Use Athena Judge instead of Gemini as the default playtester model

[p-forghani]

Problem\n\nRunWave currently uses Gemini as the default vision model for screenshot-to-action gameplay decisions. In repeated playtests, Gemini can spend much of the recording window on model latency while still struggling with games that require precise or multi-step control.\n\nFor example, six recent PR 1451 captures produced only 11-17 decision cycles in roughly 2-2.5 minutes. Average model response time was about 5.9-8.7 seconds, and the three golden-game runs repeatedly failed the opening obstacle despite the controller successfully applying every accepted action.\n\n## Requested change\n\nUse Athena Judge as RunWave's default agentic playtester model instead of Gemini.\n\nExplicit model overrides through the CLI, environment variables, and Python API should continue to work.\n\n## Validation\n\nCompare Athena Judge and the current Gemini default on a representative set of menu-driven, discrete-action, and real-time games. Record at least:\n\n- model latency and completed decision cycles\n- schema-valid response rate\n- successful menu entry\n- meaningful gameplay progress\n- cost per playtest\n