docs: add AGENTS.md to document the S2 evaluation system and update latest.json with a valid baseline result.

Author: Alexzjt

evals/AGENTS.md

@@ -0,0 +1,48 @@
+# S2 Evaluation System - AGENTS.md
+
+This directory contains the automated evaluation pipeline for AntV S2. It is designed to benchmark and validate code generation capabilities of AI models (specifically Claude Code) against expert S2 patterns.
+
+---
+
+## 🤖 Agent Role: S2 Evaluation Orchestrator
+The goal of the agents in this directory is to execute, monitor, and visualize the performance of S2-specialized code generation. 
+
+- **Primary Goal**: Ensure that "Skill Injected" code (using S2 expert knowledge) outperforms "Baseline" code in correctness and aesthetic quality.
+- **Verification Method**: Every line of generated code MUST be rendered in a headless browser (Playwright) to confirm zero-runtime-error status.
+
+## 🛠 Tech Stack
+- **Engine**: Node.js + TypeScript (`ts-node`)
+- **Agent Link**: Claude Code CLI (`claude` command)
+- **Validation**: Playwright (Headless Chromium)
+- **Visualization**: React (Vite) + Ant Design 5 + VisionSnap SDK
+- **Data**: JSON-based expert test cases (`datasets/s2-expert-cases.json`)
+
+## ⚡ Key Commands
+Agents should use the following commands to interact with the system:
+
+| Command | Action | Description |
+|:---|:---|:---|
+| `npm run eval` | **Execution** | Starts the concurrent evaluation process. Spawns isolated Claude sessions. |
+| `npm run eval -- --limit 5` | **Dry Run** | Runs the evaluation for only the first 5 cases (for debugging). |
+| `npm run dashboard` | **Visualization** | Starts the local dev server to view result reports at `localhost:5173`. |
+
+## 🛡 Boundaries & Principles
+
+### Always
+- Use **Isolated Sessions**: When spawning `claude`, always include `--session-id` with a unique UUID and `--no-session-persistence` to prevent context pollution between test cases.
+- **Extract Markdown**: Use robust regex to extract code blocks between ` ```tsx ` markers from the model's stdout.
+- **Concurrent Execution**: Utilize the built-in `pMap` concurrency (dynamically calculated based on system memory) to avoid running 100+ cases serially.
+
+### Never
+- **Never rely on local file writes** inside the model evaluation prompt. Models must be instructed via `STRICT_SYSTEM_PROMPT` to only output code to `stdout`.
+- **Never use hardcoded sleep** for iframe rendering. Always use `waitForSelector` for reliable headless validation.
+
+### Ask First
+- If `os.totalmem()` detection fails or the machine has less than 8GB RAM, ask before running high-concurrency tasks.
+
+## 📂 Directory Structure Context
+- `run_eval.ts`: The main brain/scheduler. Orchestrates Claude, handles concurrency, and saves reports.
+- `validate.ts`: The headless judge. Boots the dashboard in `?mode=validate` to check if code rendering succeeds.
+- `datasets/`: The source of truth for benchmark prompts.
+- `results/`: The output warehouse of evaluation history.
+- `dashboard/`: The UI interface for human-in-the-loop review.