Gemini: scrub key from URL + throttle between calls

Add GeminiClient — needed for the real benchmark run

Fix Groq rate-limit handling — discovered during real eval run

test(eval): cover loader (join + subset + skip) and scorer integration

test(orchestrator): retry-off ablation mode fails on first error

test(orchestrator): cover format_observation and dataset_preview

feat(orchestrator): agent loop with sandbox execution and trace

feat(orchestrator): parse <code>/<answer> blocks from LLM output

feat(trace): JSON trace log of agent steps

feat(llm_client): exponential-backoff retry on 429/5xx

feat(llm_client): add Groq client with mocked-HTTP tests

test(sandbox): verify kernel state persists across cells

test(sandbox): verify matplotlib figures are captured as PNG

test(sandbox): verify timeout interrupts long-running cells

test(sandbox): cover exception traceback capture

feat(sandbox): execute code in a Jupyter kernel and capture stdout

chore: scaffold Python package and dev environment