router/semantic: legacy baseline eval test (slice 13)

TestLegacyBaseline builds a minimal but complete router (stage-0
grammars, hash-embedder classifier seeded from models/seeds, deployed
confidence threshold) and runs it against the full 136-example corpus.

Reports macro F1, false-action rate, fast-path/residual/router-residual
accuracy, per-route P/R/F1, confusion matrix, and contrast family
breakdown (negation, question, reported_speech, quotation, hypothetical,
capability_question). Pre-route consumed cases are tracked separately
from the fast-path bucket.

The hash embedder is deterministic, so this baseline is reproducible.
The ONNX embedder would score higher; measure both before drawing
conclusions.

internal/router/semantic/eval_test.go

@@ -3,6 +3,7 @@ package semantic
 import (
 	"context"
 	"testing"
+	"time"
 
 	"github.com/kami/maven/internal/router"
 )
@@ -94,3 +95,71 @@ func TestShadowHarnessNilModel(t *testing.T) {
 		t.Errorf("Agree = %d, want 0 (nil model → uncertain)", rep.Agree)
 	}
 }
+
+// legacyRouterAdapter wraps a *router.Router to satisfy LegacyRouter.
+type legacyRouterAdapter struct {
+	router *router.Router
+}
+
+func (a *legacyRouterAdapter) Route(ctx context.Context, input router.NormalizedInput, now time.Time) (router.Decision, error) {
+	return a.router.Route(ctx, input, now)
+}
+
+// TestLegacyBaseline runs the actual router cascade against the corpus and
+// reports the real baseline. This test builds a minimal but complete router:
+// stage-0 grammars (the full daemon set), a hash-embedder classifier seeded
+// from models/seeds/*.txt, and the deployed confidence threshold.
+//
+// The hash embedder is deterministic, so this baseline is reproducible. The
+// ONNX embedder would score higher; measure both before drawing conclusions.
+func TestLegacyBaseline(t *testing.T) {
+	exs, err := LoadCorpus()
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	r := buildMinimalRouter(t)
+	now := time.Now()
+
+	rep := ScoreLegacy(context.Background(), r, exs, now)
+	t.Log("\n" + rep.String())
+
+	// Print contrast family breakdown.
+	families := ContrastFamilies(rep)
+	if len(families) > 0 {
+		t.Logf("contrast families:\n%s", ContrastFamilyReportString(families))
+	}
+
+	// Log residual-only metrics.
+	t.Logf("residual: %d/%d (%.1f%%)",
+		rep.ResidualPassed, rep.ResidualTotal,
+		100*float64(rep.ResidualPassed)/maxf(float64(rep.ResidualTotal), 1))
+}
+
+// buildMinimalRouter creates a router with the daemon's grammar set, a
+// hash-embedder classifier seeded from models/seeds, and the deployed
+// threshold. This is the minimal real router that can score the corpus.
+func buildMinimalRouter(t *testing.T) LegacyRouter {
+	t.Helper()
+	// Use the same act matcher and seed loading as the eval package.
+	acts := router.DefaultActMatcher{Fns: actVerbList()}
+	cls := buildSeededClassifier(t, router.NewHashEmbedder(1024))
+	r := router.New(router.Config{
+		Grammars:   router.StageZeroGrammars(acts),
+		Classifier: cls,
+		Extractor: router.Extractor{
+			Time:  router.StubDateTimeParser{},
+			Acts:  acts,
+			Facts: router.DefaultFactParser{},
+		},
+		Threshold: 0.55,
+	})
+	return &legacyRouterAdapter{router: r}
+}
+
+func maxf(a, b float64) float64 {
+	if a > b {
+		return a
+	}
+	return b
+}