Add case 05 Turf War (kind: agent) + agent-case contribution guide

One-shot JS policy agents battle head-to-head on a 40x40 grid; ranked by
objective wins (Bradley-Terry), refereed server-side. Inspired by Jiayi
Weng's Learning Beyond Gradients.

AGENTS.md

@@ -130,6 +130,27 @@ there is no per-project dependency setup:
   file** that the normal viewer serves verbatim. The anonymity rule above still
   applies: no model / vendor / harness identity anywhere in the code.
 
+## Agent cases — a single JS policy file
+
+Some cases (marked `kind: "agent"` in `cases.json`, currently `turf-war`) ask
+the model to write a game-playing agent instead of a page. For these your
+artifact is **one plain JavaScript file** at
+`outputs/<artifactDir>/<case-id>.js`:
+
+- Feed the case prompt to the model **verbatim, one-shot** — no test runs, no
+  iteration, no extra strategy hints of any kind. Extra context beyond the
+  canonical prompt invalidates the run.
+- Save the model's output **as-is**. The file must define `createAgent` at top
+  level (the interface is fully specified in the prompt).
+- The site runs the file inside an instrumented sandbox with a deterministic
+  compute budget; `Math.random`, `Date`, network and imports are blocked. CI
+  smoke-runs one match against a built-in baseline to check API compliance and
+  budget — it does not judge strength.
+- Battles are refereed server-side and ranked by real wins (Bradley-Terry), so
+  there is nothing to screenshot: the CI check is the proof for these cases.
+- The anonymity rule applies as usual: no model / vendor / harness identity in
+  the code or comments.
+
 ## Review guidelines
 
 NAGI BENCH stores raw benchmark artifacts. Reviewers, including Codex Review,