Add Harbor Dockerfile and multi-step support

[gabeorlanski]

Description

Adds Harbor Dockerfile and multi-step task support across the Harbor implementations.

Key changes:

• Adds shared Harbor parsing, tar, reward, and multi-step aggregation helpers.
• Supports Harbor multi-step task layouts in harbor-v1 and experimental HarborEnv.
• Supports Dockerfile-backed Harbor tasks with deterministic local image tags and configurable Dockerfile policy.
• Adds local Harbor dataset path support for harbor-v1.
• Records per-step Harbor rewards in rollout trace info.
• Adds focused unit and rollout-level Docker tests for Harbor scoring and multi-step behavior.

Type of Change

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Documentation update
• Test improvement

Testing

• All existing tests pass when running uv run pytest locally.
• New tests have been added to cover the changes

Ran:

uv run ruff check verifiers/harbor.py verifiers/v1/tasksets/harbor_v1 verifiers/envs/experimental/harbor_env/env.py tests/v1/test_harbor_taskset.py
uv run pytest tests/v1/test_harbor_taskset.py tests/test_harbor_env_mcp.py -q
uv run pytest tests/v1 -q -k 'not e2e'

Note: I did not run the full uv run pytest suite locally because the full v1 e2e matrix includes environment-dependent cases such as Modal.

Checklist

• My code follows the style guidelines of this project as outlined in AGENTS.md
• I have performed a self-review of my own code
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• Any dependent changes have been merged and published

Additional Notes

This keeps framework changes out of the PR and confines implementation to Harbor/shared-Harbor code paths plus tests/docs.

Note

Add Dockerfile build support and multi-step task scoring to Harbor taskset

• Adds DockerfilePolicy to taskset.py with build, ignore, and error modes; build constructs a deterministic local Docker image (tagged vf-harbor-*) from the task's Dockerfile context with content-hash caching.
• Extends HarborTask and parse_task to carry multi-step metadata (HarborStep list, multi_step_reward_strategy, workdir), with aggregate harness/scoring timeouts computed as sums (unbounded if any step has no timeout).
• Refactors HarborTaskset.solved and introduces run_verifier/record_step_results to evaluate steps sequentially, support early stopping on min_reward failure, and aggregate results using mean or final strategy into trace.info.
• Adds shared helpers in verifiers/harbor.py for config loading, reward parsing (scalar and JSON), tarball creation, step info serialization, and prompt synthesis from steps when instruction.md is absent.
• Mirrors multi-step support in HarborEnv for the verifiers-side evaluation path.
• Risk: default dockerfile_policy changes to build; existing callers using --taskset.ignore-dockerfile must migrate to --taskset.dockerfile-policy ignore.

^{Macroscope summarized fbc6288. (Automatic summaries will resume when PR exits draft mode or review begins).}