feat(eval): add isolated ROCm sanitizer plugins to main

[irvineoy]

Summary

This PR brings the isolated ROCm sanitizer/evaluation-tool implementation from #78 onto main.

PR #78 was merged into its stacked base, codex/quality-loop-agent, rather than main. Its parent PR #76 remains an open draft, so the sanitizer implementation never reached the default branch.

This branch starts from the current origin/main and cherry-picks only the six commits belonging to #78, with -x provenance. It intentionally excludes the stacked branch's unrelated curated-task and quality-loop commits.

Conflict resolution

The first cherry-pick was reconciled with current main as follows:

• kept both the current benchmark-method consistency fields and the sanitizer tool-gate fields in src/evaluator.py;
• retained the evaluation-tool socket/artifact isolation in src/scripts/docker_benchmark.sh while excluding quality-loop-only mounts and commands;
• excluded modifications to agents/quality_loop/*, because that agent is not present on main;
• added only /.eval-tool-artifacts/ to .gitignore;
• preserved current main runner cleanup while adding evaluation-tool temporary-path cleanup.

Validation

• git diff --check origin/main...HEAD
• Python compile checks for main.py, src/eval_tools, src/evaluator.py, and tests/eval_tools
• Bash syntax checks for the Docker runner and its shell test
• bash tests/test_docker_benchmark.sh
  • PASS: docker_benchmark runtime, agent-selection, and eval-tool isolation tests
• Pinned scoring image, focused suite:
  • 193 passed in 5.01s
• Pinned scoring image, full repository suite:
  • 459 passed, 4 skipped, 1 warning in 17.47s

Lineage

Original PR: #78

Original commits:

• 844761cc feat(eval): add isolated ROCm sanitizer plugins
• 25c3ddcd fix: harden evaluation tools fail-closed behavior
• e8d594a5 docs: document planned TSAN and UBSAN support
• 7ba42fe0 feat(eval): add Waitcheck and ConSan sidecars
• ea0d0812 fix(eval): harden sanitizer startup controls
• 6980af57 test(eval): require ConSan launcher provenance