loop: shared bounded-turn core accepts history + knobs (ServeTurn)

Turn machinery refactored so the upcoming serve front door reuses it
instead of copying it: the round loop is now runTurn over arbitrary
message history with an optional tool palette and forwarded max_tokens/
temperature. The serve path (ServeTurn) is stateless chat over client
history with tools hard-off: a tool call the model produces anyway is
refused as a tool result, never executed. Also exposes the model router
(so serve can build its catalog) and a sorted class list; llm requests
can now carry temperature.

💘 Generated with Crush

Assisted-by: Crush:glm-5.2

internal/llm/client.go

@@ -50,6 +50,9 @@ type ChatRequest struct {
 	Messages  []Message `json:"messages"`
 	Tools     []Tool    `json:"tools,omitempty"`
 	MaxTokens int       `json:"max_tokens,omitempty"`
+	// Temperature is optional sampling heat; nil = provider default.
+	// Forwarded as-is (the serve front door passes the client's value).
+	Temperature *float64 `json:"temperature,omitempty"`
 }
 
 type ChatResponse struct {