fix(agent): retry transient provider throttles

[JaimeCernuda]

Summary

• add bounded transient-provider backoff around planner and answer-synthesis calls
• retry ALCF/LiteLLM rate-limit style failures without bypassing DSPy/LiteLLM
• preserve existing structured routing/provider errors when retries are exhausted

Evidence

• targeted live ALCF/Metis gpt-oss-120b memory follow-up returned error_info=null
• follow-up used six Parquet tool calls and six analysis direct-tool handoff events
• documents the broader hierarchical benchmark target in TASK.md for post-compaction continuity

Verification

• uv run pytest tests/test_core/test_agent_planner.py -q -> 138 passed
• uv run ruff check src/ tests/ scripts/run_demo_benchmark.py scripts/create_benchmark_data.py scripts/create_demo_data.py
• uv run pytest tests/ -> 1158 passed, 37 skipped