I wanted to make this a unit test but it turns out the
Sanitizer::safeEncodeAttribute() calls currently make this
impossible.
Bug: T353269
Change-Id: I5266e7b8b67db1c812dc9e4675d0c079ab1f9a40
This patch only moves existing code around without changing any
behavior. What I basically did was merging the old "guardedReferences"
method into "references", and then splitting the resulting code in
other ways. Now we see a few other concepts emerging. But the idea
something would be "guarded" (how?) is gone.
The most critical detail in this patch are the new method names, and
how the code is split. The names should tell a story, and the methods
should do exactly what the name says. Suggestions?
Bug: T353266
Change-Id: I8b7921ce24487e9657e4193ea6a2e3e7d7b0b1c3
This removes almost 200 lines from the main class.
This patch intentionally doesn't make any changes to the code but
only moves it around. Further improvements are for later patches.
Bug: T353269
Change-Id: Ic73f1b7458b3f7b7b89806a88a1111161e3cf094
> We lose useful coverage and spend valuable time keeping these tags
> accurate through refactors (or worse, forget to do so).
>
> I am not disabling the "only track coverage of specified subject"
> benefits, nor am I claiming coverage in in classes outside the
> subject under test.
>
> Tracking tiny per-method details wastes time in keeping tags
> in sync during refactors, and time to realize (and fix) when people
> inevitably don't keep them in sync, and time lost in finding
> uncovered code to write tests for only to realize it was already
> covered but "not yet claimed".
https://gerrit.wikimedia.org/r/q/owner:Krinkle+is:merged+message:%2522Widen%2522
Change-Id: Iafa241210b81ba1cbfee74e3920fb044c86d09fc
The main benefit is that the two lines that set and reset
$this->inReferencesGroup are now next to each other. More can be
done in later patches.
Bug: T353266
Change-Id: Ib3f40c40e0b1854f8e5a32af600f28931fffdb8c
This moves the actual parsing down to be done much later in the
process. This won't make any difference in production but makes it
easier to refactor the code further.
Note I tried to use a StatusValue object but couldn't because it
merges seemingly identical messages, while the plain array is fine
with containing duplicates. There is one parser test that covers
this. While we could change this it needs discussion and most
probably a PM decision.
Change-Id: I7390b688a33dace95753470a927bbe4de43ea03a
Two problems:
1. Manipulating globals directly affects all following tests. They
are not independent from each other. This problem can be seen in
CiteTest.
2. Some test cases in testValidateRef don't test what you think.
For example, the test for a conflicting "extends" + "follow" was not
failing because of the conflict but because "extends" was disabled
and disallowed.
Change-Id: Iaa4e1f3f3222155d59984e577cba3f0b8dec40c3
This patch makes only sense together with I5a64ac4 where it is split
from. See I5a64ac4 for details.
The idea is that this patch just re-arranges the code without making
any changes to how the code behaves. This leaves a minimal change
behind that's much easier to revert, if needed.
Bug: T298278
Change-Id: Ie78313b7f3ac1ec7bce5ac7512e60a3bb011480a
This patch does two things:
1. The "normalization" function was never only doing normalization,
but also all the necessary HTML encoding. This is now more visible
and split into two separate functions.
2. To make this easier we change the order slightly. Because of this
the normalization step must now consider spaces. Before spaces have
been converted to underscores by escapeIdForLink.
The results are all the exact same as before.
This is split from I5a64ac4 to make that easier to review.
Bug: T298278
Change-Id: I9435a2ddaa21559e29587c58b7523103141467f7
User-options related classes are being moved to
the MediaWiki\User\Options namespace in MediaWiki Core;
reflect that change here.
Bug: T352284
Depends-On: I42653491c19dde5de99e0661770e2c81df5d7e84
Change-Id: I22ff2effcf9b7f2162f5d57608d8ec3651b48dd7
This parser test is a bit obscure, in my opinion. We added it in
I8c4de96 to make sure we don't get thousand separators in most
places.
We continued reworking the code since then. By now it's effectively
impossible to "accidentally" get thousand separators. The
problematic methods from the Language class are not even accessible
any more from this code.
To make the tests more robust we now use createNoOpMock (done via
the previous patch) where it matters, specifically for all Language
and Parser mocks. This proves the problematic Language methods are
never called.
Bug: T253743
Bug: T238187
Change-Id: I9bfe1f4decfaf699996da63e19473c2c0d581d9d
Both Language and Parser are extremely complex classes with hundreds
of public methods. We really want to make sure we are not depending
on anything unexpected from these classes. If calls are made into
these classes we want to know exactly what is called.
Doing this also showed that some mocked methods are not even needed.
Change-Id: Icdfff6c07be78a47bf7cadb1813a72581a51272a
This reverts a very tiny part of Ib3fdc89 from 2 weeks ago. The
reasons are explained in Ib3fdc89. Short version:
* The ->parse() calls have drastic performance implications.
* Allowing wikitext and HTML in this message also makes T321217
worse.
The new message "cite_reference_backlink_symbol" is kept and still
used in the UI. Just not in these two messages any more. This is a
minor redundancy we want to get rid of at some point. But it's not
critical for the moment. This will be done as part of T321217.
Nothing will break on the wikis. Some wikis have customizations for
"cite_references_link_one" and "cite_references_link_many" in place.
This will continue to work as before Ib3fdc89.
Bug: T339973
Change-Id: I933771e3ad67cd530bcf5ee8469cef35ea1070d2
This is a mistake that exists in this codebase for who knows how
long.
Cite mis-uses the messaging system a lot for internal things we still
want to customize somehow, but are not labels that will ever be shown
on the screen. The prefix/suffix messages in this patch are meant to be
part of the HTML in id="…" attributes. Prefix/suffix must be a static
plain text strings. Using e.g. {{GENDER}} or {{PLURAL}} in these
messages is not even possible because there is no $1 parameter to use.
Note how all other similar messages already use ->plain().
A few wikis override these messages, but stick to the plain-text
convention, as they should:
https://global-search.toolforge.org/?q=.®ex=1&namespaces=8&title=Cite.*reference.*fix
This will continue to work.
This has minor performance implications. Fetching these messages is
faster if we can skip transformations.
Bug: T321217
Change-Id: I7969c255fe4ce897e904897081da5f52678721aa
The WikiEditor extension has a button and some help text that
is only applicable if the Cite extension is enabled. Move
that (with some modifications) to the Cite extension instead.
Bug: T339973
Depends-On: I8256660f9c6886d6764b45735284e00308fc56e5
Change-Id: Ib3fdc897dd3330f69c5832003d4c3cb1e6dba2f3
This is mostly because recent IDEs can understand createMock() quite
good. We usually don't add such hints every time we use createMock().
We would have a million of them. ;-)
Change-Id: If9e37807a6945c4408d374fc97664cd636020ffd
The best practice for message keys is to use dashes, not underscores.
This codebase is quite old and traditionally uses underscores. I
think we can make it flexible enough to work with both.
Required for Ie64f4ab.
Change-Id: I6f0584299a4f279ed929784927392eb0f72cbc80
Extensions using Phan need to be updated simultaneously with core due
to T308443.
Bug: T308718
Depends-On: Id08a220e1d6085e2b33f3f6c9d0e3935a4204659
Change-Id: Iebc5768a3125ce2b173e9b55fc3ea20616553824
If a ResourceLoaderFileModule is constructed with no arguments, it
accesses global variables, so this is not allowed from a unit test.
(This is probably a bug in ResourceLoaderFileModule, but one thing at a
time.)
This blocks If005958c76bbfabba74def4215c48fe94f297797.
Change-Id: I84056024b0d3a9dcddb1ab4dc8596118bb3fe8ea
Note how only the HTML5 mode behavior changes, but nothing in
legacy mode.
Also note this does not 100% fix the issue. The esample with a
non-breaking space is still broken. But it's already much better
than before.
Bug: T298278
Change-Id: Idf50dad4219ff4c594a0cc15f63cb10fdac5ffb7
This is only to document the current status quo and make later patches
smaller and easier to review.
Bug: T298278
Change-Id: I6c78f4d3ee32de596f2b5ee081d56eaffb1cc7bd
* Remove the overhead of serializing and then re-parsing client-side,
instead assign it directly as native object literal.
* Move code for array slicing to PHP.
Change-Id: Iedcc8d57d3bddd3fa32a78b4e7ecc25615d94277
Rather than depending on a separate module with one line of generated
JS code, generate it as a prepended statement to the same module.
Should be a no-op.
Depends-On: I809951d34feb2dbd01b7ae0f4bd98dac7c3f6fe2
Change-Id: I5886bf9f82025048976b7750e8cb751681021fb4
There is the patch(I4297aea3489bb66c98c664da2332584c27793bfa) which will
add DeprecationHelper trait to Parser class in order to deprecate public
Parser::mUser. DeprecationHelper trait has appropriate magic methods
which help to use dynamic properties. In order not to mock them via
createMock(), so getMockBuilder() and onlyMethods() was used.
onlyMethods() method helps to specify methods which need to be mocked.
Now we can use dynamic properties in Parser related tests of Cite
extension.
Bug: T285713
Change-Id: Ie75c9cd66d296ce7cf15432e2093817e18004443
The structure of this class changed a lot in If2fe5f5
(T239572). This was never reflected in the @covers tags. I
suggest to go with a trivial @covers for the entire class
and let PHPUnit figure it out. The alternative is to kind
of repeat many private implementation details, and this
doesn't feel right.
Change-Id: Ie414876489133ab9aca934c19a5e403cd339abf1
This is done to make the discussion in If3dcfd7 easier.
When we introduced this code we actually used it to format
entire numbers. We had to change this later to *not* localize
digits, but only separators. Language::formatNum is and always
was able to do this, so we just continued to use it.
This is discussed now.
It turns out there is only a single place left where we use
formatNum, and it does nothing but localizing the decimal
point. There is another way to do the same.
Bug: T237467
Change-Id: I89b17a9e11b3afc6c653ba7ccc6ff84c37863b66
We broke this feature in December 2019 because it was never covered
by any tests. Full explanation in T245376.
All the features we care about are covered by tests. If all existing
tests succeed, that should be proof enough that this patch does not
introduce any new regression.
Bug: T245376
Change-Id: I1a447884bdc507ac762d212466496b4591c18090
CI already ensures that VisualEditor is loaded alongside Cite, so
the defensive check in the code isn't needed; ext.cite.visualEditor is
defined statically, it's just injected into the page dynamically in the
VisualEditor code handling VisualEditorPluginModules.
Bug: T232875
Change-Id: Ie5e096feca92f9c3ef13c732f3f1ae491e2b7d03
This change does have two effects:
1. Instead of prepending a newline individually in every possible
code path, we do it one time at the end. But only if there is
something in the output. This does not change anything, as proven by
the unchanged parser tests.
2. I removed the newline between the <h2> and the generated
<references> element. Note that both these elements are created in
the same method, next to each other. So there is no way this can
influence other wikitext. Unfortunately this code path is executed
only when using the *preview* function, and impossible to be covered
by parser tests because of this. However, it's covered by unit tests.
This refactoring is motivated by, but not required for T148701.
Bug: T148701
Change-Id: I6691c70f8e3fa3f21e2d11035bed9cdc2dc87093
Previously the reflist was added at the end of the last line of text,
which messes up paragraph wrapping (as seen in many test cases), and
generated invalid HTML when the last line was a list item (T148701).
(second try, previously reverted in 8c933d03c5)
Note this affects only pages where the <references /> tag is missing,
and the references section is auto-generated at the very end of the page.
Bug: T148701
Change-Id: Ib2101346434a4e317b5fc7379215b60c7020cb2b
Previously the reflist was added at the end of the last line of text,
which messes up paragraph wrapping (as seen in many test cases), and
generated invalid HTML when the last line was a list item (T148701).
Bug: T148701
Change-Id: Ifc873fc913e717026d80d54b570c594d1073fb42
This removes a few tiny pieces of code, and a large chunk related to
incomplete follow="…" attributes (see T240858). It turns out we don't
need to insert elements at the top of the ReferenceStack::$refs
array, because this array is reordered anyway in
ReferencesFormatter::formatRefsList()!
Incomplete follow refs don't have a number, and are ordered to the top
because of this, as before. This doesn't change with this patch.
Change-Id: I43036420be22feb8f0f287d9ccee2afd317df2a9
The isModuleRegistered() method was introduced a few years ago,
when the load order in ResourceLoader was undergoing a change.
It used to be that hooks like were run first to register modules, and then
wgResourceModules was registered afterwards. This was reversed to disallow
mutating the config at run-time from foreign modules and to allow better
caching and error detection.
It's been several years since then, so this redundant check is no longer
needed. ServiceWiring.php in MW core for ResourceLoader always processes
config and extension.json first before this hook is called.
Bug: T247265
Change-Id: I466f1fa70b8f0e9fe5e8e8df90bb0001b3329b87
Html5 fragment mode now bans whitespace per html5 spec
See Ie2b7c9429691e2c491c3359d5b400d8f078aa789
Change-Id: Ie6fa40798f06a358f6082110b4d8cc0028c80321
This patch is mostly moving code around without changing the behavior.
Exceptions:
* The ErrorReporter creates a <span> container. This was previously
parsed. The only benefit might be error checking and escaping. Rather
pointless. The code just created this HTML. With this patch, it is not
parsed any more. The unit test reflects this change. The output in
production will not change, as the parser tests show.
* Parsing of the message key (to detect it's type and id) is simplified
a lot, using explode. With this the code can, in theory, support more
types.
Bug: T239572
Change-Id: If2fe5f55db46dfc7e0ce445348608bef00bec64e
Perform the validation in validateRef, and display a new error message for
broken "follow" refs. This changes existing behavior, where broken folow
ref content is arbitrarily displayed at the top of the references list and
no error is rendered.
Thanks to weasely wording, the new error can later be reused for "extends"
errors.
Bug: T240858
Change-Id: I506e4dcd1151671f5302ecd99581145d979d8124
This exception was introduced very late in the patch I38c9929. It
already caused trouble. This here is essentially a revert. It restores
the previous behavior where this edge-case was silently ignored. The
worst thing that can happen is that appendText() creates an incomplete
entry in the $this->refs array, which will be rendered at the end. The
user can see it then.
As of now we are not aware of a code path where this would even be
possible. Still this does make the code *more* robust by not making it
explode, but give the user something they can work with.
Bug: T243221
Change-Id: I2e2d29bbd557090981903fcc2ece8796fafa4aa4
This resolves another TODO. Since this is an intentional limitation in
the design of the feature, I find it pretty signigicant to give it it's
own error message.
Note that the text does not need to be perfect, just good enough for now.
We will review all error messages later via T238188.
Bug: T242141
Change-Id: Id9c863061e855350320131e81f6702c8810736f4
… if possible. In most cases it's possible to use the real object, and
reach into it's private parts via TestingAccessWrapper. This is almost
the same as using a mock, but I feel it's much more "light-weight".
The main change is that there is no strict assertion any more for the
number of ReferenceStack::pushInvalidRef() calls. Before this was mixed
into the same array as the valid references, as elements set to "false".
I think the test is as valueable as before without this extra check. If
the rollback stack works or not is already covered by other tests.
Change-Id: I90213557b164b3e43233a3dc393ee3f3d3d556a9
Allow a ref with `name=""` for backwards-compatibility.
Partially reverts I07738cce2641026dfaa92ba263ed6f9834be0944
Bug: T242437
Change-Id: Iaed2d1c41be377a4961aff39838b0965f6c00616
The difference between the two is that isOK() only reports "fatals",
while isGood() also reports "warnings" and "errors". I believe we
*want* to report all of these the same way.
Change-Id: I3be832c5db7aba3c03bd2ad8cfbba42362c093fd
A fun edge case where `name=""` fools both validation branches after
a references rollback, and triggered a LogicException. Stop these
freak refs.
Bug: T242437
Change-Id: I07738cce2641026dfaa92ba263ed6f9834be0944
* Always have an empty line between @param and @return to improve
readability as well as consistency within this codebase (before, both
styles have been used).
* Flip parameter order in validateRefInReferences() for consistency with
the rest of the code.
* In Cite::guardedRef() the Parser was now the 1st parameter. I changed
all related functions the same way to make the code less surprising.
* Same in CiteUnitTest. This is really just the @dataProvider. But I feel
it's still helpful to have the arguments in the same order everywhere, if
possible.
* Add a few strict type hints.
* It seems the preferred style for PHP7 return types is `… ) : string {`
with a space before the `:`. There is currently no PHPCS sniff for this.
However, I think this codebase should be consistent, one way or the other.
Change-Id: I91d232be727afd26ff20526ab4ef63aa5ba6bacf
The rollback feature was not able to properly restore a __placeholder__.
That's why a specific use case was behaving different. This already
worked just fine:
<ref extends="a">…</ref>
<references>
<ref name="a">…</ref>
</references>
But this didn't, even if it is the exact same from the users
perspective:
<ref extends="a">…</ref>
{{#tag:references|
<ref name="a">…</ref>
}}
Bug: T239810
Change-Id: I163a1bffb9450a9e7f776e32e66fb08d0452cdb9
One of the most significant changes is when I noticed that the $group
can never be null. We set it to DEFAULT_GROUP before. That's an empty
string.
I'm not very happy with the two @phan-suppress-next-line. Is there a
better way to fix these lines?
Change-Id: I33c1681e2f3857cb6701da71f4ed8893caff4d1e
I hope this is more readable. This patch does two things: It uses
array keys to name all elements in the data provider. (Note these
array keys don't actually do anything, PHPUnit ignores them.) And this
patch merges two parameters into a single $expectedResult.
Change-Id: Ib7adc32bf8bfd523735591d35d0bcabd3b853cfc
Since I3db5175 the ParserCloned hook handler does not rely on cloning
the Cite object any more. There is no cloning any more. This is dead
code and we could remove it. Just to be sure I propose to keep the
method, but let it throw an exception.
Bug: T240248
Change-Id: I2057ea652ca25f4c7031c28a6e713671738f5e22
These should be impossible conditions, we don't want to continue with
processing.
I hate this patch, it's a temporary workaround until someone rewrites
or replaces the rollback logic, for example with a two-pass parse.
Change-Id: I6a1327e397d4272fa412c3f290c2107d867d2854
I hope this patch is not to horrifying and can be reviewed. It's
possible to split this into a sequence of smaller patches. Please
tell me.
Change-Id: I4797fcd5612fcffb0df6c29ff575dd05f278bd4d
The main benefit is this nifty call: `$this->rollbackRef( ...$call )`
To make this possible, the minimal change I needed to do was to move
the two $argv and $text arguments to the end.
I also tried to order all other arguments as good as I could: Required
first, optional later. Group and name together. Name and extends
together.
All this is private implementation and should not affect anything.
Change-Id: I7af7636c465769aa53122eb40d964eabdd1289ba
I feel this is a little better than before. It looks like we never need
to *replace* a text that existed before.
This depends on I4a156aa which fixes one of the last remaining trimming
issues. Outside of <references>, a <ref> </ref> with no other content
but some whitespace was already forbidden. But not inside of <references>.
This is relevant for appendText(). It should not be called with null, but
was because of the inconsistent behavior.
Change-Id: I38c9929f2fa6e69482e45919e2f8dbf823cb1c8b
Note that this patch changes behavior, an invalid "dir" will result in
a cite reference at the point where the <ref> is declared rather than
in the references section. This is consistent with other errors.
Bug: T15673
Change-Id: Id10db40aa0b391f2f1d9274aa09d22a7278d65e3
The name of the base class in tests is guaranteed to only occur a
single time in a file. There is not much value in making it relative,
and requiring it to appear in the use section. Especially because it
is in the root namespace.
This reflects what I once encoded in the sniff
https://github.com/wmde/WikibaseCodeSniffer/blob/master/Wikibase/Sniffs/Namespaces/FullQualifiedClassNameSniff.php
I wish we could pick this rule and use it in our codebases. But it
seems it is to specific and can't be applied on all codebases, hence
it can't become part of the upstream MediaWiki rule set. At least not
at the moment.
Change-Id: I77c2490c565b7a468c5c944301fc684d20206ec4
* Use the Html class to safely create HTML code.
* $this->referenceStack can not be null any more.
* $this->inReferencesGroup is not needed during output, only when
parsing tags.
* Replace ReferencesStack::getGroupRefs() as well as deleteGroup()
with a combined popGroup() that does both things.
* Extract the code responsible for the "responsive" behavior to a
separate function.
* Some TestingAccessWrapper are not needed.
Change-Id: Ie1cf2533d7417ae2f6647664ff1145e37b814a39
Finishes breaking the circular reference between Cite and Parser.
This patch also demonstrates how evil it is to allow the error reporter
to be called from anywhere, and have side-effects. At least it's explicit
now.
Also fixes a bug where the inner error message would not be in the
interface language.
Bug: T240431
Change-Id: Ic3325cafb503e78295d72231ac6da5c121402def
This begins our journey of breaking the circular reference between
Cite and Parser. In later patches the child objects will also take
Parser as a parameter.
Bug: T240431
Change-Id: Ic672bb4bae19ac5f1e1f5817de171d76b3bd8786
Only create a Cite object if we need one. Never clearState, just
destroy and recreate later.
This makes it less likely that we leak state between parsers, and
saves memory and processing on pages without references.
It's also preparation to decouple Cite logic from state.
Change-Id: I3db517591f4131c23151c76c223af7419cc00ae9
* All classes are in a Cite\ namespace now. No need to repeat the word
"Cite" all over the place.
* The "key formatter" is more an ID or anchor formatter. The strings it
returns are all used in id="…" attributes, as well as in href="#…" links
to jump to these IDs.
* This patch also removes quite a bunch of callbacks from tests that
don't need to be callbacks.
* I'm also replacing all json_encode().
* To make the test code more readable, I shorten a bunch of variable
names to e.g. $msg. The fact they are mocks is still relevant, and still
visible because these variable names are only used in very short scopes.
Change-Id: I2bd7c731efd815bcdc5d33bccb0c8e280d55bd06
We are *so* close to 90%.
This patch should raise the coverage for the CiteDataModule to 100%.
I'm also adding a pure unit test for the clone() behavior. Note the
later is already covered by the CiteDbTest.
Question: Do we want the CiteDbTest to @cover anything?
Change-Id: I40763d01e18991f509bc30b6655aa57b23412fd9
Fixes a bug introduced in Icf61c9a27fd, which would cause a parser
cache split any time the Cite extension was initialized. The
`setLanguage` interface is regrettable, but I'm hoping it will only
be around temporarily.
Converts an integration test into a unit test and completes coverage.
Bug: T239988
Change-Id: I4b1f8909700845c9fa0cbc1a3de50ee7d42f69a5
Tickle very particular edge case in which a recursive parse corrupts
the $parser->extCite object.
Bug: T240248
Change-Id: I70d100e88fa72825194ed9c477b030bbf0b6b486
Because that is what it does. Note our method is different from the one
in the Language class. We only accept strings.
Change-Id: I39107e837cc29f2d7c8867c1e602aa643f9e1a57
This class renders a <references> tag and everything inside. The
previous name sounds like it is responsible for rendering the contents
of a <ref>…</ref> tag. I mean, the class contains a method that does
exactly this. But this method is private.
Change-Id: I1cd06c9a11e0a74104f2874a34efa3e0843a0f70
This adds a test for numbers like "1.2.0" that appear when an extended
reference (e.g. "1.2.") is reused multiple times.
The first separator is from the extended reference. We decided to never
localize it. However, the second seperator is from reusing a reference.
This was always localized. We believe this is a bug, but haven't fixed
it yet.
The test is documenting the status quo "1.2,0" with a comma. This kind
of makes sense, one could argue, because the "1.2" appears like this up
in the text, but the ",0" is a different indicator for a reuse, which
*never* occurs in the text.
Change-Id: Ie3d26bcadd8929b906bfbcac4806af2150d61f2a
This partly reverts Ied2e3f5. I haven't properly tested this before.
Rendering a bad extends (that extends a <ref> that's already extended)
not indented messes the order up and rips other extended <ref>s out of
context.
For now it might be better to stick to the previous, "magic" behavior:
Such an extends behaves like it is extending the *parent*, and is
ordered and indented as such. This is still not correct, but I feel
this is much better than rendering such a bad extends on the top level.
This patch also makes the code fail much earlier for a nested extends,
if this decision can be made already. In this case the error message is
rendered in the middle of the text (as other errors also are), not in
the <references> section.
Change-Id: I33c6a763cd6c11df09d10dfab73f955ed15e9d36
This partly reverts Id7a4036e64920acdeccb4dfcf6bef31d0e5657ab.
The message "cite_section_preview_references" says "Preview of references".
This line is not meant to be part of the content, but an interface message.
It should use the users (interface) language, not the content language.
Change-Id: I1b1b5106266606eb0dfaa31f4abd3cee9ba92e8c
We never access Language directly, so proxy its method instead of
returning the full object.
I believe I've found a bug, but not fixing here: the footnote body
numeric backlinks like "2.1" behave as if they were decimals rather
than two numbers stuck together with a dot. So they are localized
to "2,1".
Bug: T239725
Change-Id: If386bf96d48cb95c0a287a02bedfe984941efe30
