Use it to verify everything mechanical: authentication, order types accepted, rejection messages, field formats, reconnection behaviour, rate limits. These are the things that break on day one live and are cheap to find here.
Do not use it to estimate performance. Sandbox fill engines are typically generous, filling at the quote instantly and in any size, which makes any strategy look like it executes perfectly. That is a simulator artefact, not evidence.
The strongest version of this practice is a live account with minimum size running the real code path. It costs a little in commissions and gives you honest slippage numbers, which is precisely what incubation needs to measure.
Related: incubation, paper-trading, fill-assumption, live-vs-backtest-divergence