Always rounds up, because in the exact output case (increasing price) we need to move the price at least far enough to get the desired output amount, and in the exact input case (decreasing price) we need to move the price less in order to not send too much output. The most precise formula for this is liquidity * sqrtPX96 / (liquidity +- amount * sqrtPX96), if this is impossible because of overflow, we calculate liquidity / (liquidity / sqrtPX96 +- amount).
Always rounds down, because in the exact output case (decreasing price) we need to move the price at least far enough to get the desired output amount, and in the exact input case (increasing price) we need to move the price less in order to not send too much output. The formula we compute is within <1 wei of the lossless version: sqrtPX96 +- amount / liquidity