Tests for the line-buffered reader added in 7b1ff38 were subtly
dependent on the execution environment due to differing end-of-line
markers on Windows and Unix-based systems.
Because StreamReader discards all newlines when reading line-by-line,
LineBufferedReader used a StringBuilder to patch the peeked lines
back together with the remaining contents of the file being read.
As StringBuilder.AppendLine uses the environment-specific newline
delimiter, the delimiters after the peeked-but-unconsumed lines can
therefore be substituted by the platform-specific variants, causing
the test failures due to the overly-simplified way they were written.
Reformulate the test to avoid such issues from resurfacing again
by splitting lines by \r or \n and then testing each line individually.
Additionally remove all raw literals in favour of explicitly mixing
various line delimiter character sequences for additional coverage.
LegacyDifficultyCalculatorBeatmapDecoder was registered as a fallback
decoder in commit ffde389 for future use in the server-side difficulty
calculation components. Due to the pre-existing fallback registrations
this causes a runtime crash when the diffcalc components are started.
Add a test reproducing this scenario to prevent the issue from
resurfacing in the future.
The "un-filter" step causes a `SelectNextRandom` invocation. If this happens to select a difficulty in set 3 other than the previously buffered difficulty #2, the subsequent test would fail.
I've split this test out to remove the random element, but added a new assert to ensure buffered (previously visited?) difficulty is re-selected on return to the same set.
Due to non-deterministic test failures in TestSceneBeatmapCarousel,
migrate the checkSelected helper step from AddAssert to AddUntilStep.
This adds more leniency for performance-related issues while still
checking the desired behaviour.
