feat(multimodal): 视觉分层路由（多模态检测→主/视觉/OCR 工具）+ 语音 STT→TTS 降级链（异步）

internal/runtime/multimodal/doc.go

@@ -0,0 +1,17 @@
+// Package multimodal groups the runtime's non-text input/output handling:
+// image understanding (vision) and speech (voice).
+//
+// It follows DevDocs/design/06-多模态.md:
+//
+//   - The vision subpackage routes image messages through a layered strategy
+//     (main model first, then a tool-shaped vision/OCR adapter) and registers
+//     those adapters as sdk.Tool capabilities so the M2 tool loop can invoke
+//     them without putting the image in the main-model context.
+//   - The voice subpackage provides asynchronous STT/TTS through a degrading
+//     provider chain (cloud → local → edge-tts/espeak → silent).
+//
+// Both subpackages keep real provider calls behind interfaces with an HTTP
+// default implementation, so they are testable with a stdlib httptest server
+// and leave the real cloud/local providers as adaptation points for later
+// milestones.
+package multimodal