fix(anthropic): avoid Fable OAuth system prompt refusal

[heyrry]

Summary

• use the native billing + Claude Code identity system-block shape for Anthropic Fable OAuth requests
• omit the generic CLI expansion block that makes Fable stop with refusal and zero output tokens
• preserve the original client system instructions through the existing message-history migration
• keep the current configured/default system-block behavior unchanged for non-Fable models

Problem

For a non-Claude-Code client routed through an Anthropic OAuth account, claude-fable-5 can return HTTP 200 with an empty public response. The upstream Anthropic SSE contains no content block, reports stop_reason=refusal, and has zero output tokens. The same account, model, and prompt work when the request uses the native billing + Claude Code identity shape.

This affects both the OpenAI-compatible Chat Completions and Responses forwarding paths because both call applyClaudeCodeOAuthMimicryToBody.

Validation

• make test-unit
• added a regression test covering the Fable-specific system-block shape and preservation of the original client system instructions
• sanitized live Anthropic OAuth verification: Chat Completions returned non-empty OK with output tokens; Responses returned non-empty output_text with output tokens

No credentials, account identifiers, deployment configuration, or response content beyond the fixed test token are included.

[github-actions]

All contributors have signed the CLA. ✅
_{Posted by the CLA Assistant Lite bot.}