[OsamaAnsar]

Problem

inscriptis._get_html_tree parses the input with lxml.html.fromstring. For fragments that consist of several root elements, or of top-level text next to elements, fromstring returns a synthetic <div> (if a block-level tag is present) or <span> as the root. The relaxed CSS profile styles those tags (div: padding_inline=2, span: prefix/suffix " "), so markup that is not in the input shows up in the text. In particular, words get split: <b>un</b>believable is converted to un believable.

Observed on master (a863dd7, 2.7.5, lxml 6.1.3 and 5.4.0, default/relaxed profile), compared with the same markup inside an explicit <html><body>...</body></html>:

input	before (get_text)	after	same markup in explicit <body>
<p>a</p>x	' a\n\n x'	'a\n\nx'	'a\n\nx'
<p>a</p><p>b</p>	' a\n\n b\n'	'a\n\nb\n'	'a\n\nb\n'
<div>a</div><div>b</div>	' a\n b'	' a\n b'	' a\n b'
<b>a</b>b	'a b'	'ab'	'ab'
<b>bold</b>text	'bold text'	'boldtext'	'boldtext'
<b>un</b>believable	'un believable'	'unbelievable'	'unbelievable'
<td>a</td><td>b</td>	' ab'	'ab'	'ab'

get_annotated_text('<b>un</b>believable', ParserConfig(annotation_rules={'b': ['bold']})) returned text 'un believable' before and returns 'unbelievable' with label (0, 2, 'bold') now.

Change

_get_html_tree uses lxml.html.document_fromstring instead of fromstring (also in the <pre> fallback). document_fromstring always returns the <html> root, so no synthetic wrapper is created; this is the same tree fromstring already returns for full documents.

Behaviour changes

• Output for fragment inputs with multiple roots or top-level text changes (see table above). This is the intended fix, but anyone relying on the old indentation/spacing for such inputs will see a difference. Only the relaxed (default) profile is visibly affected; I found no difference with the strict profile.
• Existing expectation changed: tests/test_broken_table_handling.py::test_forgotten_td_close_tag parses <body>hallo<table>...</table>echo</body>. fromstring turned that <body> into a <div>, which gave "hallo\n1 2\n3 4\n\necho". With the real <body> the expected text is now "hallo1 2\n3 4\n\necho". This is the same output that <html><body>hallo<table>... already produces on master, so I treated the old "hallo\n" as an artifact of the wrapper rather than intended table behaviour (the test is about the missing </td> handling). Please say if you would rather keep the line break and handle this differently.
• No other existing expectation changed. All other existing tests pass unmodified.

Not changed

• Single-element fragments (<p>a</p>, <div>a</div>, <span>a</span>, <li>a</li><li>b</li>), full documents, empty/whitespace-only/comment-only input: output identical before and after.
• Explicit <span> elements keep the relaxed profile's intended padding (e.g. fi<span>r</span>st is still fi r st, see Presentation of internal spans seems a bit odd #75).
• Fragments that start with bare text (Hello <b>world</b>!, plain text) already ended in a \n before this change and still do; libxml2 puts such text into an implied <p>, which is separate from the lxml wrapper.

Tests

New tests/test_fragment_wrapper.py (29 cases): exact expected text for the fragments above, equality with the same markup inside <html><body> for the relaxed and strict profiles, annotation offsets, and unchanged outputs.

With only the src/inscriptis/__init__.py change reverted, 15 tests fail (14 in the new file, plus the updated test_forgotten_td_close_tag); with the change restored, all 109 tests pass.

Verified

• pytest (full suite): 109 passed on Python 3.12 with lxml 6.1.3 and with lxml 5.4.0 (the lower bound in pyproject.toml).
• ruff check and ruff format --check: clean.
• Compared old and new output on all 44 tests/html/*.html files passed in raw, without a <body> wrapper (88 comparisons, text and annotations, both profiles): 16 relaxed-profile results differ (the leading indentation of the synthetic wrapper is gone); strict-profile output is identical.

Not tested

• tox: there is no tox.ini in the repository, so I ran what CI runs (ruff check) plus ruff format --check and pytest. I did not run uv build, ty, or safety.
• Python versions other than 3.12, and lxml versions other than 5.4.0 and 6.1.3.
• On Windows, two existing tests (test_html_snippets, test_html_snippets_annotations) fail on master unless PYTHONUTF8=1 is set, because they open() UTF-8 files without an encoding. Unrelated to this change; I ran with PYTHONUTF8=1.
• Direct use of Inscriptis(fromstring(...)) (as in the docs and examples) is not affected by this change and still has the old behaviour.

[AlbertWeichselbraun]

Thank you for spotting and fixing this. Your pull request has been merged and included in the Inscriptis 2.7.6 release.