A method-level data provider returning an empty array or an empty
Traversable made runTest() return without calling the test method, and
run() printed a passing mark. It now throws TestCaseException, the same
way as an empty file-level data provider fails the test.

Method-level annotations only; attributes take precedence over the docblock,
which stays fully functional. File-level .phpt annotations cannot be replaced
(no file-level attributes in PHP, runner reads them statically before spawn).

DressCode replaces Nette Coding Standard in CI, which now runs on PHP 8.4.
It is installed from dev-master, as it has no tagged release yet.
.gitattributes normalizes line endings of all text files (text=auto).

Changed messages:
- "This test forgets to execute an assertion." is now "No assertions were
executed in this test."
- a test method with arguments but without a data provider reports "no data
provider is configured" instead of "@dataProvider is missing"
- an empty data provider file reports the data file, not the test file
- DomQuery names the unsupported pseudo-class or combinator instead of
"Not implemented."
- Helpers::prepareTempDir() shows the given path instead of an empty one
- fixed wording of CLI option errors ("does not accept an argument",
"more than once") and a stray apostrophe
- Assert::contains() and notContains() keep the custom description when the
needle is not a string

Assert::throws() declares the same type of $code as exception(). Expect
lists all supported constraints. The readme example is runnable, lists all
assertions and matches the current output and options.

Dumper: replaced spl_object_hash() with spl_object_id(), the former is
deprecated. TestCase: ReflectionMethod::invoke() no longer receives an
object for a static test method. Assert::type() maps the 'integer' alias
to is_int(), because is_integer() is deprecated. FileMock.phpt: PHP 8.6
no longer shows the path in the warning about a failed stream_open().

The number of lines was the count of newlines plus one, and the code
view listed the numbers up to that count exclusively, so the last line of
a file not ending with a newline had no line number and coverage mark.
The number of lines is now counted correctly, which also fixes the total
shown in the report header.

The template declared a global function, so rendering a second report in
the same process failed on its redeclaration.

Names were inserted into the page as HTML, so a path containing HTML
special characters broke the report and could inject active content when
opening a report generated from a foreign source tree. Directories in
the tree view use generated element ids, as a path may contain
characters like # that break the links.

Not covered by a test, as the names are inserted by JavaScript of the
report template.

With -o tap, junit or none on stdout, the final "% covered" line was
still printed into the TAP or XML output; the error about an empty
coverage file goes to stderr in that case. When the requested report
cannot be generated, tester now exits with code 1.

Metrics of a file were the sum of its classes and traits only, so free
functions and top-level statements were missing although their <line>
elements were reported. Functions are now counted among methods, and the
statements of a loaded file are all its executable lines. For a never
loaded file the statements are still estimated from methods and
functions.

Methods of an enum were reported as free functions, and a function or
method declared as function &name() was not reported at all. Enums are
listed among classes, so they are included in Clover metrics.

Such a file has nothing left uncovered and is shown with 100 % coverage.

The guard checked only the literal form of the path, so e.g. a path
ending with '..' or a symbolic link pointing to the root passed.

Failed tests are reported as <failure> elements, but the summary counted
them in the errors attribute and failures was missing. The testsuite
element now has failures with their number and errors="0". Consumers
reading only the errors attribute will see 0 now.

A failure message comes from the output of the test process, and control
characters or invalid UTF-8 in it made the whole report unparsable.
They are now replaced by U+FFFD, in test names as well.

@multiple 0 or a non-numeric value ran the test twice, because the
value was cast to int and used as range(0, $count - 1). It now fails
with a message that a positive integer is expected.

css2xpath() converted [attr$="val"] to an expression that never matched
a longer value. An attribute value containing an apostrophe produced an
invalid XPath expression, and a quoted value lost all its surrounding
quotes instead of one pair.

On PHP < 8.4, find() and matches() with a selector list like 'a, b'
applied the axis only to the first alternative, so find() returned
elements from the whole document. The output format of css2xpath() for
selector lists is unchanged.

Appending remembered the end of the file at open time, so two handles
opened in 'a' mode overwrote each other's data. A mode like 'rb+' was
treated as read-only, because the parser did not reach the '+' after a
flag; like PHP, only the first character and the presence of '+' matter
now. touch() creates a missing file.

The detection of a UTF-8 mask was broken, so characters were always
matched as bytes and e.g. %c% could not match a multibyte character.
A mask containing non-ASCII characters is now matched as UTF-8 when the
actual value is valid UTF-8 as well; binary values are still matched as
bytes.

%f% no longer accepts invalid numbers like '.1.2'.

isMatching() left the raised pcre.backtrack_limit set when the regular
expression threw (e.g. a warning escalated to an exception).

run() without a previous setTempDirectory() call failed on an
uninitialized property, as did getDuration() before run(). Without a temp
directory, stderr was connected to a pipe that was closed immediately,
so the output was lost; it is now captured to a tmpfile().

Variables passed to the test process were removed from the runner's
environment after proc_open() instead of being restored, so a variable
the runner itself had (e.g. set by --setup) was lost. They are restored
even when the process fails to start.

With --stop-on-fail and parallel jobs, run() returned while the other
jobs were still running, after end() had been called on output
handlers, so they could write files or coverage data late or overlap
with the next run in watch mode. Now no new job is started after the
failure, but the running ones are finished and their results reported.

When an exception leaves the loop (Ctrl+C, a failing output handler),
running processes are terminated. Test discovery is inside the
try/finally block too, and end() is called on all output handlers even
if one of them throws.

With follow: true, curl returns header blocks of all responses in the
redirect chain (and interim responses like 100 Continue), and they were
merged into one array, so e.g. Location of the 302 response was visible
after following it. Now each status line resets the parsed headers.

A header sent multiple times (e.g. Set-Cookie) is combined into one
comma-separated value instead of keeping only the last one.

expectHeader() without an expected value and denyHeader() for a missing
header did not increment Assert::$counter, so a test consisting only of
such checks failed with "This test forgets to execute an assertion".

A float reachable through a PHP reference (array item, object property)
was converted to a string in the original variable, so a failed
assertion could change data used later by the test or its teardown.
INF, -INF and NAN are now dumped as valid PHP instead of INF.0 etc.

A message pattern '' or '0' was treated as "do not check", so any
message passed. Now only null skips the check. Tests that used '' to
disable the message check must pass null instead. The failure message
also shows an exception or error message '0' instead of omitting it.