feat: support macOS local benchmark runs

[waple0820]

What changed

• add a process-tree profiler for macOS with CPU lifetime accounting, RSS/Chrome RSS, host-memory guard, and throughput
• make the benchmark wrapper Bash 3 compatible and select the correct platform profiler
• run local comparisons with an explicit system Chrome executable and accurate machine labels
• continue evaluation/summarization when individual benchmark tasks fail, while still stopping on a resource guard
• use Lexmount pagination totals when verifying active-session cleanup

Why

The 5090 host was not available, so the GPT-5.5 LexBench comparison needed a reproducible macOS fallback without presenting RSS as PSS or controller metrics as server-side utilization. The local adapter fix is already merged in browseruse-agent-bench#100.

Validation

• 10 tests passed
• Ruff passed
• Bash syntax check passed
• real local LexBench task completed and judged successfully
• paired 8-task smoke and paired 20-task pilot completed on both backends
• c5 local pilot stayed above the 6 GiB host-memory guard

[github-actions]

Claude finished @waple0820's task in 2m 56s —— View job

---

PR Review

• Gather context
• Review changed files for bugs and security issues
• Post findings

---

Findings

1 bug found. No security issues.

---

Bug: len(response) is evaluated unconditionally — probe_lexmount_sessions.py:77

"page_size": int(getattr(pagination, "page_size", len(response))),

Python evaluates all arguments before calling a function, so len(response) is computed every time pagination is not None — even when page_size exists on the pagination object. If the real Lexmount SDK response doesn't implement __len__, this raises TypeError on every paginated list call.

The test masks this because FakeResponse defines __len__, but the real response object likely doesn't.

Fix: don't use len(response) as an eager default. Either fall back to a safe sentinel or compute it lazily:

"page_size": int(getattr(pagination, "page_size", 0)),

Fix this →

---

[github-actions]

Automated Claude review completed. Review run: https://github.com/lexmount/LexBrowserEnv/actions/runs/29203667689. Please check Claude's inline feedback and summary comment for details.