feat(llm): провайдер за протоколом — OpenAI-совместимый API и mock, фабрика по конфигу

- LLMProvider: complete(messages, schema) -> LLMResult (текст, разобранный JSON, токены)
- openai_compat: /chat/completions, structured output через response_format json_schema,
  JSON из ```-обёртки тоже разбирается; ретраи на 429/5xx и сетевые ошибки, 4xx — сразу ошибка
- mock: детерминированные правила для тестов и демо без сети
- LLMSettings (LLM_PROVIDER/BASE_URL/MODEL/API_KEY…), CLI python -m app.llm.cli

Refs #17

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

app/llm/mock.py

@@ -0,0 +1,36 @@
+"""Deterministic offline provider for tests and demos without network or tokens.
+
+Rules are checked in order: the first whose `contains` is found in the last user message wins.
+Without a match it echoes the message (or returns `{}` when JSON was requested).
+"""
+
+import json
+from dataclasses import dataclass
+from typing import Any
+
+from app.llm.provider import ChatMessage, JsonSchema, LLMResult
+
+
+@dataclass(frozen=True)
+class MockRule:
+    contains: str
+    reply: str | dict[str, Any]
+
+
+class MockProvider:
+    def __init__(self, rules: list[MockRule] | None = None) -> None:
+        self.rules = list(rules or [])
+        self.calls: list[tuple[list[ChatMessage], JsonSchema | None]] = []
+
+    async def aclose(self) -> None:
+        pass
+
+    async def complete(self, messages: list[ChatMessage], schema: JsonSchema | None = None) -> LLMResult:
+        self.calls.append((messages, schema))
+        last = next((m.content for m in reversed(messages) if m.role == "user"), "")
+        reply: str | dict[str, Any] | None = next((r.reply for r in self.rules if r.contains in last), None)
+        if reply is None:
+            reply = {} if schema is not None else f"mock: {last}"
+        text = json.dumps(reply, ensure_ascii=False) if isinstance(reply, dict) else reply
+        data = reply if isinstance(reply, dict) and schema is not None else None
+        return LLMResult(text=text, data=data, prompt_tokens=len(last.split()), completion_tokens=len(text.split()))