test(mvp): gate Collector correctness, performance, and resources E2E

[zzylol]

Summary

• sample real Collector self-telemetry counters across each paired four-node load window
• fail closed on missing/restarted counter evidence, refused points, or exporter failures
• gate sustained accepted-point throughput, points per CPU-second, per-agent CPU, and peak RSS
• include Collector runtime evidence in the machine-readable result and MVP report
• add evaluator and Prometheus parser regression tests

Acceptance contract

• both agents sustain at least 90% of the declared base workload
• at least 25k accepted metric points per CPU-second
• zero refused and zero failed-export metric points
• at most 4 CPU cores and 4 GiB peak RSS per agent

Validation

• python3 -m unittest discover -s deploy/mvp-multinode/scripts/tests -v (20 tests)
• python3 -m py_compile deploy/mvp-multinode/scripts/*.py
• bash -n deploy/mvp-multinode/scripts/run_demo.sh
• git diff --check

The physical four-node run is intentionally left for review/cluster execution; missing runtime evidence cannot pass the evaluator.