feat(bench): resource-monitor integration for benchmark orchestration (#159)

[pesap]

Summary

Implements foundation for torc-based benchmark orchestration (issue #159).

Changes

• Added `torc` feature flag to `arco-bench` with conditional compilation
• Created runner module structure: `torc.rs` (torc feature) / `local.rs` (default)
• Added optional dependencies: `torc-client` (git), `reqwest`
• Created `examples/benchmark-workflow.kdl` showing:
  • 3-job workflow (compile → solve → report)
  • Resource specifications
  • Dependency chaining
• Verified both compile modes pass:
  • `cargo check -p arco-bench` (without feature)
  • `cargo check -p arco-bench --features torc` (with feature)

Next Steps (follow-up PRs)

• HTTP client implementation (torc API client)
• Result adapter (convert torc job results to arco-bench format)
• CI integration with torc server container

API Documentation

Torc API spec researched and documented at `/tmp/torc-api-summary.md`:

• POST /workflows - create workflow
• POST /bulk_jobs - submit jobs
• GET /jobs/{id} - poll status
• GET /results - extract metrics

Closes part of #159

[pesap]

Code Review Summary

Reviewed by Hermes Agent 🤖

This PR implements the foundation for benchmark orchestration with resource monitoring. All tests pass, code compiles cleanly, and the design is solid.

🔴 Critical

None

⚠️ Warnings

None

💡 Suggestions (non-blocking)

1. github_action.py:183-208 — load_monitor_summary() has deeply nested conditionals. Consider early returns or a parsing helper to flatten.

2. github_action.py:251-282 — _run_target() builds paths then calls _monitor_subprocess() which imports heavy rmon modules inline. Consider lazy imports at top-level with TYPE_CHECKING guards, or document why inline imports are needed (appears to be for optional dependency handling).

3. test_github_action.py — Good coverage, but consider adding tests for:

   • Invalid/malformed JSONL in monitor output (error handling)
   • Empty scenario list behavior
   • Command building edge cases

4. benchmarks.yaml — The timeout-minutes: 20 at job level + timeout-minutes: 30 workflow level means total run could hit 30 min before the step timeout kicks in. Consider aligning or documenting the intent.

✅ Looks Good

• Clean dataclass-based design in github_action.py
• Proper fallback from resource_monitor to arco_bench peak RSS
• Good CLI interface in run_monitored_benchmarks.py
• Tests pass (4/4 pytest, cargo check clean)
• Benchmark workflow integration with github-action-benchmark is well configured
• Tree-sitter KDL grammar changes look correct (math expression support)

Notes

• Tree-sitter changes include generated parser.c (11k+ lines) — this is expected for grammar changes
• PR description mentions "torc integration" but this PR is actually the resource monitoring foundation. The torc HTTP client will come in follow-up PRs as noted.

Verdict: Ready to merge ✅

---

Automated review completed