feat(loop): wire speak tool into the loop (M57)

- speak_tool.go: the speak tool itself — rune-derived timeout budget, structured
  failure codes, persona/emotion stamped on the synthesis ctx, audio persisted
  through the store seam, structured result for the model.
- voice_deps.go: the wiring seam (VoiceSynth / VoiceStore + VoiceStoreFunc
  adapter), voiceEnabled, withSpeakSpec (appended after every schema shrink),
  and captureVoice (success beats failure, last one wins).
- options.go / loop.go: VoiceSynth, VoiceStore, SpeakBudget injection points.
- request.go: advertise speak on both toolSchemas return paths.
- tools.go / tools_invoke.go: carry the batch voice outcome; split invokeOne /
  emitToolExecution / spanStatusForToolPhase / invokeParallel into their own
  file (tools.go was at the 500-line ceiling) and let speak run without a
  dispatcher.
- run.go: keep the voice outcome across tool batches (a later batch without
  speak must not erase it).

Known, deliberate: speak bypasses the capacity allow-set and skill-round
narrowing (it belongs to no built-in toolset), so it is also advertised in
quiet/awake; execution still passes the tool security gate, where speak is not
registered and non-owner actors are denied fail-closed.

internal/runtime/loop/tools_invoke.go

@@ -0,0 +1,107 @@
+package loop
+
+// 本文件承载**单个/一批工具调用的执行**（span、流式帧、结果归一化、
+// 有界并行），与 tools.go 的分类、准入与回渲职责分开 —— 从 tools.go
+// 抽出，使两边都留在 500 行以内（AGENTS.md §10）。
+
+import (
+	"context"
+	"fmt"
+	"sync"
+
+	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/agent"
+	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/model"
+	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/tool"
+	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/trace"
+)
+
+// invokeOne 在其每工具截止时间下执行单个工具调用，发出一个
+// tool span、以 tool_execution.start/end 事件流式输出执行进度
+// 并归一化结果。执行全程由单一 phase 决策驱动：start 帧带
+// running，Invoke 返回后按成败取 done/error，同一取值既决定
+// end 帧的 phase，也决定 span 终态（design/19 的 tool 节点
+// running/done/error 状态即由此驱动）。
+func (l *Loop) invokeOne(ctx context.Context, d *tool.Dispatcher, tc agent.ToolCall) ToolOutcome {
+	// speak 由循环自己承载（见 invokeTool），没有 dispatcher 也能执行；其余
+	// 工具缺 dispatcher 时仍按原样短路，不为此多开 span 或流式帧。
+	if d == nil && !(l.voiceEnabled() && tc.Name == speakToolName) {
+		return ToolOutcome{Err: fmt.Errorf("no tool dispatcher configured for %q", tc.Name), ErrKind: tool.KindExecution}
+	}
+	_, toolSpan := l.span(ctx, "tool."+tc.Name, trace.SpanKindTool)
+	l.emitToolExecution(ctx, tc, model.StreamKindToolExecutionStart, model.ToolPhaseRunning, "")
+	// 分派见 invokeTool：路标识桥（tool.WithLegID）与 speak 的循环内承载都在
+	// 那里。路标识（design/15 §2.5.7、主人裁决 3）：网关在**每一条在途路**上把
+	// 互斥身份经 loop.WithToolLegID 挂在 ctx 上（普通回合 / 并行路 / 队首续跑路
+	// 都挂，取值 = 该路的 turnID），工具层凭 tool.LegIDFrom 读到「我是哪一路」
+	// 做文件级互斥；未挂载 ⇒ 空串 ⇒ 完全不做互斥，行为逐字不变（兼容硬要求）。
+	//
+	// 刻意**不**读 ParallelLegFrom：那条通道只有真并行路才有、且只决定 checkpoint
+	// 归属（checkpoint.go 的 checkpointAgentID），拿它当互斥身份会让普通回合在
+	// 「1 普通回合 + 1 并行路」时完全不受保护。
+	result, err := l.invokeTool(ctx, d, tc)
+	phase := model.ToolPhaseDone
+	if err != nil {
+		phase = model.ToolPhaseError
+	}
+	l.emitToolExecution(ctx, tc, model.StreamKindToolExecutionEnd, phase, toolEndFrameSummary(result, err))
+	out := ToolOutcome{Result: result, Err: err, ErrKind: errorKind(err)}
+	l.spanProp(toolSpan, "phase", phase)
+	l.recordToolDone(toolSpan, tc, out)
+	if err != nil {
+		l.spanProp(toolSpan, "error", err.Error())
+		l.spanProp(toolSpan, "error_kind", out.ErrKind)
+	} else {
+		recordToolResult(toolSpan, result)
+	}
+	l.finishSpan(toolSpan, spanStatusForToolPhase(phase))
+	return out
+}
+
+// emitToolExecution 在循环处于流式路径（配置了 OnStream）时把
+// tool_execution 三事件（start / update / end）转发给前端；否则为
+// 空操作（被静音的内部流程也会跳过它）。phase 是结构化字段
+// （running/done/error），绝不塞进 Delta；工具名走 Tool 字段，
+// ID 复用 provider 的 tool call id，使执行帧能与其 tool_call.* 帧关联；
+// resultSummary 只在 end 帧、且只在结果真正拿到时非空（口径见 tool_frame_summary.go）。
+func (l *Loop) emitToolExecution(ctx context.Context, tc agent.ToolCall, kind model.StreamKind, phase, resultSummary string) {
+	if l.onStream == nil || isMutedBroadcast(ctx) {
+		return
+	}
+	ev := model.StreamEvent{Kind: kind, Phase: phase, Tool: tc.Name, ID: tc.ID, ResultSummary: resultSummary}
+	ev.SessionID = loopSession(ctx)
+	ev.TurnID = loopTurn(ctx)
+	l.onStream(ev)
+}
+
+// spanStatusForToolPhase 把工具执行的终态 phase 映射为 span 终态，
+// 使 trace/流程图与 WS 帧从同一个 phase 取值得出一致结论。
+// 非终态 phase（running）不会走到这里 —— span 只在执行结束时收尾。
+func spanStatusForToolPhase(phase string) trace.SpanStatus {
+	if phase == model.ToolPhaseError {
+		return trace.SpanStatusError
+	}
+	return trace.SpanStatusOK
+}
+
+// invokeParallel 并发运行 idxs 处的调用，上限为循环的
+// 并行度，并把每个结果写回 out 的对应调用索引处，使调用方
+// 能按原始顺序渲染结果。信号量限制并发，在途
+// goroutine 完成时释放其槽位，因此被取消的 context
+// （它使每个 Invoke 迅速返回）不会泄露 goroutine。
+func (l *Loop) invokeParallel(ctx context.Context, d *tool.Dispatcher, calls []agent.ToolCall, idxs []int, out []ToolOutcome) {
+	if len(idxs) == 0 {
+		return
+	}
+	sem := make(chan struct{}, l.toolParallelism)
+	var wg sync.WaitGroup
+	for _, i := range idxs {
+		wg.Add(1)
+		sem <- struct{}{}
+		go func(i int) {
+			defer wg.Done()
+			defer func() { <-sem }()
+			out[i] = l.invokeOne(ctx, d, calls[i])
+		}(i)
+	}
+	wg.Wait()
+}