docs: fix factual errors in retool example page

[Shi-Dong]

Summary

Fixes factual errors in docs/examples/retool.md found during a fact-check of every page under docs/examples/ against the actual code in examples/retool/.

• Reward walkthrough: the page showed an invented reward function (parse_trace, r_ans / r_fmt / r_use weighted sum) that doesn't exist. Replaced with the real reward_func from generate_with_retool.py: math_dapo_compute_score(..., strict_box_verify=True) is the dominant correctness signal, and the only shaping is a small tool-call bonus that softens the penalty of incorrect samples (capped at −0.6) so tool use isn't extinguished early.
• Tuning table: replaced the row referencing the nonexistent reward weights with the actual knob — the tool-call bonus inside reward_func.
• What to watch: the doc listed fabricated retool/* and sandbox/timeout_rate metrics. Replaced with the real logged metrics (rollout/raw_reward, perf/rollout_time, train/loss, train/grad_norm) and keyed the timeout advice on the real sandbox error string (Error: Code execution timed out after ...).
• Troubleshooting: "EM = 0 forever" doesn't apply — this example scores \boxed{...} answers via math_dapo, not exact match. Row rewritten accordingly.

Test plan

• Verified against examples/retool/generate_with_retool.py and examples/retool/tool_sandbox.py on current main
• Docs render check

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.