feat(bot): LLM gateway with GigaChat and offline mock providers

Single entry point app/llm/gateway.py: extract_item, extract_checkin,
risk_classify, phrase, summarize_for_doctor. Two-level design:
LLMProvider (transport, GigaChat) and LLMTasks (PromptedTasks over a
provider, or rule-based MockTasks). Shared policy in the gateway:
assert_scrubbed backstop, confidence threshold 0.5, risk failure -> NONE,
forbidden-output filter with config/texts/phrases.yaml fallbacks.

Refs #7

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

app/llm/base.py

@@ -0,0 +1,49 @@
+"""Provider and task protocols of the LLM layer.
+
+Two levels, so the offline mock does not have to reverse-engineer prompts:
+
+* `LLMProvider` — transport: "send system prompt + scrubbed user text, get JSON/text back".
+  Implemented by GigaChat (`giga.py`). `PromptedTasks` (`tasks.py`) turns it into tasks.
+* `LLMTasks` — task level: extract_item / extract_checkin / risk / phrase / summarize.
+  Implemented by `PromptedTasks(provider)` and by the rule-based `MockTasks` (`mock.py`).
+
+The gateway (`gateway.py`) wraps an `LLMTasks` with the shared policy: scrub guard, confidence
+threshold, forbidden-output filter, template fallbacks. Tasks may return `None` on any failure.
+"""
+
+from typing import Any, Protocol, TypeVar
+
+from pydantic import BaseModel
+
+from app.core.types import ScrubbedText
+from app.llm.schemas import CheckinFacts, ItemScore, RiskVerdict
+
+M = TypeVar("M", bound=BaseModel)
+
+
+class LLMProvider(Protocol):
+    async def complete_json(
+        self,
+        system: str,
+        user: ScrubbedText,
+        schema: type[M],
+        validation_context: dict[str, Any] | None = None,
+    ) -> M | None: ...
+
+    async def complete_text(self, system: str, user: ScrubbedText) -> str | None: ...
+
+
+class LLMTasks(Protocol):
+    name: str
+
+    async def extract_item(
+        self, scale_item_id: str, item_text: str, max_score: int, answer: ScrubbedText
+    ) -> ItemScore | None: ...
+
+    async def extract_checkin(self, text: ScrubbedText) -> CheckinFacts | None: ...
+
+    async def risk_classify(self, text: ScrubbedText) -> RiskVerdict | None: ...
+
+    async def phrase(self, intent: str, payload: ScrubbedText, base_text: str) -> str | None: ...
+
+    async def summarize_for_doctor(self, structured: ScrubbedText, base_text: str) -> str | None: ...