feat: 完成 F 群組 — 對話生成管線與 LLM Provider 抽象

- LLMProvider 介面（generate/stream）+ LLM_PROVIDER 切換（llm.module.ts useFactory，
  LLM_PROVIDER=mock|claude 環境變數決定注入哪個實作）
- MockProvider：依性格原型 × 情緒狀態 × 親密度從模板庫選填回應，seeded PRNG 確保同 seed 可重現，
  支援 ＊動作描寫＊ 標記抽取（目前只有元氣原型模板，G 群組需補齊其餘五種原型）
- ClaudeProvider 空殼：LLM_PROVIDER=claude 時啟動即輸出 ERR log，呼叫 generate/stream 時拋出例外
- OutputFilterService：全域＋角色禁則詞彙過濾（攔截並改寫）
- FastChannelDetector：固定問候與危險輸入（含求助資源)走快速通道，跳過記憶檢索
- BehaviorReinforcementService：稱讚加權、糾正後修正版繼承舊權重以確保排序領先（實作時抓到並修正一個
  「降權後打平」的真實 bug）
- ContextAssemblerService：組裝人設／情緒／關係／檢索記憶／對話歷史為統一上下文快照
- DialogueService/Controller：串起以上所有元件，作為 H-1 正式對話端點的基礎
- scripts/smoke/F.mjs：涵蓋可重現性、上下文快照完整性、雙速通道、行為強化、輸出過濾、
  以 child process 驗證 Provider 切換

npm run restart && npm run smoke -- F 皆通過（F-V），A/B/C/D/E 群組冒煙測試無回歸。

apps/api/src/llm/dialogue.service.ts

@@ -0,0 +1,79 @@
+import { Inject, Injectable } from "@nestjs/common";
+import { PrismaService } from "../prisma/prisma.service.js";
+import { WorkingMemoryService } from "../memory/working-memory.service.js";
+import { LLM_PROVIDER, type LLMProvider } from "./llm-provider.js";
+import { ContextAssemblerService } from "./context-assembler.service.js";
+import { OutputFilterService } from "./output-filter.service.js";
+import { FastChannelDetector, DEFAULT_FAST_REPLIES } from "./fast-channel.service.js";
+import { extractActions } from "./action-markup.js";
+import type { GenerationContext } from "./types.js";
+
+export interface HandleMessageOptions {
+  seed?: number;
+  now?: Date;
+}
+
+export interface HandleMessageResult {
+  text: string;
+  actions: string[];
+  isFastChannel: boolean;
+  context: GenerationContext;
+}
+
+function fastChannelSituation(reason: string): string {
+  return `fast-channel:${reason}`;
+}
+
+// F 群組整合入口：串起雙速通道判斷、上下文組裝、Provider 生成、輸出過濾與工作記憶寫入。
+// H-1 的 POST /chat/:characterId 會以此為基礎擴充（見 todo.md 實作記錄）。
+@Injectable()
+export class DialogueService {
+  constructor(
+    @Inject(LLM_PROVIDER) private readonly llmProvider: LLMProvider,
+    private readonly fastChannel: FastChannelDetector,
+    private readonly contextAssembler: ContextAssemblerService,
+    private readonly outputFilter: OutputFilterService,
+    private readonly workingMemory: WorkingMemoryService,
+    private readonly prisma: PrismaService,
+  ) {}
+
+  async handleMessage(
+    characterId: string,
+    userId: string,
+    sessionId: string,
+    userInput: string,
+    options: HandleMessageOptions = {},
+  ): Promise<HandleMessageResult> {
+    const now = options.now ?? new Date();
+    this.workingMemory.append(sessionId, { role: "user", content: userInput, timestamp: now });
+
+    const { isFastChannel, reason } = this.fastChannel.detect(userInput);
+
+    const context = await this.contextAssembler.assemble(characterId, userId, sessionId, userInput, {
+      skipMemoryRetrieval: isFastChannel,
+      seed: options.seed,
+      now,
+    });
+
+    let text: string;
+    let actions: string[];
+
+    if (isFastChannel && reason) {
+      const rule = await this.prisma.client.proceduralRule.findFirst({
+        where: { characterId, situation: fastChannelSituation(reason) },
+        orderBy: { weight: "desc" },
+      });
+      text = rule?.responsePattern ?? DEFAULT_FAST_REPLIES[reason];
+      actions = extractActions(text);
+    } else {
+      const generated = await this.llmProvider.generate(context);
+      text = generated.text;
+      actions = generated.actions;
+    }
+
+    const filteredText = this.outputFilter.filter(text);
+    this.workingMemory.append(sessionId, { role: "character", content: filteredText, timestamp: now });
+
+    return { text: filteredText, actions, isFastChannel, context };
+  }
+}