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/signals.go

@@ -1,6 +1,11 @@
+// signals.go holds the intent layer vocabulary and the L1/L2/L3 layers; the
+// package doc lives in capacity.go.
 package capacity
 
-import "strings"
+import (
+	"context"
+	"strings"
+)
 
 // defaultQuietTools is the companion-safe face (design §1): terminal + file edit
 // + memory/emotion companion. It is a function building the slice fresh so no
@@ -16,12 +21,13 @@ func defaultAwakeTools() []string {
 }
 
 // intent classifies the round. It runs L1 on the current turn, falls to L2 on
-// the recent context trend, and folds in the L4 learned-habit verdict (design §2.2)
-// at each layer: L4 may correct a decided result (boost / quiet→awake) or fill in
-// when L1/L2 are undecided. With no matching habit L4 is silent, so a turn with no
-// recall behaves exactly as before; otherwise the result stays conservative (低置信
-// still rests via applyAwakening).
-func (d *Decider) intent(in Input) (State, float64, string) {
+// the recent context trend, then to the model-assisted L3 layer when it is
+// enabled, and folds in the L4 learned-habit verdict (design §2.2) at each layer:
+// L4 may correct a decided result (boost / quiet→awake) or fill in when L1/L2/L3
+// are undecided. With no matching habit L4 is silent, so a turn with no recall
+// behaves exactly as before; otherwise the result stays conservative (低置信 still
+// rests via applyAwakening).
+func (d *Decider) intent(ctx context.Context, in Input) (State, float64, string) {
 	if len(in.Messages) == 0 {
 		return "", 0, ""
 	}
@@ -33,12 +39,44 @@ func (d *Decider) intent(in Input) (State, float64, string) {
 	if s, c, t, ok := l2(in.Messages); ok {
 		return d.mergeL4(s, c, t, l4s, l4c, l4t, l4ok)
 	}
+	if s, c, t, ok := d.l3(ctx, current); ok {
+		return d.mergeL4(s, c, t, l4s, l4c, l4t, l4ok)
+	}
 	if l4ok {
 		return l4s, l4c, l4t
 	}
 	return "", 0.3, "" // 拿不准 → conservative fallback
 }
 
+// l3 runs the model-assisted intent layer (design/14 §2.2 L3; 裁决 14.4.4): the
+// configured classifier judges the current turn and its verdict is trusted only
+// at or above the wake threshold — a low-confidence verdict is treated as
+// undecided, so a hesitant model can never wake tools on its own. A classifier
+// error is reported and treated the same way. With the layer off this returns
+// ok=false immediately, leaving L1/L2/L4 behavior byte-identical.
+func (d *Decider) l3(ctx context.Context, current string) (State, float64, string, bool) {
+	if d.l3cache == nil {
+		return "", 0, "", false
+	}
+	j, err := d.l3cache.classify(ctx, current)
+	if err != nil {
+		d.log.Warn("capacity L3 classifier unavailable, keeping the conservative default", "error", err)
+		return "", 0, "", false
+	}
+	if j.Confidence < d.cfg.MinConfidence {
+		d.log.Debug("capacity L3 verdict below threshold, treated as undecided",
+			"confidence", j.Confidence, "threshold", d.cfg.MinConfidence, "is_task", j.IsTask)
+		return "", 0, "", false
+	}
+	if !j.IsTask {
+		return StateQuiet, j.Confidence, "", true
+	}
+	if j.TaskKind == taskKindComplex {
+		return StateFull, j.Confidence, inferTool(current), true
+	}
+	return StateAwake, j.Confidence, inferTool(current), true
+}
+
 // l1 is the fast heuristic on the current turn (design §2.2 L1): a strong
 // action signal wakes (awake, or full for complex work); a clear companion
 // signal (weak + no task object) stays quiet; otherwise it is undecided and the