bug#51101: 29.0.50; read-char-from-minibuffer accepts Enter even when not a choice.

[Juri Linkov <juri@linkov.net>]

close 51101 28.0.60
quit

>> Execute:
>>
>>  (read-char-from-minibuffer "Answer y or n" '(?y ?n))
>>
>> and press Enter. The form returns Enter (13) rather than re-prompting
>> for a y or n.
>
> It seems undocumented what RET is supposed to do in this function --
> I've added Juri to the CCs, perhaps he has some comments.

RET is supposed to do exactly the same what it did in
read-char-choice-with-read-key, or when these variables are non-nil:
read-char-choice-use-read-key and y-or-n-p-use-read-key,
i.e. to ignore RET and read the character again.

So this is fixed accordingly now in 28.0.

>> This causes a dataloss threat due to read-char-from-minibuffer being
>> called through ask-user-about-supersession-threat.
>>
>> Even if the read-char-from-minibuffer bug is quickly fixed, I'd
>> sleep better if the following patch were applied to userlock.el:
>>
>> @@ -194,7 +194,9 @@ ask-user-about-supersession-threat
>>  		       (list "File reverted" filename)))
>>  	      ((eq answer ?n)
>>  	       (signal 'file-supersession
>> -		       (list "File changed on disk" filename)))))
>> +		       (list "File changed on disk" filename)))
>> +              ((eq answer ?y))
>> +              (t (setq answer nil))))

I installed the patch provided by David as well.

> But I think ask-user-about-supersession-threat is working correctly here
> already (almost by chance).  RET means "yes" in functions like
> `y-or-n-p', which this is basically an extended version of, so it should
> work as "yes" here, too.

Maybe it works here because ask-user-about-supersession-threat
is called from C with some flag that disables signaling 'quit'.
But when trying to type RET after (y-or-n-p "Answer y or n: ")
it terminates with the 'quit' signal.  And indeed in the map
used by y-or-n-p, RET is bound to 'exit':

    (define-key query-replace-map "\r" 'exit)

I noticed the recent commit ec9f25bd356c7c81d94c78f11100b97d6d52ce97
saying that RET means "yes" in y-or-n-p.  But anyway since RET
now does the same that read-char-choice-with-read-key does,
so I removed mentions of RET from the doc string.  Or should
the fixed behavior be mentioned?