Add Kernel-backed Harbor control arm with agent stop propagation and numeric verifier metrics

[rgarcia]

Summary

Upstream TIGER-AI-Lab/ClawBench now ships the native Kernel browser runtime (TIGER-AI-Lab#311), so this PR is reduced to what upstream does not have: the Harbor integration. The branch is rebased on the current upstream main; the two runtime commits that upstream accepted are dropped as duplicates.

What remains:

• add a Kernel-backed mode to clawbench-harbor-adapt (--browser-runtime kernel) that attaches ClawBench instrumentation to one pre-created Kernel browser per task, with headful stealth creation and no provider-default viewport override
• configure stock Harbor agents with pinned Playwright MCP (@playwright/mcp@0.0.79) through the credential-free CDP bridge; the Kernel API key never reaches the agent
• stop stock Claude Code and Codex cleanly when ClawBench records the final request, and persist stop timing
• intercept the Red Cross ZIP lookup request rather than blocking the chapter-finder landing page
• emit numeric verifier metrics (reward, intercepted, judge_match) and timestamped interception evidence
• finalize the replay, download recording.mp4, delete the browser idempotently (including failure paths via a setup trap) and verify deletion through the provider API
• docs: new Harbor control-arm section plus changelog entry; tests for adapter wiring, lifecycle scripts, stop wrapper, and provider session checks

Validation

• uv run pytest -q — 219 passed
• uv run ruff check src/clawbench tests — clean
• uv run ruff format --check src/clawbench tests — clean
• uv run --frozen pyright src/clawbench tests — 0 errors

End-to-end Harbor runs recorded during development of these commits (patches unchanged by the rebase): 3/3 tasks with intercepted=1, judge_match=1, reward=1 on stealth Kernel browsers; median agent-start-to-interception 40.8s; median stop propagation 47ms; all replays retained and all browser deletions verified.