fix: 记忆抽取关闭 Qwen 思考模式 + 测试模型与生产配置对齐

[lishuceo]

Summary

• 生产 extractor 和测试中添加 enable_thinking: false，记忆抽取是结构化输出任务不需要 CoT 推理
• 测试硬编码的 qwen3.5-plus（思考模型）改为 process.env.MEMORY_EXTRACTION_MODEL || 'qwen3.5-flash'，与生产配置对齐
• quality.test.ts 抽取 case 超时兜底从 120s 放宽到 240s

Test plan

• npx vitest run src/memory/__tests__/quality.test.ts — 抽取 case 全部通过，耗时从 >310s 降至 ~6s
• 运行 npx vitest run 全量测试通过

🤖 Generated with Claude Code

[claude]

Claude finished @lishuceo's task in 49s —— View job

---

PR Review: 记忆抽取关闭 Qwen 思考模式

• Check for previous review comments — none found
• Read changed files and understand context
• Review changes
• Post summary

---

✅ Approved — Clean, focused fix. No issues found.

Changes reviewed:

1. src/memory/extractor.ts — Added enable_thinking: false to the DashScope API call with as never type assertion. This is correct: memory extraction is a structured JSON output task where CoT reasoning adds latency without benefit. The as never pattern is consistent with the existing codebase (utils/quick-ack.ts:105).

2. src/memory/__tests__/integration.test.ts — Model changed from hardcoded qwen3.5-plus to process.env.MEMORY_EXTRACTION_MODEL || 'qwen3.5-flash', aligning with production config. Also added enable_thinking: false + as never to match production behavior.

3. src/memory/__tests__/quality.test.ts — Same model and thinking mode alignment. Timeout bumped from 120s → 240s as a safety margin for CI, though with thinking disabled the actual runtime should be much faster (~6s per PR description).

Assessment: All three changes are minimal, correct, and well-motivated. Tests mirror production config, as never is the established pattern for DashScope-specific extensions not in OpenAI's type definitions, and the performance improvement (310s → ~6s) is significant.