llm: round-trip identity test, anthropic → anthropic

[jfallows]

Summary

Mirrors the openai → openai round-trip identity test (M2) for the Anthropic dialect: llm server (dialect: anthropic) → llm client (dialect: anthropic) against the mock Anthropic backend reproduces the original stream with no payload parsed.

Scope

• Same assertion shape as the openai round trip: framing decode/re-encode fidelity (byte-identical or defined-semantic-equivalent, per the SSE decoder's resolved open question) and no payload parsed.
• Cover both streaming and non-streaming shapes, multi-block content, and tool-call streaming (input_json_delta).

Acceptance criteria

• IT test passes against the mock Anthropic backend, structured to mirror the openai round-trip test for easy comparison.

Depends on

Mock Anthropic backend (this milestone), dialect: anthropic (this milestone).