[codex] Refactor benchmark execution orchestration

[glennmatlin]

Summary

• Refactor the Thriller CLI path into reusable experiment preparation and run_config execution helpers.
• Add a paper-style benchmark runner with planning, resume checks, manifests, preflight API support, and result-quality failure detection.
• Update API/parsing behavior and optional adversarial dependencies so local tests and smoke execution do not require the full NLP stack unless those augmentations are used.
• Expand tests around runner planning, output safety, parsing behavior, prompt generation, and API response extraction.

Why

The previous run path was hard to orchestrate safely for full benchmark runs: it mixed CLI-only behavior, prompt snapshot writes, API key attachment, output directory creation, and execution. The new runner gives us a repeat-aware, manifest-backed workflow that can dry-run plans, resume completed runs, and fail fast on partial or malformed outputs.

Validation

• uv run ruff check src tests conftest.py scripts/run_benchmark.py
• uv run pytest tests/ -> 20 passed, 1 existing pydantic deprecation warning

Notes

Generated benchmark outputs and quarantined partial runs remain local and are not included in this PR.