feat(harbor): upload standard verifier output

[David Leen (dleen)]

Summary

• ingest Harbor standard verifier output (test-stdout.txt, test-stderr.txt, and ctrf.json)
• show a bounded summary on the verification span and each score, with one shared verifier-output.json attachment
• support multi-step trials and scoreless verifier failures while respecting attachments=none
• apply the integration standard structured-key normalization and configured redact_patterns before upload
• enforce safe bounded reads, reject symlinks and non-regular files, and share the trial attachment budget

Complexity

• verifier output collector: 14 to 9
• bounded file reader: 13 to 10
• _sync_final_result: 56 to 48, matching upstream

Testing

• nox -s test_harbor(latest): 34 passed with Harbor 0.22.0
• nox -s test_harbor(0.16.0): 34 passed
• focused pre-commit hooks: passed
• pylint --errors-only on changed Python files: passed
• strict live backfill confirmed verifier_output_summary and verifier-output.json on the verification span

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 02ccbe3675

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "Codex (@codex) review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "Codex (@codex) address that feedback".

[chatgpt-codex-connector]

Include trajectory images in the attachment budget

When trajectory_mode="atif", import_trajectory() can add image Attachments from atif.py to the agent spans, but this counter starts with only selected_artifacts and is not updated after the trajectory import. For example, with both size limits set to 5 MB, a 4 MB trajectory image followed by a 2 MB verifier attachment is accepted even though the trial uploads 6 MB, so the newly shared max_total_attachment_bytes limit is not actually enforced. Include ATIF attachment sizes in the remaining budget before collecting verifier and reward-detail evidence.

Useful? React with 👍 / 👎.