feat(benchmark): add test-time scaling workflow

[elyasmnvidian]

uv run --no-sync python benchmark/test_time_scaling_config.py \
  --dataset benchmark/datasets/swebench-verified-closed-book \
  --task <task-id> \
  --model <model-id> \
  --output benchmark/tb_runs/test-time-scaling

cargo run -p switchyard-test-time-scaling-runner -- \
  benchmark/tb_runs/test-time-scaling/config/runner.json

What

This PR adds an executable version of the paper's two-iteration test-time scaling workflow.

switchyard-test-time-scaling owns the schedule, candidate records, voting, and saved results. The separate runner owns model requests and external commands. Harbor-specific Python adapters create fresh task copies, save patches, and run official verification only after candidate selection finishes.

The runner rejects unknown config fields, changed model IDs, invalid votes, reused output directories, and missing API keys.

Why

The repository has no command that can run the paper's schedule while keeping candidate selection independent from benchmark grades. This split keeps the reusable selection code separate from Harbor and model-provider details.

Existing benchmark commands keep their defaults. run-baseline.sh adds an optional custom-agent import and a pinned mini-swe-agent version for this workflow.

Current status

A small model-only smoke completed initial and refined candidates, summaries, votes, and run.json output. A real Harbor task and the full paper-size run are still pending, so this PR is a draft.

Notes for reviewers

Please check the boundary between selection and verification. Candidate selection must not read Harbor reward or verifier output.

[github-actions]

PR Preview Action v1.8.1
🚀 View preview at https://NVIDIA-NeMo.github.io/Switchyard/pr-preview/pr-478/
Built to branch gh-pages at 2026-08-18 23:34 UTC. Preview will be ready when the GitHub Pages deployment is complete.