feat(03-04): retire flat-file settings; wire per-user AI config via DB lookup

- config.py: Remove SETTINGS_FILE, DEFAULT_SYSTEM_PROMPT, DEFAULT_SETTINGS
  constants; add system_prompt, default_ai_provider, default_ai_model to Settings
- services/classifier.py: Add _DEFAULT_SYSTEM_PROMPT module constant; classify_document
  and suggest_topics_for_document accept ai_provider/ai_model kwargs; no longer calls
  storage.load_settings() — uses app_settings defaults with DB-supplied overrides (D-14, D-15)
- services/storage.py: Delete load_settings, save_settings, mask_api_key, settings_masked;
  remove from __all__; remove import copy, json, DEFAULT_SETTINGS, SETTINGS_FILE (D-12)
- tasks/document_tasks.py: _run resolves user.ai_provider/ai_model via session.get(User,
  doc.user_id) and passes through to classifier; task signature unchanged (T-03-19)
- api/settings.py: Deleted — /api/settings endpoint removed (D-12)
- main.py: Remove settings_router import and include_router call
- tests/test_settings.py: Replace all tests with test_settings_endpoint_removed (404, green)
- tests/test_classifier.py: Implement test_per_user_provider, test_celery_task_uses_user_provider,
  test_default_provider_fallback; remove xfail markers (DOC-03, DOC-05)

backend/config.py

@@ -1,5 +1,3 @@
-from pathlib import Path
-
 from pydantic_settings import BaseSettings, SettingsConfigDict
 
 
@@ -13,9 +11,6 @@ class Settings(BaseSettings):
         env_list_separator=",",
     )
 
-    # Data directory — used only for the flat-file settings.json path (Phase 1)
-    data_dir: str = "/app/data"
-
     # PostgreSQL
     database_url: str = "postgresql+psycopg://docuvault_app:changeme_app@postgres:5432/docuvault"
     database_migrate_url: str = "postgresql+psycopg://docuvault_migrate:changeme_migrate@postgres:5432/docuvault"
@@ -56,39 +51,10 @@ class Settings(BaseSettings):
     # Frontend URL — used to build password reset links (D-02, D-03)
     frontend_url: str = "http://localhost:5173"
 
+    # AI classification defaults (Phase 3 — D-13, D-15)
+    system_prompt: str = ""              # SYSTEM_PROMPT env var; hardcoded fallback lives in classifier.py
+    default_ai_provider: str = "ollama"  # DEFAULT_AI_PROVIDER env var
+    default_ai_model: str = "llama3.2"   # DEFAULT_AI_MODEL env var
+
 
 settings = Settings()
-
-# SETTINGS_FILE: still flat-file in Phase 1; migrates to users.ai_provider in Phase 2
-SETTINGS_FILE = Path(settings.data_dir) / "settings.json"
-
-DEFAULT_SYSTEM_PROMPT = """You are a document classification assistant. When given a document's text content and a list of existing topics, you must:
-1. Assign the document to one or more relevant topics from the list.
-2. If no existing topics fit well, suggest new topic names.
-Return ONLY valid JSON in this exact format, with no additional text or explanation:
-{"assigned_topics": ["topic1"], "new_topic_suggestions": ["new topic name"]}
-If the document fits no topics and you have no suggestions, return: {"assigned_topics": [], "new_topic_suggestions": []}"""
-
-DEFAULT_SETTINGS = {
-    "system_prompt": DEFAULT_SYSTEM_PROMPT,
-    "active_provider": "lmstudio",
-    "providers": {
-        "anthropic": {
-            "api_key": "",
-            "model": "claude-sonnet-4-6"
-        },
-        "openai": {
-            "api_key": "",
-            "model": "gpt-4o",
-            "base_url": None
-        },
-        "ollama": {
-            "base_url": "http://host.docker.internal:11434",
-            "model": "llama3.2"
-        },
-        "lmstudio": {
-            "base_url": "http://host.docker.internal:1234",
-            "model": "gemma-4-e4b-it"
-        }
-    }
-}