[ Bounty $4k ] [ Sandbox ] Summarize cleanup_all failures - bulk cleanup

[MyTH-zyxeon]

Summary

• make sandbox destroy report filesystem removal failures instead of swallowing them
• return the list of failed agent ids from cleanup_all while continuing through the rest of the registry
• add sandbox regression tests for partial and full cleanup success
• unblock full local pytest by exporting AgentStatus and using an RLock for timer observations

Root cause

destroy() removed the sandbox from the registry before attempting deletion and called shutil.rmtree(..., ignore_errors=True), so cleanup failures were hidden and cleanup_all() could not report failed agent ids. While validating the full suite, imports from src.agent also missed AgentStatus, and stop_timer() re-entered the collector lock through observe().

Validation

• ./.venv/bin/python -m pytest tests/test_agent_sandbox.py -q
• ./.venv/bin/python -m pytest tests/test_agent_sandbox.py tests/test_agent_registry.py tests/test_metrics.py -q
• ./.venv/bin/python -m pytest -q
• ./.venv/bin/python -m flake8 src/agent/sandbox.py src/agent/__init__.py src/common/metrics.py tests/test_agent_sandbox.py
• ./.venv/bin/python -m compileall -q src tests
• git diff --check

Bounty issue: #4066