If a matching element isn't the first child (e.g. preceded by a whitespace
text node), it may sit at a key other than 0 after some cleanup, leading to
a "Undefined array key 0" error.
This fix is now covered by 6 tests.
Fixes#96
Signed-off-by: Kevin Decherf <kevin@kdecherf.com>
Because of PHP 8.2 deprecation, in f14428e4c0, we stopped converting non-ASCII characters to HTML entities. Instead, we started to explicitly insert `meta[charset]` tag at the start of the document.
Later, we discovered that was breaking `html[lang]` so, in efbbc86df9, we made the insertion smarter. One of the improvements was that it would not insert the `meta[charset]` tag when it was already present.
That, however, broke websites that had `title` tag before `meta[charset]`. On those, libxml2 would decode the `title` contents as ISO-8859-1.
We could improve the logic (e.g. check that there is not text content before `meta[charset]`) or insert the tag unconditionally but it will probably be simplest to just go back to converting the non-ASCII characters to entities, just using non-deprecated function variant.
`DOMDocument::loadHTML` will parse HTML documents as ISO-8859-1 if there is no `meta[charset]` tag. This means that UTF-8-encoded HTML fragments such as those coming from JSON-LD `articleBody` field would be parsed with incorrect encoding.
In f14428e4c0, we tried to resolve it by putting `meta[charset]` tag at the start of the HTML fragment. Unfortunately, it turns out that causes parser to auto-insert a `html` element, losing the attributes of the original `html` tag.
Let’s try to insert the `meta[charset]` tag into the proper place in the HTML document.
We do not need to use the same trick with `JSLikeHTMLElement::__set`.
That expects smaller HTML fragments, not `html` documents, so creating `html` and `head` elements will not be a problem.
`parse_url($this->url, \PHP_URL_HOST)` will return `null` for local filesystem path.
Casting it to `string` will produce an empty regular expression,
which would match any link when computing link density.
`DOMNodeList` implements `Traversable`.
There are some `for` loops left but we cannot simply replace those:
PHP follows the DOM specification, which requires that `NodeList`
objects in the DOM are live. As a result, any operation that removes
a node list member node from its parent (such as `removeChild`,
`replaceChild` or `appendChild`) will cause the next node
in the iterator to be skipped.
We could work around that by converting those node lists to static arrays
using `iterator_to_array` but not sure if it is worth it.
It would fail for e.g. `<div> <p>foo</p> </div>`.
mozilla/readability uses children for the tag lookup, which return only elements.
PHP does not have children property so b580cf216d
mistakenly used `childNodes` instead, but that can return any node type.
Let’s filter the children ourselves.
Also add comments from mozilla/readability’s `_hasSingleTagInsideElement`.
Once we bump minimum PHP version, we will get newer PHP-CS-Fixer,
which will try to apply this cleanups.
Also manually tweak anonymous functions so that they are cleanly formatted
once we switch to `fn` syntax.
This is deprecated since PHP 8.2:
Deprecated: mb_convert_encoding(): Handling HTML entities via mbstring is deprecated; use htmlspecialchars, htmlentities, or mb_encode_numericentity/mb_decode_numericentity instead
It was used because `DOMDocument`, which uses libxml2 internally, will parse the HTML as ISO-8859-1, unless the document contains an XML encoding declaration or HTML meta tag setting character set.
Since first such element wins, putting the `meta[charset]` up front will ensure the parser uses the correct encoding, even if the document contains incorrect meta tag (e.g. when the document is converted to UTF-8 without also updating the metadata by the software passing it to Readability).
https://stackoverflow.com/a/39148511/160386
1) src/Readability.php (braces, no_unneeded_control_parentheses, single_line_comment_spacing, global_namespace_import, no_unused_imports, phpdoc_align)
2) src/JSLikeHTMLElement.php (phpdoc_separation)
Switch code blocks to Markdown syntax to work around `phpdoc_separation`, ApiGen uses Markdown these days anyway.
Readability was previously removing (was trying to actually, see next
section) invisible nodes using a pattern from `unlikelyCandidates`. This
was quite hacky and was removed during a backport of logics from
mozilla/readability. There is still a need to remove them so here we
are. We still use a pattern but specifically against the style
attribute. We also remove nodes with the attribute `hidden`.
The clean feature of tidy actually replaces inline style attributes
with css classes thus preventing readability to detect invisible nodes,
see https://github.com/htacg/tidy-html5/blob/5.6.0/src/clean.c#L1488
We therefore set clean configuration to false.
Signed-off-by: Kevin Decherf <kevin@kdecherf.com>
This change backports several things from mozilla/readability:
- Add child score to all ancestors instead of the first parent only
- Check 5 top candidates and try to find alternative candidates within
ancestors, this can help to find a better parent and grab more content
- Reduce patterns from `unlikelyCandidates` to the one used by Mozilla
as ours tend to remove useful nodes
- Score headers (h2 to h6) by default in addition to div, p, td and
section
Signed-off-by: Kevin Decherf <kevin@kdecherf.com>
> Method "Psr\Log\LoggerAwareInterface::setLogger()" might add "void" as a native return type declaration in the future. Do the same in implementation "Readability\Readability" now to avoid errors or add an explicit @return annotation to suppress this message.
HTML 4.01 Strict only allows block-level elements within noscript, form and
blockquote. The `enclose-block-text` option fixes the instances when those
elements contain inline elements or text by wrapping the children in paragraphs.
HTML 5 has looser content model and allows noscript elements basically anywhere,
including paragraphs, making the noscript elements inherit the parent element’s
content model. This means that tidy will produce invalid HTML nesting paragraphs
for `p > noscript > text`, a structure that would be invalid on two counts
in HTML 4 Strict profile but is completely valid in HTML 5.
Popular WordPress image lazy-loading code produces precisely that structure
so tidy “corrects” it to invalid code. In a proper HTML parser, the produced
code would force close the outer paragraph, making the noscript element
its sibling instead of a child. The only reason this does not break Graby’s code
for stripping the lazy-loading HTML is that libxml2 contains a bug
counteracting this:
https://gitlab.gnome.org/GNOME/libxml2/-/issues/205
Since all three elements allow flow content in HTML 5, it does not make much
sense to enable this option any more. The only possible issues that could occur
is producing HTML code not conforming to 4.01 Strict but that was never guaranteed,
as our example shows, and having blockquotes contain text nodes not wrapped
in paragraphs, which might be expected by some ancient stylesheets
but that is only minor and easily fixable visual backwards incompatibility.
`electrolinux/php-html5lib` was quite old and incompatible with the upcoming Composer 2.0.
Jumping to `masterminds/html5` for the same result. Also the lib is maintained.
Also:
- keep README in vendors
- use new Scrutinizer engine
- test with lower deps
- remove php-coveralls dev deps and download the phar during the CI build
A change released in tidy 5.6.0 breaks php-tidy when using
tidy_parse_string+tidy_clean_repair and wrap=0, incorrectly wrapping
every single word. Also it seems that $tidy->value should not be used to
retrieve the repaired html as far as it is undocumented and for internal
use.
We replace the call with tidy_repair_string which directly returns the
repaired string.
Relates to https://github.com/htacg/tidy-html5/issues/673
Relates to https://bugs.php.net/bug.php?id=75947
Tests pass.
Signed-off-by: Kevin Decherf <kevin@kdecherf.com>
Some contents have a `infocontent` node (ot sth different) and they are real content.
Using only `info` as regex is too agressive and remove legitimate content.
Matching the whole word `info` (or `infos`) should be a better choice
Kevin Decherf