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/run.go

@@ -187,6 +187,11 @@ func (l *Loop) driveTurn(ctx context.Context, sessionID string, msgs []model.Mes
 		}
 		msgs = batch.messages
 		res.ToolCallsExecuted += batch.executed
+		// speak 的出站结论逐批回填（M57）：后面批次里没有 speak 时保留
+		// 前面那批的结论 —— 这条回复想用语音这件事发生在哪一批都算数。
+		if batch.voice != nil {
+			res.Voice = batch.voice
+		}
 		l.saveCheckpoint(ctx, sessionID, step, msgs, agent.StopReasonToolCalls)
 
 		// 关键工具循环安全网（或终止钩子）不等真正的最终答案