Use benchmark Python for checkpoint evaluation

[tmusser]

Summary

• preserve the configured benchmark Python path, including virtual-environment paths;
• use the same interpreter for checkpoint verification and hidden evaluators;
• validate the evaluation environment once per phase;
• distinguish invalid evaluation environments from agent functional failures;
• preserve tri-state environment evidence for historical bundles;
• keep setup and usage bootstrap-safe when the project venv does not yet exist;
• surface interpreter and environment-validity metadata in traces and scorecards.

Validation

• focused observer and scorecard tests: 39 passed
• full benchmark harness suite: 264 passed
• shell syntax checks: passed
• git diff --check: passed

Evidence

The repaired Task 7 Codex pair produced complete checkpoint coverage and evaluator-backed first-functional and first-bench-ready provider-item boundaries without the prior missing-pytest failure.

Generated benchmark runs and reports remain local and are not included.