Emit numeric per-task scores in inference tasks map

Output score-only task metrics in inference results

Clamp emitted per-step rewards to strict open interval

Use 1e-3 epsilon for all submission score clamps

Clamp inference output scores to strict open interval

Force proxy-visible LLM probe in inference

Probe the LLM proxy at startup

Use injected API_KEY for baseline LLM calls

align inference env vars with submission checklist

merge hfspace history and sync root runtime/docs with latest fixes

Sync full clinical-trial-triage project into Space