We should discuss if we need really:

1. guesstimate centurys (read year 70 as 1970). This is problematic with old years and can imho abandoned.
2. Year BC detection. It's difficult to test it in all languages. I have used the wikidata object Q39085.
3. Variable yearBC is stored in mw.language.data because it is language specific. We can use later wgUserLanguage instead of wgPageContentLanguage? Used variable wgDefaultDateFormat and tableSorterCollation is stored in mw.config - why? (RFC)

I'm write now some tests for jquery.tablesorter.test.js.

I'm working on the date parser. I think we need a language setting like mw.language.months.names for BC era to detect years before year 0.

... oh sorry, i have tested with an older version offline. We can transform more than one letter ("aa" into "å") since Aug 2016.
But the tableSorterCollation shoud contain only lowercase letters. He work after

s = $.trim( s.toLowerCase() );
if ( ts.collationRegex ) { ...

In Norway is since Nov. 2017 a tableSorterCollation in Common.js, in Nynorsk (nn:), Islandic (is:) also. Danish (da:) and Faroese (fo:) Wikipedia have this not yet.
But the difficult task is to transform "aa" into "å". Today we can replace only one letter with one (or more) letter. Need we this really?
w:no:Spezial:All pages from "Saa.." sort 'aa' normal after 'a'.

This task is now resolved with Change 287449 jquery.tablesorter: Improve detection and handling of isoDate.
Because the parser isoDate build the date instead with new Date("...")
now with isodate.setUTCHours(h,m,s,ms).

I have tested the combination "sortable" and "mw-collapsible" and found only mistake if the header have more than one row and tfooter used.
If no caption written the script hides only the tbody, that is not enough.

Firefox needs a "T" in strings for isoDate:

new Date( "2010-02-06T19:12:00" ).toISOString();

is ok.
I have a solution for this - but it's today not ready for commit. See mw:Tablesorter/Improved parsers ...

This ins not a bug. Before sorting rowspanned cells must split into multiple cells with same content.
We can think about in this case - all rows are rowspanned - that after sorting the rowspan will restore to table.

In the past cssChildRows are read for auto detection of parser only for the first row. Now cssChildRows are scipped for auto detection and sorting.
After change I518029 i have tested the example in https://en.wikipedia.org/wiki/Help:Sorting#Keeping some rows together. But this example have only text in the first row and in this case the second row was not read.

A good patch, it works good for ltr- and rtl-direction. But why starts the classname with mw- ? I suggesst use of "autonumber" instesd "mw-autonumber". Every user read this HTML-code in the wiki-editor. It should be simple and short. For the counter i suggest use of "row":

table.autonumber tr td:first-child:before {

counter-increment: row;
content: counter(row);
}

Thee reason is, that bud0011 used | instead ! and so no <TH> generated. The function buildHeaders() use in headers the last row with the most columns.

I have worked on this problem and have a solution comming soon, after check some other problems with colspanning/rowspanning.