[mstaats-nvidia]

Implements one-agent execution from #185 using the portable references published by #173. helix_execute.py selects one prepared agent and case taskset, requests missing image builds, verifies registered contents, and saves its retry key and exact request before submission. Retrying the same receipt reuses the accepted run; an intentional new run requires a new receipt.

Execution works without the original local agent/task directories. Run receipts retain the actual resolved configuration and source hashes. The internal native collector preserves full artifacts and job specifications, checks complete matrix accounting, keeps failed trials, and publishes evidence atomically. #174 adds the public command that collects and verifies directly into the standard results format.

Stack: #185 → #173 → #162 → #174. Shared client settings control the deployment; execution settings control dispatcher, concurrency and sandbox resources. Agent behavior and task selection remain in the prepared experiment.

Validation: full CI passed on Python 3.12/3.13 (4,496 tests per version, 110 optional/environment skips). Live per-agent submissions worked without original upload directories; identical replay left receipts unchanged and server listing showed exactly one run per retry key. Pending collection returned exit 2 without a final output directory. Completed runs retained recursion-limit, agent-timeout and sandbox-start errors as unavailable observations. A serial Oracle retry reproduced sandbox startup failure before agent execution; this is not a clean performance qualification.

[asutermo]

I would link the Helix repo if not done already (either here or somewhere higher level). Maybe longer-term a link to compass/eval-author inside helix, etc