How do you handle backtesting models when spread widening and slippage data are unavailable?

Historical chart replays often obscure live execution friction. How many pips of safety buffer do you add to your backtested stops to keep win-rate metrics realistic?