Surface fork CTA and sandbox limits in Guide Step 3

[PunchTheDev]

Summary

Guide Step 3 (Write your agent) now front-loads the fork CTA and surfaces the sandbox-time constraints first-timers hit when their agent fails at eval.

Changes

• Fastest-path fork prose tightened around the ↗ code chip; new routed Browse the leaderboard → link (137-char tooltip) drives first-timers straight to a forkable winning agent
• "Or write from scratch" path leads with forge new your-name (CLI scaffolding) instead of manual mkdir
• generate(spec, llm) signature tooltip (274 chars) explains harness injection + inspect.signature rejection
• New Sandbox per problem: line with 3 tooltips: 180 s wall-clock (251) · 4 GB RAM (168) · network enabled for LLM calls only (236) — the eval-time constraints that aren't visible until your run times out or tries to fetch off-network
• Docstring fix: Sandbox: 60s, 4GB RAM, network enabled for LLM calls → Sandbox: 180s wall-clock, 4GB RAM, network enabled for LLM calls. The 60s claim was factually wrong — TIMEOUT_SECONDS = 180 in benchmark/sandbox.py.

Verification

Puppeteer at 1440×900 on /guide#write: routed /rankings href present, browse-link tooltip 137 chars, 3 sandbox tooltips render with lengths 251/168/236, forge new your-name chip present, 60s docstring claim gone, 180s wall-clock present. 0 new console errors.