draft: add steering eval experiment on provisional MC

[williamli-15]

Summary

This PR adds a formal steering evaluation entrypoint on the current SynthPersona provisional multiple-choice benchmark.

Changes

• added experiments/01_steering_eval.py
• added src/persona_vectors/eval.py for MC answer-space scoring
• added qid metadata to activation artifacts and used it for safer steering-vector alignment when available
• updated steering docs and README to point to the formal experiment entrypoint

Result

The repo now supports a benchmark-style steering loop instead of only saving vectors:

• extract templated / biography activations
• compute biography - templated steering vectors
• evaluate bare, templated, biography, and steered on the same MC items
• report target-answer probability, accuracy, and flip statistics

Validation

• uv run python -m py_compile main.py src/persona_vectors/artifacts.py src/persona_vectors/extraction.py src/persona_vectors/steering.py src/persona_vectors/eval.py experiments/01_steering_eval.py tests/smoke_test.py
• uv run python tests/smoke_test.py
• uv run python experiments/01_steering_eval.py --help

Note

A remote dry-run was not completed from this environment because NDIF_API_KEY is not configured here.