Claude API adds streaming refusal stop_reason

This is a small API change with outsized product impact: it turns a safety event into a first-class stream outcome, which is cleaner for Anthropic but can break clients that only expect normal completions or hard errors.

• –Streaming clients now need explicit refusal handling, not just post-hoc text parsing or status checks.
• –Anthropic says the refused turn must be removed or updated before continuing, which means this can poison chat state if you retry blindly.
• –The change also makes safety behavior more visible in developer UX, since refusals can arrive mid-stream with partial output already emitted.
• –Billing still applies up to the refusal point, so teams should monitor prompt patterns that trigger it frequently.
• –This looks like an incremental platform update, but it matters for anyone building agent loops or chat apps on Claude.