v1: model Harbor artifacts as durable outputs, not only grading transport

[hallerite]

Context

#2144 adds artifact collection/restoration for isolated grading and imports Harbor's artifacts and [[verifier.collect]] task fields.

The current implementation uses artifacts only as an in-process transport channel:

• HarborTask.finalize() runs collect hooks and archives /logs/artifacts/ plus declared paths for every Harbor rollout.
• The tar bytes are stored in trace.state.artifacts, which is excluded from trace serialization.
• Isolated agentic-judge restores those bytes into the fresh judge sandbox.
• Single-agent and shared-judge runs still archive the files, but nothing consumes or persists the bytes.

Current code: Harbor collection, transient state, isolated restore.

Harbor semantics

Harbor describes artifacts as durable trial outputs: collected files are written under <trial_dir>/artifacts/, accompanied by a manifest.json, and exposed in the results viewer. With a separate verifier, those same artifacts are additionally restored into the verifier at their original source paths.

Harbor collection is best-effort and records failures in the manifest rather than failing the trial. Its destination field controls host-side placement without changing verifier-side restoration.

References:

• https://www.harborframework.com/docs/run-jobs/results-and-artifacts
• https://www.harborframework.com/docs/tasks#verifier-environment-shared-vs-separate

Problem

We currently conflate two responsibilities:

1. Result archival — preserving rollout outputs for inspection and downstream use.
2. Grading transport — moving selected files into an isolated verifier sandbox.

Consequences include:

• shared and single-agent Harbor rollouts pay archive/network/memory costs for bytes that are then unused;
• artifacts are not durably available in eval output or a viewer;
• the grading-transport 32 MB cap and strict failure behavior also define the imported Harbor artifact surface;
• destination has no effect because there is no host-side artifact directory;
• transient artifact bytes remain attached to live trace objects unless explicitly released.

Follow-up scope

Do not block #2144 on resolving the full artifact-storage model. Treat its transport-only behavior as an interim implementation and decide the durable artifact contract separately.

A follow-up design should consider:

• a host-side/eval-output artifact sink with a manifest and per-source status;
• streaming or downloading artifacts to disk rather than retaining every payload in Trace.state;
• durable collection for shared and single-agent runs without redundant grader upload;
• reusing collected outputs as the source for isolated-verifier restoration;
• separate policies for archival failures versus required grading-input failures;
• whether transport limits differ from archival limits;
• support for Harbor's destination, service provenance, collisions, and best-effort behavior;
• explicit cleanup of any transient transport buffers;
• lifecycle ordering: collect hooks, artifact capture, task scoring, and verifier execution.

Done when

• The durable artifact contract is documented independently of grading topology.
• Shared, single-agent, and isolated modes have intentional, tested behavior.
• Artifact outputs can be inspected after an eval without relying on live trace state.
• Isolated grading can restore required inputs without retaining unused payloads for every rollout.