feat: migrate agent to GPT-5.6 Luna

[vrtornisiello]

Migrate agent to GPT-5.6 Luna

Swaps the agent's LLM from Gemini 3 Flash to OpenAI GPT-5.6 Luna, and adjusts the surrounding config, prompts, and schemas to fit the new model.

Model & provider

• Replace langchain-google-genai with langchain-openai; instantiate via ChatOpenAI instead of init_chat_model.
• New settings: OPENAI_API_KEY, MODEL_URI=gpt-5.6-luna, and REASONING_EFFORT (none…max) replacing MODEL_TEMPERATURE + THINKING_LEVEL. Reasoning is configured via reasoning={effort, summary: "auto"}. .env.example updated.

Context handling

• Retune SummarizationMiddleware for the larger context window: token-based thresholds (trigger at 500k, keep last 100k) instead of fractions, and summarize from the full discarded history (trim_tokens_to_summarize=None).

Prompts & tools

• Rewrite/condense the system prompt and the tool docstrings (search_datasets, get_dataset_details, get_table_details, BigQuery tools) for the new model.
• Drop the {language_directive} injection from the prompt middleware and remove language_directive from app/i18n (prompt now only fills {current_date}).

Structured response schema

• Remove TemporalCoverage / TemporalGranularity and the temporal_coverage field.
• Rename follow_up_questions → follow_up_prompts, reframed as next messages in the user's own voice.

Tests

• Update middleware, agent_runner, and i18n tests to match the removed/renamed symbols.