Generate and release open agent trajectory traces on Hugging Face

[gagb]

The repo provides the marketplace environment and agent implementations but does not publish any agent trajectory data. The infrastructure to capture traces already exists (database logging, extract_agent_llm_traces.py), but there is no pipeline to export complete, structured trajectories suitable for sharing.

We should generate trajectory datasets from simulation runs across different LLM providers and configurations, and publish them on Hugging Face Datasets. This would fill a gap — most agent benchmarks release environments but not traces.

Related: #156 (static input data release)

[gagb]

We have a few paths for publishing Magentic Marketplace trajectories to HF, with different effort/structure tradeoffs.

1. Lowest effort: upload the exported SQLite DB plus extracted markdown traces. This requires no new conversion code and preserves the full raw simulation state.
2. Best minimal structured dataset: export one row per LLM call from the logs table, since prompt is already stored in OpenAI-style chat format. This gives us HF-ready JSONL with almost no engineering.
3. Most complete but highest effort: stitch logs and actions into full tool-use trajectories (assistant -> tool call -> tool result -> assistant). This is the best long-term format for agent training, but it requires a real postprocessing pass.

Recommendation: start with option 2. It is the cleanest balance of usability, compatibility, and implementation cost.