fix(multimodal): 冻结推理环境并隔离迟到导入错误

backend/app/local_models/runtime.py

@@ -69,9 +69,10 @@ def configure(request):
     return request
 
 
-def interpreter():
+def interpreter(config=None):
     from app.local_models import components
-    if not os.getenv("APP_MODEL_PYTHON") and configuration().device == "cuda" and components.ready():
+    requested_device = (config or configuration()).device
+    if not os.getenv("APP_MODEL_PYTHON") and requested_device == "cuda" and components.ready():
         return components.ROOT / "Scripts/python.exe"
     return Path(os.getenv("APP_MODEL_PYTHON", str(BACKEND_DIR / ".venv-models" / ("Scripts/python.exe" if os.name == "nt" else "bin/python"))))
 
@@ -160,7 +161,8 @@ class Runtime:
     async def _execute(self, key, operation, payload, config, diagnostics):
         if read_state(key)["status"] != "installed":
             raise ProviderError("LOCAL_MODEL_NOT_INSTALLED", "请先下载本地模型。")
-        if not interpreter().is_file():
+        executable = interpreter(config)
+        if not executable.is_file():
             raise ProviderError("LOCAL_RUNTIME_NOT_INSTALLED", "请先安装本地模型运行环境。")
         from app.services.usage_service import UsageAttempt
         attempt = UsageAttempt("local-models", CATALOG[key].repository, "local", operation, source="local")
@@ -170,7 +172,7 @@ class Runtime:
             env = {**os.environ, "HF_HUB_OFFLINE": "1", "TRANSFORMERS_OFFLINE": "1",
                    "HF_HUB_DISABLE_TELEMETRY": "1", "OMP_NUM_THREADS": str(config.cpu_threads),
                    "PYTHONIOENCODING": "utf-8"}
-            args = (str(interpreter()), str(Path(__file__).with_name("worker.py")))
+            args = (str(executable), str(Path(__file__).with_name("worker.py")))
             options = {"env": env, "limit": 16 * 1024 * 1024,
                        **({"creationflags": 0x08000000} if os.name == "nt" else {})}
             try:

backend/tests/test_local_models.py

@@ -47,7 +47,7 @@ def test_local_model_missing_is_explicit():
 def test_cancel_reaps_active_model_process(monkeypatch):
     import app.local_models.runtime as module
     monkeypatch.setattr(module,'read_state',lambda key:{'status':'installed'})
-    monkeypatch.setattr(module,'interpreter',lambda:Path(sys.executable))
+    monkeypatch.setattr(module,'interpreter',lambda *_:Path(sys.executable))
     class Input:
         def write(self, value):
             request = json.loads(value)
@@ -92,7 +92,7 @@ def test_subprocess_fallback_runs_and_reaps_real_worker(monkeypatch, tmp_path, c
     import app.local_models.process as process_module
 
     monkeypatch.setattr(module, 'read_state', lambda key: {'status': 'installed'})
-    monkeypatch.setattr(module, 'interpreter', lambda: Path(sys.executable))
+    monkeypatch.setattr(module, 'interpreter', lambda *_: Path(sys.executable))
     worker = tmp_path / 'worker.py'
     worker.write_text(
         'import json,sys,time\n'

backend/tests/test_multimodal_finalization.py

@@ -21,7 +21,7 @@ def test_cuda_retries_only_device_failures_in_reaped_process(monkeypatch, code,
     from app.services.usage_service import connection
     monkeypatch.setattr(module, 'configuration', lambda: module.RuntimeConfig(device='cuda'))
     monkeypatch.setattr(module, 'read_state', lambda key: {'status': 'installed'})
-    monkeypatch.setattr(module, 'interpreter', lambda: Path(sys.executable))
+    monkeypatch.setattr(module, 'interpreter', lambda *_: Path(sys.executable))
     events = []
 
     class Process:

backend/tests/test_runtime_components.py

@@ -79,7 +79,8 @@ def test_interpreter_keeps_cpu_default_and_respects_explicit_override(monkeypatc
     (components.ROOT / 'ready.json').write_text('{}')
     monkeypatch.setattr(runtime, 'configuration', lambda: runtime.RuntimeConfig(device='cpu'))
     assert runtime.interpreter() != python
-    monkeypatch.setattr(runtime, 'configuration', lambda: runtime.RuntimeConfig(device='cuda'))
-    assert runtime.interpreter() == python
+    # A queued attempt keeps its frozen device even after the saved setting changes.
+    assert runtime.interpreter(runtime.RuntimeConfig(device='cuda')) == python
+    assert runtime.interpreter(runtime.RuntimeConfig(device='cpu')) != python
     monkeypatch.setenv('APP_MODEL_PYTHON', 'explicit-python.exe')
     assert str(runtime.interpreter()) == 'explicit-python.exe'

docs/development/阶段F收尾验收记录.md

@@ -13,7 +13,7 @@
 | 项目 | 结果 |
 | --- | --- |
 | 后端全量 `python -m pytest -q -p no:cacheprovider` | 559 通过；1 条已有 Starlette/httpx 弃用提示 |
-| 前端全量 `npm test -- --run` | 29 个文件、102 项通过 |
+| 前端全量 `npm test -- --run` | 29 个文件、103 项通过 |
 | 类型与生产构建 `npm run build` | vue-tsc 与 Vite 构建通过，仍有既有大 bundle 提示 |
 | `git diff --check` | 通过 |
 | 真实页面 | 模型卡片读取实际大小；音频统计显示真实缺失；提供商表单展示请求编辑、恢复默认、导入/导出及推理验证入口 |

docs/retrospectives/阶段F-Embedding与知识库问题与解决方案.md

@@ -138,6 +138,14 @@ embedding_local_only: true
 
 ## 9. 工程经验
 
+### F-18：设备快照与迟到导入错误未完全隔离
+
+问题：推理任务虽然冻结了 RuntimeConfig，但启动子进程时又从数据库读取最新 device 来选择 Python 环境；排队期间修改设置会改变已提交任务的运行环境，CUDA → CPU 重试也可能继续使用 CUDA 环境。请求规则的迟到成功响应已失效，但迟到失败仍会把旧错误显示到新草稿。
+
+实际方案：`interpreter` 接收本次 attempt 的冻结配置，`_execute` 在检查前解析一次可执行路径并复用；CUDA attempt 使用已验证的独立 CUDA 环境，CPU attempt 使用默认 CPU 环境，显式 APP_MODEL_PYTHON 仍保持最高优先级。导入异常与成功响应使用同一 generation 条件，只允许当前操作更新界面。
+
+验证：增加保存设置变化后仍按显式 attempt 选择环境、CPU 重试环境，以及旧导入失败晚于新编辑的回归。最终后端 559 项、前端 103 项和生产构建通过。
+
 ### F-16：CUDA 选装只有脚本，前端缺少安装入口
 
 问题：上一轮完成了独立 CUDA 环境安装和 GPU 实测，但页面只有设备下拉框及脚本说明。用户无法从前端下载组件，工程收尾遗漏了可操作入口。

frontend/src/features/settings/RequestJsonEditor.spec.ts

@@ -1,5 +1,5 @@
 // @vitest-environment happy-dom
-import { mount } from '@vue/test-utils'
+import { flushPromises, mount } from '@vue/test-utils'
 import { expect, it, vi } from 'vitest'
 import RequestJsonEditor from './RequestJsonEditor.vue'
 import { apiClient } from '@/services/apiClient'
@@ -38,6 +38,21 @@ it('ignores an imported configuration that finishes after a newer edit', async (
   wrapper.unmount()
 })
 
+it('ignores an old import failure after a newer edit', async () => {
+  let fail!: (reason: Error) => void
+  vi.mocked(apiClient.post).mockReturnValue(new Promise((_resolve, reject) => { fail = reject }))
+  const wrapper = mount(RequestJsonEditor, {props:{modelValue:[]}})
+  const input = wrapper.get('input[type="file"]')
+  Object.defineProperty(input.element, 'files', {value:[new File(['{}'], 'old.json')], configurable:true})
+  await input.trigger('change')
+  await wrapper.findAll('button').find(button => button.text() === '添加请求规则')!.trigger('click')
+  fail(new Error('旧导入失败'))
+  await flushPromises()
+  expect(wrapper.text()).not.toContain('旧导入失败')
+  expect(wrapper.findAll('textarea')).toHaveLength(1)
+  wrapper.unmount()
+})
+
 it('restores defaults even from an invalid draft and reflects replacement configurations', async () => {
   const wrapper = mount(RequestJsonEditor, {props:{modelValue:[{capability:'chat', body:{enable_thinking:false}}]}})
   await wrapper.get('textarea').setValue('{invalid')

frontend/src/features/settings/RequestJsonEditor.vue

@@ -51,7 +51,7 @@ async function importRules(event: Event) {
     if (current !== generation) return
     rules.value = validated.request_overrides.map(rule => ({...rule, draft: JSON.stringify(rule.body, null, 2), error: ''}))
     publish()
-  } catch(e) { transferError.value = (e as Error).message }
+  } catch(e) { if (current === generation) transferError.value = (e as Error).message }
 }
 async function exportRules() {
   transferError.value = ''