Phase 0 eval spec for prompt ops

Draft Phase 0 evaluation spec for Factor/Compress/Compose experiments and create initial spec doc.

Goal

Define a minimal, concrete spec format and metrics so t-394/395/396 can be tested before theory work.

Deliverables

1. Spec document (Markdown) describing:

• Task suite format (YAML/JSON)
• Validator types (command, regex, json schema, file diff)
• Metrics (fidelity, trace drift, cost, latency, variance)
• Run protocol (baseline vs variants, repeat count)

2. Example spec file with 3-5 test cases.

Notes

• Prefer to align with existing eval harness conventions in Omni/Agent/Eval.
• Keep it minimal and executable via agentd or Op runner later.
• Use ASCII only.