feat(session,runtime): 多角色并存—按会话选角（迁移0011 + 按回合解析 + agent池 + 提取器修正）

按 DevDocs/design/22 批次 1 后端（§2/§3/§4.1-4.5/§4.7/§5 验收 1-6）。
- 迁移 0011：chat_sessions.persona_id + chat_messages.persona_id（均 DEFAULT ''，无回填 → 升级零风险）；同步包内幂等 DDL；**真实库升级验证**（拷贝 data/cyrene.db 到临时目录跑 MigrateUp：既有 1 会话/109 消息 persona 仍为 ''，原库未被触碰）
- history：SessionMeta/Message 加 PersonaID + SetSessionPersona；ListSessions/List/scan/Append 带上
- runtime：SessionEnv.PersonaID + loop.personaFor(ctx) 替换固定 l.personaID（memory/request/turn）+ **agent 池**（Options.Agents，按角色取 instructions，未命中回落）+ **提取器不再把 persona 烤死在构造期**（多角色下记忆不再永远记在 cyrene 名下）+ **thinking 旋钮改为按会话角色解析**（否则角色 Y 会用角色 X 的预算/展示）
- gateway：会话 DTO/POST/PATCH 支持 personaId；ws/rest/voice 的 OwnerEnv 硬编改为**按会话解析角色**（actor 元数据语义不变）；admin memories 搜索去掉硬编 cyrene
- Docs/api-client v0.42；**两处有意的行为收紧**已写明：POST /sessions body 非法 JSON → 400（原先完全忽略 body；空/null/{} 仍 201）；POST /admin/memories/search 省略 personaId 且无默认角色 → 400（原先默认查 cyrene）
- 测试：design/22 §5 验收 1-6（记忆隔离/提示词首条 system/情绪按角色/提取 persona 归属/迁移兼容）+ 进程内 e2e（断言 adapter 收到的 model.Request.Messages[0] 即该角色的 systemPrompt）
- **已知未做（留批次 2 / 待确认）**：① IM 与渠道级未做 → **§7-3「owner 私聊镜像随主会话角色」尚未生效**；② 请求级覆盖未实现（本批只做会话级）；③ **§7-5 的「上下文发送角色标记」只完成存储+DTO 半边**——消息 persona 已落库且 DTO 已带，但模型上下文不从 chat_messages 重建，「第 N 条是谁说的」尚未进上下文（需确认是否属批次 2）；④「切角色重建摘要」当前是空操作（压缩未接线）；⑤ thinking 旋钮仅在流式路径生效（非流式 agent.RunStep 不带，A 批次既有性质）
- 验收：go build/vet/test 全绿 + linux/amd64 与 darwin/arm64 交叉编译（既有 flake 本轮未触发；同批 session_orchestrator.go 一度 501 行 → 抽出 session_tail.go）

internal/runtime/loop/spanlog.go

@@ -23,10 +23,12 @@ const (
 const maxLogSummaryRunes = 240
 
 // recordLLMDone appends a desensitized summary of a finished model call to its
-// span: the configured model, both stop reasons, how many tool calls were
+// span: the model that served the turn, both stop reasons, how many tool calls were
 // requested and how long the call took. No prompt or completion text is recorded,
-// so an llm node's log can never carry conversation content.
-func (l *Loop) recordLLMDone(span *trace.Span, out *stepOutcome) {
+// so an llm node's log can never carry conversation content. ag is the turn's agent
+// (the per-role pool entry), so a multi-role process reports the model that actually
+// served the call rather than the default role's.
+func (l *Loop) recordLLMDone(span *trace.Span, out *stepOutcome, ag *agent.Agent) {
 	if span == nil || out == nil {
 		return
 	}
@@ -35,7 +37,7 @@ func (l *Loop) recordLLMDone(span *trace.Span, out *stepOutcome) {
 		"tool_calls":  len(out.toolCalls),
 		"elapsed_ms":  elapsedMS(span),
 	}
-	if m := l.modelName(out); m != "" {
+	if m := modelName(out, ag); m != "" {
 		props["model"] = m
 	}
 	if out.modelStopReason != "" {
@@ -72,16 +74,16 @@ func (l *Loop) recordToolDone(span *trace.Span, tc agent.ToolCall, out ToolOutco
 }
 
 // modelName reports the model that served a step: the routed request model when
-// the streaming path knows it, else the loop's configured model. It is empty when
-// neither is available.
-func (l *Loop) modelName(out *stepOutcome) string {
+// the streaming path knows it, else the turn's agent's configured model. It is
+// empty when neither is available.
+func modelName(out *stepOutcome, ag *agent.Agent) string {
 	if out.model != "" {
 		return out.model
 	}
-	if l.agent == nil {
+	if ag == nil {
 		return ""
 	}
-	return l.agent.Model()
+	return ag.Model()
 }
 
 // elapsedMS reports how long a span has been open so far. A span's Duration is