[codex] add isolated judge QA execution

[greenthree]

What changed

• add the lazy top-level judge_qa_problem(...) Core API and an isolated Judge QA runtime
• run Validator, Checker/Interactor fixtures, controlled contestants, and automatic robustness probes from a byte-consistent snapshot under the existing per-problem Judge lock
• enforce an overall deadline, default cancel-file and caller cancellation callbacks, live source/data/fixture hash fences, bounded structured results, and full process-tree cleanup
• atomically publish judge-qa-evidence-v1.json only after a complete successful run; failures, cancellation, timeout, input drift, lock contention, cleanup failure, or publication failure preserve the previous evidence
• share Validator policy and cancellation primitives with the formal Judge, and keep Interactor resource-failure attribution stable during simultaneous exits
• require official Validator/Checker/Interactor programs to remain reproducible C++ sources, matching the formal Judge and DOMjudge package path; controlled Interactor contestants may use C++ or Python

Why

The P0 schema can describe adversarial Checker and Interactor fixtures, but it could not execute them. This P1 slice provides the isolated execution and evidence layer needed to actively test each problem's special Judge without mutating formal PDF, ZIP, metadata, Manifest, source, or data artifacts.

This PR intentionally does not add the judge-qa CLI, seal integration, lint/status/report evidence state, Skill/reference updates, WebUI, or Legacy support. Those remain the P2/P3 work.

Validation

• targeted Python/unit/integration suite: 154 passed, 13 skipped
• npm run check: 643 passed, 30 skipped
• npm run pack:check: probhub=217, probhub-skill=5
• Skill quick validation passed
• git diff --check passed
• local phtest nine-problem lint passed
• local phtest L02 judge --no-cache passed: 33 inputs, 3 solutions, 99 case runs, all_expectations_met
• final same-model read-only review found no blocking issues