eval-spec-vs-build

# eval-spec-vs-build Compare a recorded run's app against the spec snapshot it was built from. **Spec fidelity only.** Course alignment is `/eval-materials-vs-build`. ## Steps 1. **Verify cwd is a course folder.** Confirm `spec.md`, `materials/`, `builds/`, `evals/` exist. Refuse otherwise: "Run /eval-spec-vs-build from inside a course folder." 2. **Require a run.** If `$ARGUMENTS` is empty: ```bash ls -1 builds/ 2>/dev/null | grep -E '^run-[0-9]+$' | sort ``` List those runs and ask the user: "Which run should I evaluate? Pick one — I won't default to the latest." Stop until they answer. 3. **Validate the run.** Confirm `evals/$ARGUMENTS/spec.md` (the snapshot) exists and `builds/$ARGUMENTS/` is a non-empty directory. Refuse otherwise. The report reflects whatever the build folder contains — evaluating an in-progress run is fine; just re-run after the build finishes for a clean score. 4. **Read the inputs.** - Snapshot spec: `evals/$ARGUMENTS/spec.md` (NOT the top-level `spec.md` — that may have been regenerated since this run). - Build: walk `builds/$ARGUMENTS/` (skip node_modules, .venv, dist, build, .git, and anything matching its own .gitignore). - Resolved decisions: `builds/$ARGUMENTS/resolved-decisions.md`, if present. Judge the build against the values it records, not the Ledger Defaults. If it is missing, say so in the Summary and judge against the Defaults. 5. **Produce the report at `evals/$ARGUMENTS/spec-vs-build.md`** with sections: - **Summary** — one-paragraph verdict. - **Coverage matrix** — each spec section → present / partial / missing, with file references into `builds/$ARGUMENTS/`. - **Divergences** — things the build did that the spec did not specify. - **Missing** — things the spec specified that the build did not do. - **Not applicable** — rules and criteria retired by the chosen options, per the snapshot's "Tests by option" table; never listed under Missing. Criteria that table says change are judged in their changed form. - **Adapted** — criteria the build reports as adapted. Judge whether each adaptation follows the chosen options; never score it against its Default form. - **Score** — your judgment on a 0–10 scale, with reasoning. 6. **Report the path written, then hand off.** After echoing the report path, tell the user: "Spec-fidelity report written. The companion eval is /eval-materials-vs-build $ARGUMENTS — different reference, different question (course alignment, not spec fidelity). Say 'go ahead' to run it next, or skip if you've already done it." ## Don't - Don't read `materials/` — that's the *other* eval. Mixing them defeats the point of having two. - Don't read the top-level `spec.md` — always the per-run snapshot. - Don't pick a default run. List the runs in `builds/` and ask the user which one.