theory-vs-reality

Compare a plan (stories / tasks / subtasks / acceptance criteria) against the actual built code. Produce an interactive HTML artifact a HUMAN can pick up cold and run, with per-item verdicts pre-filled by code-checking subagents.

The artifact is a test script, not a summary. Anyone who reads a row should know exactly which URL to open, which command to run, which file to upload, and what to look for on screen — without reading the plan or the code.

Two purposes:

1. Plan sanity — spot when the plan itself doesn't match what the user wanted.
2. Build sanity — spot misalignments between plan and implementation.

Stage 0 — Get the source

Accept any of:

• Direct paste — plan text in the conversation
• File paths — e.g. stories/*.md, subtasks.json, plan.md, tasks/*.json
• A retrieval recipe — e.g. "run aaa ralph subtasks list", "read all .md under docs/plans/"