Thanks for the heads-up. I also found a contemporaneous English-language source,
the Brooklyn Daily Eagle Almanac (1912).

Unfortunately, we have no good way to model stopped clocks in tzdb, as tzdb
clocks are always running. So we need a transition of some sort, presumably
either one like 2020a (from 00:01:00 local time to 23:51:39 the previous day),
or the other, more-logical one (from 00:09:21 local time to 00:00:00 the same day).

Now that I'm looking into it, the line "0:09:21 - PMT 1911 Mar 11 0:01 # Paris
MT" that's currently in tzdb is surely a typo. It's not what's in Shanks, who
gives a transition time of "0:00". Formerly tzdb had no time for that
transition, which defaulted the time to 00:00. My patch dated 2001-03-13 changed
this to "0:01", but this part of the patch patch was in response to an email
dated 2000-12-20 (or -19) from Ciro Disceopolo
<https://mm.icann.org/pipermail/tz/2000-December/011284.html> which says nothing
about that "0:01". Evidently I mistakenly copied the "0:01" from the previous
line "Zone Europe/Paris 0:09:21 - LMT 1891 Mar 15 0:01", where Shanks does say
"0:01".

Anyway, the bottom line is that the stopped clocks make this an oddball
transition that cannot be modeled exactly by tzdb, and that given the story you
mentioned you are correct that it's better modeled using an ordinary transition
(from 00:09:21 to 00:00:00) than the unusual transition we're currently using. I
installed the attached proposed patch into the development database.

Given the above, we have a new trivia question: when and where could a stopped
clock have been correct *more* than three times in a single day? Answer: a clock
stopped at 00:00 was correct an infinite number of times that day in France.