This pull request resolves #2258 by adding a normalized view for RunResult.to_input_list() that exposes the same continuation input the runner uses after nested handoff history or handoff input filters rewrite model history.

• Keep mode="preserve_all" as the default so existing callers still get the full converted history and current to_input_list() behavior remains compatible.
• Add mode="normalized" as an opt-in path that switches to _model_input_items only when the runner recorded a real divergence between model-facing continuation and richer session history.
• Propagate that normalized replay signal across sync, interruption, and streaming result builders so follow-up runs behave the same regardless of execution path.
• Cover the real failure modes behind .to_input_list() provides unparsable data if handoff happens in that run w/ nest_handoff_history: true #2258: nested handoffs, custom handoff filters, and cases where the delegated agent adds its own tool calls that must be preserved in follow-up input.