feat(model,capacity,vision,io): per-persona 路由 + 本地三级(默认关) + L3 分类 + OCR 独立模型 + image-crop 工具（裁决 03.4.1/03.4.3/14.4.1/14.4.4/14.4.5）

- 03.4.1：per-persona 路由载体选 models.json（组内新增 personaModels/personaProviders + 校验 + PersonaRoute()），
  Router/RuntimeLLM 快照绑定、loop 流式路径按该 turn 的 persona 路由；主步用路由结果、降级步钉自己的模型
- 03.4.3：本地第三级降级链（LLM_LOCAL_*，Configured() 含 Enabled ⇒ **默认关**），warn 标注来源
- 14.4.4：capacity 接入 L3（分类器接口 + TTL 缓存 + 严格 JSON 解析），CAPACITY_L3_ENABLED **默认 false**
- 14.4.1：语义修正 —— group.multimodal 才是「主模型能否收图」（env 路径不再臆断），LLM_VISION_MODEL 只配视觉工具；
  vision/OCR 可各用不同模型；runtime 不再因有图抢路由（vision.Router 是唯一分层权威）
- 14.4.5：新增 image-crop 工具（工具数 5→6，策略默认 owner-only），schema 支持按区域裁切
- 组合根：wire_runtime 读 group.Multimodal；capacity 装配拆出 wire_capacity.go（守 500 行）
- 复核：`go build`/`go vet`/各包 `go test -count=1` 全绿；**6 条变异反证**先红后还原（含作者自查出的 2 条假阳性变异）
- 未落：按 provider 的 thinking 开关（需 openai_encode.go/openai.go 文件面）、persona awakening（需 loader 文件面）

internal/runtime/capacity/capacity.go

@@ -110,6 +110,17 @@ type Config struct {
 	// Awakening carries the persona-level override (design §3). Zero = no
 	// override.
 	Awakening Awakening
+	// L3Enabled turns on the model-assisted intent layer (design/14 §2.2 L3;
+	// 裁决 14.4.4). It is off by default: the layer costs one model call per
+	// undecided turn, so it must be asked for explicitly. When false the
+	// classifier is never consulted, whatever L3Classifier holds.
+	L3Enabled bool
+	// L3Classifier is the model-assisted classifier the L3 layer consults when
+	// L3Enabled is true. Nil keeps the layer silent.
+	L3Classifier IntentClassifier
+	// L3CacheTTL is how long a verdict for identical turn text is reused (裁决
+	// 14.4.4 的 L3 输出缓存). Zero or negative disables reuse.
+	L3CacheTTL time.Duration
 }
 
 // Input is everything Decide needs to classify one round (design §4).
@@ -140,10 +151,13 @@ type Input struct {
 type Option func(*Decider)
 
 // Decider is the awakened-capacity decision engine. It is immutable after New
-// and safe for concurrent Decide calls.
+// and safe for concurrent Decide calls (the L3 cache is mutex-guarded).
 type Decider struct {
 	cfg Config
 	log *slog.Logger
+	// l3cache is the model-assisted intent layer's classifier + verdict cache.
+	// Nil when the layer is off, which is the default (裁决 14.4.4).
+	l3cache *l3Cache
 }
 
 // sourcedLog returns a logger tagged with the module source (project rule §10),
@@ -175,6 +189,9 @@ func New(cfg Config, opts ...Option) *Decider {
 	}
 	d := &Decider{cfg: cfg}
 	d.log = sourcedLog(nil)
+	if cfg.L3Enabled {
+		d.l3cache = newL3Cache(cfg.L3Classifier, cfg.L3CacheTTL)
+	}
 	for _, opt := range opts {
 		if opt != nil {
 			opt(d)
@@ -196,7 +213,7 @@ func (d *Decider) Decide(ctx context.Context, in Input) Decision {
 	if !d.cfg.Enabled {
 		return Decision{State: in.CurrentState, AllowedTools: append([]string(nil), in.AllToolNames...)}
 	}
-	state, conf, tool := d.intent(in)
+	state, conf, tool := d.intent(ctx, in)
 	j := judgmentFor(in, state, conf, tool)
 	state, conf = d.applyAwakening(state, conf, in)
 	allowed := d.allowedTools(state, in.AllToolNames)