feat: adversarial restraint eval (Inspect) under evals/

[Olawoyin007]

Summary

Adds an adversarial restraint eval under evals/empathysync_restraint/, built on Inspect. It red-teams empathySync's core promise - restraint on sensitive topics - and produces findings for a human to review. It never edits the pipeline, prompts, or corpus.

How it works, per sample:

1. Solver runs the adversarial prompt through the real empathySync pipeline, headless, with storage redirected to a throwaway dir so ./data is never read or written.
2. Scorer is a model-graded judge whose rubric rewards restraint (deferring, handing off, steering toward exit), not helpfulness. It is fail-closed: an unparseable judge verdict is NOANSWER, never a silent pass.
3. Preflight refuses to start unless the models exist and there is real RAM headroom - unified-memory over-commit reboots the target box, so it checks rather than assumes.

Model roles are chosen deliberately: the engine under test is the small model empathySync actually ships (test reality), and the judge is a stronger, different-family model (a model is a lenient grader of its own output).

inspect_ai/openai are an optional evals extra, not core deps - a normal install stays lean, and the tests importorskip the extra so CI stays green without it.

Changes

• evals/empathysync_restraint/ - the eval: config, dataset loader + frozen starter set, offline generator, pipeline solver, restraint scorer, preflight guard, task, runner, README.
• tests/test_restraint_eval.py - pure-logic + mocked tests for the risky bits (fail-closed judge parsing, scorer verdict mapping, dataset loader, preflight).
• docs/adversarial-eval-loop.md - rewritten from the pre-Inspect draft to match the built design; now the rationale companion to the module README.
• pyproject.toml - new optional evals extra.
• CHANGELOG.md, TESTING_CHECKLIST.md, MERGE_CHECKLIST.md - entries/rows for the new eval and the new evals/ top-level dir.

Not touched: docs/architecture.md. That doc maps the runtime safety pipeline inside src/; this harness exercises the pipeline from outside and is documented by its own README + design doc, so adding it to the component table would be noise.

Testing

• pytest tests/test_restraint_eval.py -q - 15 passed
• pytest tests/ -m "not conversation" - 1143 passed, 23 deselected, no new failures
• black --check --line-length=100 evals/ tests/test_restraint_eval.py - clean
• flake8 (project args) on evals/ and the test - clean
• Verified the eval tests skip without the evals extra (CI installs core only, so importorskip guards them)
• Live eval run - deliberately not run here; it loads the 65 GB judge and produces findings for human review, which is a you-in-the-loop action, not part of landing the code.