feat(llm): provider-abstract gateway with cost and failure controls

DeepSeek (OpenAI-compatible) as primary with a 智谱 GLM slot that is a real
routing path but raises until the 算法备案 model list is confirmed. Adds scene
routing, failover, a per-provider circuit breaker, structured output with
retry-on-validation-failure, prompts loaded by id+version from prompts/, and
cost attribution written to ob_ai_call for every attempt -- successes,
retries and failures alike.

Two controls exist because they are cheap now and expensive later: a
cache-prefix validator that rejects timestamps, uuids and nicknames in the
cached system prefix (a volatile prefix silently multiplies input cost), and
a hard platform-wide daily budget that rejects non-critical scenes instead of
merely alerting, since a retry storm outruns any human on-call.

packages/llm/providers/base.py

@@ -0,0 +1,34 @@
+"""Provider abstraction.
+
+架构文档 §4.3 requires the gateway interface to be designed as if it were a
+remote call, so providers take a fully-specified request (explicit timeout,
+serialisable payload) and return a fully-specified response. No provider is
+allowed to reach back into gateway state.
+"""
+
+from __future__ import annotations
+
+from typing import Protocol, runtime_checkable
+
+from packages.llm.types import ProviderRequest, ProviderResponse
+
+
+@runtime_checkable
+class LLMProvider(Protocol):
+    @property
+    def name(self) -> str:
+        """Stable identifier used in routing config and in ``ob_ai_call``."""
+        ...
+
+    @property
+    def default_model(self) -> str:
+        """Model used when the scene config does not pin one."""
+        ...
+
+    async def complete(self, request: ProviderRequest) -> ProviderResponse:
+        """Perform one completion. Must raise ``packages.llm.errors.ProviderError``."""
+        ...
+
+    async def aclose(self) -> None:
+        """Release connections."""
+        ...