feat(agent): support Responses steering and project-aware tools

Use the dedicated TJWATER_AGENT_URL and Responses stream protocol. Keep running
turns active while users steer or queue follow-ups and edit pending delivery.
Refresh the model catalog before generation and preserve approval mode selection.

Integrate scoped investigation messaging/presentation, geographic place selection,
current Agent composer styles and supporting workbench contracts.

Validation: 77 focused Vitest checks, 21 browser cases, lint, typecheck and
production build. Preserve direct Server/Agent routing and project authentication.

tests/browser/agent-voice-hydration.e2e.ts

@@ -97,7 +97,7 @@ test("speech recognition appends every final transcript in one result batch", as
 
   await page.getByRole("button", { name: "语音输入" }).click();
 
-  await expect(page.getByPlaceholder("输入调度问题，Agent 将通过后端会话流式响应")).toHaveValue("检查城南泵站 水位");
+  await expect(page.getByPlaceholder("询问供水问题，或描述需要处理的任务")).toHaveValue("检查城南泵站 水位");
 });
 
 test("speech recognition pulse stays transparent across its loop boundary", async ({ page }) => {