Add rigorous local model benchmark evidence

[alan-d-smith]

Summary

• resolve memory/CPU metrics from the worker's actual nested cgroup
• record CPU quota, period, derived quota cores, and effective cpuset
• add repeatable --task-kind filtering and a balanced 40-case sentiment stress corpus
• document the evidence-based rejection of all tested Phi GGUF routes
• refresh both reproducibility manifests without consuming sealed holdout-v2

No local weights, llama.cpp runtime dependency, production route, or container change is included.

Verification

• 579 tests passed, 1 Windows symlink skip
• Ruff check and format check passed
• strict mypy passed across 72 source files
• Python 3.10 source compilation passed
• track1-v1 and track1-v2 manifests verify