Support external eval fixture files

Quickstart.md

@@ -265,6 +265,24 @@ cargo run --release -- --eval read-and-explain --model qwen2.5-coder:7b
 cargo run --release -- --eval fix-failing-test --json
 ```
 
+You can also point `--eval` at a data-only fixture JSON file. Workspaces are
+resolved relative to that fixture file and may not escape that root:
+
+```json
+{
+  "id": "external-readme-check",
+  "prompt": "Update README.md so it mentions the release version.",
+  "workspace": "workspace/readme-check",
+  "checks": [
+    { "type": "fileContains", "path": "README.md", "needle": "version" }
+  ]
+}
+```
+
+```bash
+cargo run --release -- --eval ./evals/local/external-readme-check.json --json
+```
+
 ## A Good First Session
 
 Here is a simple sequence that exercises the whole product:

README.md

@@ -837,7 +837,9 @@ runtime.
   `printf '…\n' | small-harness` for scripts and CI. Approval-gated tools
   are denied by default; pass `--allow-tools` to allow them.
 - **Agent eval CLI** — `small-harness --eval fix-failing-test [--model M] [--json]`
-  runs a bundled eval fixture and exits 0/1 (for CI scripts).
+  runs a bundled eval fixture and exits 0/1 (for CI scripts). `--eval` can
+  also point at a data-only fixture JSON file; its workspace is resolved
+  relative to that file and rejected if it escapes the fixture root.
 - **Warmup.** Small Harness sends a 1-token request with the full system
   prompt + tools at startup so llama.cpp-derived engines have a hot
   prompt-eval cache before your first prompt. Disable with `WARMUP=false`.