chore(evals): store eval cases as .ndjson so review tools skip them

evals/mcp-tool-use/README.md

@@ -14,7 +14,7 @@ pnpm mcp-tool-use --model claude-opus-5-5 --out runs/opus
 
 - `ANTHROPIC_API_KEY` in the repo root `.env` pays for Claude models and the Haiku shell simulator.
 - Codex models (`gpt-*` in `MODELS` in `run.ts`) go through a CLIProxyAPI server, set up as the `vibeproxy` model provider in `~/.codex/config.toml`. Keep the proxy under about 12 samples in flight across all runs.
-- Flags: `--case <ids>`, `--cases <file.jsonl>`, `--samples <n>` (default 2, and a 3rd when the first 2 disagree), `--concurrency <n>` (samples in flight, default 8), `--max-steps <n>`, `--effort <level>`, `--dry-run` (prints request bodies, no model calls), and `--context` (prints the size of T3's text and schemas, no key needed).
+- Flags: `--case <ids>`, `--cases <file.ndjson>`, `--samples <n>` (default 2, and a 3rd when the first 2 disagree), `--concurrency <n>` (samples in flight, default 8), `--max-steps <n>`, `--effort <level>`, `--dry-run` (prints request bodies, no model calls), and `--context` (prints the size of T3's text and schemas, no key needed).
 - `--baseline <results.jsonl> --stop-on-regression` stops as soon as a case that passed every baseline sample fails. Use it to test one copy change against a baseline run.
 - Cost: a full Sonnet run is about $9, and a full Opus run about $17. Codex models through the proxy cost only the shell simulator.
 

evals/mcp-tool-use/run.ts

@@ -1678,7 +1678,7 @@ const main = Effect.gen(function* () {
   }
   const fs = yield* FileSystem.FileSystem;
   const path = yield* Path.Path;
-  const readJsonl = <A, E>(file: string, decode: (line: string) => Effect.Effect<A, E>) =>
+  const readNdjson = <A, E>(file: string, decode: (line: string) => Effect.Effect<A, E>) =>
     fs.readFileString(file).pipe(
       Effect.map((content) => content.split("\n").filter((line) => line.trim() !== "")),
       Effect.flatMap((lines) => Effect.forEach(lines, (line) => decode(line))),
@@ -1687,10 +1687,10 @@ const main = Effect.gen(function* () {
   const caseFiles =
     flags.cases ??
     (yield* fs.readDirectory(casesDir))
-      .filter((name) => name.endsWith(".jsonl"))
+      .filter((name) => name.endsWith(".ndjson"))
       .toSorted()
       .map((name) => path.join(casesDir, name));
-  const cases = (yield* Effect.forEach(caseFiles, (file) => readJsonl(file, decodeCase))).flat();
+  const cases = (yield* Effect.forEach(caseFiles, (file) => readNdjson(file, decodeCase))).flat();
   const wanted = flags.case?.flatMap((ids) => ids.split(","));
   const codex = isCodex(flags.model);
   const harness = codex ? "codex" : "claude";
@@ -1764,7 +1764,7 @@ const main = Effect.gen(function* () {
     shellModel: flags["shell-model"],
   };
   const previous =
-    flags.baseline === undefined ? [] : yield* readJsonl(flags.baseline, decodeBaselineRow);
+    flags.baseline === undefined ? [] : yield* readNdjson(flags.baseline, decodeBaselineRow);
   const baseline = new Map(groups(previous, (row) => row.id).map(([id, r]) => [id, passRate(r)]));
   // Only a case that passed every baseline sample can stop a run; flaky ones are just reported.
   const passedBefore = (id: string) => baseline.get(id) === 1;