Section 4.7.5 Formatting the Number says, under item 6:

"The mantissa is converted (if necessary) to an xs:decimal value, using an implementation of xs:decimal that imposes no limits on the totalDigits or fractionDigits facets. "

And, on xs:double/float specifically:

"the one that is chosen should be one with the smallest possible number of digits not counting leading or trailing zeroes (whether significant or insignificant). For example, 1.0 is preferred to 0.9999999999, and 100000000 is preferred to 100000001" 

I fail to see how 100000000 is preferred over 100000001 (both have the same amount of digits, the second having in part non-zero trailing digits). I also don't see how this is supposed to work. Consider:

format-number(1E26, '#')

This can return '10000000000000000905969664' because converting the double value 1E26 to xs:decimal is 10000000000000000905969664.

However, the text may imply that this should in fact return '10000000000000000000000000', but I don't see how that is supposed to happen.

The way I understand the text is that the example above should format the same way as format-number(xs:decimal(1E26), '#'), where there's no limit in the decimal digits. This seems in line with other parts of the spec where conversion from double to decimal retains the insignificant digits.

Relevant tests that show a different interpretation are numberformat60a, numberformat60m, numberformat60n, numberformat60o, numberformat60p, numberformat60q. Taking the latter:

format(1E100 div 3, '#####################')

I think this should (may, ought to, must?) return 
'3333333333333333224453896013722304246165110619355184909726539264904319486405759542029132894851563520', 

and not 
'3333333333333333000000000000000000000000000000000000000000000000000000000000000000000000000000000000'

But apart from the relevant tests, I raised this as a spec bug because, even if the text is considered correct, perhaps an example could be given for this particular entry.