Fix Harbor reward details output

[neubig]

Summary

• Keep reward.txt as the Harbor-readable scalar result.
• Write pr_runtime diagnostics to Harbor's reward-details.json sidecar.

Test plan

• uv run pytest -q tests/test_pr_runtime_verifier.py tests/test_pipeline_pr_runtime.py
• uv run ruff check src/repo2rlenv/pipelines/_pr_runtime_verifier.py src/repo2rlenv/pipelines/pr_runtime.py tests/test_pr_runtime_verifier.py tests/test_pipeline_pr_runtime.py
• uv run ruff format --check src/repo2rlenv/pipelines/_pr_runtime_verifier.py src/repo2rlenv/pipelines/pr_runtime.py tests/test_pr_runtime_verifier.py tests/test_pipeline_pr_runtime.py

Out of scope

• Regenerating published Hub dataset revisions.

Closes #74

Live evidence

Run on 2026-07-13 UTC with the real standalone production _pr_runtime_verifier.py CLI in each checkout. Both runs used the identical parsed pytest log and F2P/P2P oracle files:

python src/repo2rlenv/pipelines/_pr_runtime_verifier.py \
  --log pytest.log \
  --f2p f2p.json \
  --p2p p2p.json \
  --runner pytest \
  --exit-code 0 \
  --out-dir "$OUT_DIR"

The input contained one passing FAIL_TO_PASS test and one passing PASS_TO_PASS test, so the production verifier computed a scalar reward of 1.0 and resolved=true.

Current main: Harbor-preferred file has an invalid schema

Commit: a8cf74d71a0f6de6bbc9adfb43a00249108610c1

The CLI wrote both reward.txt and reward.json. Although reward.txt contained the valid scalar 1.000000, Harbor gives reward.json precedence. The emitted JSON was neither a number nor a flat numeric map: it contained booleans (resolved, command_resolved), arrays (regressions, untracked_failed), and strings (parse_status, runner). Applying Harbor's numeric-or-flat-numeric-map contract produced:

{
  "harbor_schema_valid": false,
  "resolved": true,
  "command_resolved": true,
  "parse_status": "ok"
}

This demonstrates the issue directly: a correctly resolved verifier run still presents Harbor with a higher-precedence structured reward file that violates its reward schema.

PR head: Harbor consumes the scalar and details remain available

Commit: 056632c92d4d4d01a33ab2c337b21f8024297b94

The identical CLI invocation wrote reward.txt and reward-details.json, with no reward.json present:

files=reward.txt reward-details.json
reward.txt=1.000000
reward.json present=false
harbor_scalar_valid=true

The sidecar retained the structured evaluation data without competing for Harbor reward precedence:

{
  "reward": 1.0,
  "resolved": true,
  "command_resolved": true,
  "parse_status": "ok",
  "runner": "pytest"
}

This validates the changed scalar-plus-sidecar output contract through the actual standalone verifier process. Regenerating published Hub dataset revisions remains explicitly out of scope for this code-only PR.