[EPIC] Eval Trace Fidelity #439

Outcome

Evaluation traces contain readable, complete context needed to explain benchmark behavior.

Scope

• Capture the system prompt used for a run.
• Render worker and coordinator prompts in a readable form.
• Retain enough trace context to reconstruct evaluation inputs and key decisions.
• Bound recorded content without silently destroying diagnostic value.

General streaming backpressure, client delivery, and runtime buffering remain in #208.

Complete when

• A supported evaluation failure can be investigated from its trace without relying on terminal-only output.
• Prompt content is readable and attributed to the correct run and agent.
• Trace-content limits are explicit and tested.