feat(llm-completion): support Anthropic-specific message roles in ChatML

[marliessophie]

Describe the feature or potential improvement

Problem

Experiments using dataset items with Anthropic message roles result in unexpected behaviour. Any message role not matching OpenAI's schema gets force-converted to assistant.

Root Cause

Langfuse currently only recognizes OpenAI message roles:

• user → HumanMessage
• system/developer → SystemMessage
• Everything else → AIMessage (defaults to assistant)

This hardcoded OpenAI-first approach ignores that Anthropic uses different conventions:

• OpenAI: system, user, assistant, tool
• Anthropic: user, assistant (no system role in messages array; uses separate system parameter)

Dataset items with Anthropic's native role names are silently mangled.

Proposed Solution

Extend ChatMessageRole enum in packages/shared/src/server/llm/types.ts (line 120):

export enum ChatMessageRole {
  System = "system",
  Developer = "developer",
  User = "user",
  Assistant = "assistant",
  Tool = "tool",
  Model = "model",
  // Anthropic compatibility
  Human = "human", // Anthropic user equivalent
  AI = "ai", // Anthropic assistant equivalent
}

Update message transformation logic in fetchLLMCompletion.ts to map these roles correctly per provider.

Additional information

No response