feat(classifier): optional GigaChat-2-Lite hybrid tier classify

LLM classify via LiteLLM gigachat-classifier when heuristic confidence
is low. CLASSIFIER_MODE=heuristic|hybrid|llm. Metrics classifier_source.

router/orchestrator.py

@@ -11,6 +11,7 @@ from dataclasses import dataclass, field
 from enum import Enum
 from typing import Any
 
+from classifier_llm import LlmClassifier
 from rules_loader import load_model_matrix, load_orchestration, load_routing_rules
 
 log = logging.getLogger("orchestrator")
@@ -54,6 +55,7 @@ class RouteDecision:
     escalation_level: int
     confidence: float
     delegated_internal: bool = False
+    classifier_source: str = "heuristic"
 
 
 class SessionStore:
@@ -172,6 +174,7 @@ class Orchestrator:
     def __init__(self, session_store: SessionStore) -> None:
         self.sessions = session_store
         self.classifier = Classifier()
+        self.llm_classifier = LlmClassifier()
         self._orch = load_orchestration()
         self._matrix = load_model_matrix()
         self._models: dict[str, dict] = self._matrix.get("models", {})
@@ -217,7 +220,7 @@ class Orchestrator:
         suffix = TIER_SUFFIX[tier]
         return f"{prefix}-{suffix}"
 
-    def resolve(
+    async def resolve(
         self,
         messages: list[dict[str, Any]],
         *,
@@ -229,6 +232,17 @@ class Orchestrator:
     ) -> RouteDecision:
         mode = quality_mode or self.default_quality
         tier, confidence = self.classifier.classify(messages, has_image=has_image, text=text)
+        classifier_source = "heuristic"
+
+        if self.llm_classifier.should_use_llm(tier.value, confidence, has_image):
+            llm_result = await self.llm_classifier.classify(text)
+            if llm_result:
+                tier = Tier(llm_result[0])
+                confidence = llm_result[1]
+                classifier_source = "gigachat"
+            elif self.llm_classifier.mode == "llm":
+                classifier_source = "heuristic_fallback"
+
         ctx = self.sessions.get(session_id)
         prompt_hash = hashlib.sha256(text.encode()).hexdigest()[:16]
 
@@ -263,11 +277,13 @@ class Orchestrator:
                 escalation_level=escalation,
                 confidence=confidence,
                 delegated_internal=True,
+                classifier_source=classifier_source,
             )
 
-        delegated = confidence < self.classifier._low_conf and tier in (
-            Tier.MEDIUM_OPS,
-            Tier.SIMPLE,
+        delegated = (
+            classifier_source == "heuristic"
+            and confidence < self.classifier._low_conf
+            and tier in (Tier.MEDIUM_OPS, Tier.SIMPLE)
         )
 
         return RouteDecision(
@@ -278,6 +294,7 @@ class Orchestrator:
             escalation_level=escalation,
             confidence=confidence,
             delegated_internal=delegated,
+            classifier_source=classifier_source,
         )
 
     def after_request(