Each strategy takes as candidates the language outputted by the
previous strategy if any. This was already the case for the
Classifier and Heuristic strategies as these couldn't generate new
candidate languages (as opposed to the Modeline, Filename, Shebang,
and Extension strategies).
In practice, this signifies that if, for example, the Shebang
strategy finds two possible languages for a given file (as is
currently possible with the perl interpreter), the next strategy, the
Extension strategy, will use this information and further reduce the
set of possible language.
Currently, without this commit, the Extension strategy would discard
the results from the previous strategy and start anew, possibly
returning a different language from those returned by the Shebang
strategy.
* Add test to demonstrate Perl syntax detection bug
A Perl 5 .pm file containing the word `module` or `class`, even with
an explicit `use 5.*` statement, is recognized as Perl 6 code.
* Improve Perl 5 and Perl 6 disambiguation
The heuristics for Perl 5 and 6 `.pm` files disambiguation was done
searching for keywords which can appear in both languages (`class` and
`module`) in addition to the `use` statement check.
Due to Perl 6 being tested first, code containing those words would
always be interpreted as Perl 6.
Test order was thus reversed, testing for Perl 5 first. Since Perl 6
code would never contain a `use 5.*` statement, this does no harm to
Perl 6 detection while fixing the problem to Perl 5.
Fixes: #3637
