feat: quick experimentation harness for testing configs and prompts

Summary

There is no lightweight way to quickly test a config change, prompt edit, or model swap against a small dataset without running the full pipeline. This makes iteration slow and expensive.

Current state

The existing test suite (tests/engine/) contains 22 unit tests that mock LLM calls. There is no tool for running live end-to-end experiments with real models against a small slice of data and comparing results across runs.

Proposed harness

A CLI command or notebook (or both) that supports:

• Run a named config preset or inline config against N records from a dataset
• Diff output columns (leakage_mass, utility_score, judge, needs_human_review) across two runs side by side
• Print a summary table: avg leakage, avg utility, avg judge, # needing repair, # needing human review, # errors
• Optionally pin a model or override defaults without editing source files
• Works with any of the existing test datasets (TAB, BIO, RAT-Bench, Enron)

Why this matters

Right now testing a prompt change requires: editing code → reinstalling → running a full notebook → manually inspecting output. A harness would let Amy and others iterate on prompts and configs directly without needing to touch the library internals.

Start by reviewing the existing tests in tests/engine/ and the available TAB, BIO, RAT-Bench, and Enron datasets. Determine whether a CLI command, notebook, or both best fits the proposed workflow, then define how configs, model overrides, run comparisons, and summary statistics will be exposed. Done means a small dataset can be run against a config and results can be compared with the listed output columns and metrics.