perf(decode): cache struct schemas per type

Assisted-by: GLM 5.3

decode.go

@@ -8,6 +8,7 @@ import (
 	"math"
 	"reflect"
 	"strings"
+	"sync"
 	"time"
 )
 
@@ -105,7 +106,7 @@ func (d *decoder) assignTable(tbl map[string]any, dst reflect.Value) error {
 }
 
 func (d *decoder) assignStruct(tbl map[string]any, dst reflect.Value) error {
-	schema := newStructSchema(dst.Type())
+	schema := cachedStructSchema(dst.Type())
 	for key, val := range tbl {
 		field, ok := schema.byName[strings.ToLower(key)]
 		if !ok {
@@ -253,6 +254,21 @@ type structSchema struct {
 	embedMaps [][]int
 }
 
+// structSchemaCache holds one schema per struct type. A schema is immutable
+// once published, so concurrent callers only race to build an identical value,
+// the same trade-off encoding/json's field cache makes. The cache grows with
+// the number of distinct types decoded or encoded, never per document.
+var structSchemaCache sync.Map // reflect.Type -> structSchema
+
+func cachedStructSchema(t reflect.Type) structSchema {
+	if s, ok := structSchemaCache.Load(t); ok {
+		return s.(structSchema)
+	}
+	s := newStructSchema(t)
+	actual, _ := structSchemaCache.LoadOrStore(t, s)
+	return actual.(structSchema)
+}
+
 func newStructSchema(t reflect.Type) structSchema {
 	s := structSchema{byName: make(map[string]structFieldLoc, t.NumField())}
 	// A struct may embed a pointer to itself, which is legal Go, so the walk