First note that 3.8.3 grammar.html is stated to be the actual grammar used by the old parser, and is a bit different from the more human readable grammar given in the reference manual.  It is a bit different in 3.9 and I expect will be much more different in 3.10 with the new PEG parser. 

In the grammar, the CAPITALIZED_NAMES are token names returned by the tokenizer/lexer.  This is a standard convention.  

I am pretty sure that the human readable lexing rules in lexical_analysis are not what the lexer uses.  I presume the latter uses barely readable RE expressions, as does the tokenize module.

Compare the float grammar in https://docs.python.org/3/reference/lexical_analysis.html#floating-point-literals to the float REs in tokenize.py.

def group(*choices): return '(' + '|'.join(choices) + ')'
def maybe(*choices): return group(*choices) + '?'
# The above are reused for multiple REs.
Exponent = r'[eE][-+]?[0-9](?:_?[0-9])*'
Pointfloat = group(r'[0-9](?:_?[0-9])*\.(?:[0-9](?:_?[0-9])*)?',
                   r'\.[0-9](?:_?[0-9])*') + maybe(Exponent)
Expfloat = r'[0-9](?:_?[0-9])*' + Exponent
Floatnumber = group(Pointfloat, Expfloat)

Note that this is (python) code, not a text specification.  You or someone else can look at what the C lexer does.  But I think that the proposal should be rejected.