InDelta(t, 0.4, r.Coverage, 1e-9)
+	assert.InDelta(t, 0.5, r.Precision, 1e-9)
+}
+
+func TestEvaluate_NothingAutoHasZeroPrecisionNotNaN(t *testing.T) {
+	r := wing.Evaluate([]wing.Sample{{Truth: "a", Suggested: "a", Confidence: 0.1, ScorerTop: "a"}}, 0.8)
+	assert.Zero(t, r.Auto)
+	assert.Zero(t, r.Precision)
+	assert.Zero(t, r.Coverage)
+}
+
+func TestEvaluate_EmptyInput(t *testing.T) {
+	r := wing.Evaluate(nil, 0.8)
+	assert.Zero(t, r.N)
+	assert.Zero(t, r.Coverage)
+}