DOMParser drops whitespace-only text nodes between inline elements

[Juanquintana-tech]

Summary

DOMParser (and therefore from_html) drops a text node that consists only of whitespace when it sits between two adjacent inline elements. Words end up concatenated.

Reproduction

from prosemirror.model import Schema
from prosemirror.model.from_dom import from_html

schema = Schema({
    "nodes": {
        "doc": {"content": "block+"},
        "paragraph": {"content": "inline*", "group": "block",
                      "parseDOM": [{"tag": "p"}], "toDOM": lambda n: ["p", 0]},
        "text": {"group": "inline"},
    },
    "marks": {
        "strong": {"parseDOM": [{"tag": "strong"}],
                   "toDOM": lambda m, inline: ["strong", 0]},
    },
})

doc = from_html(schema, "<p><strong>bold</strong> <strong>text</strong></p>")
print([n.get("text") for n in doc["content"][0]["content"]])

Observed (0.6.1): ['boldtext']

Expected: ['bold', ' ', 'text']

Scope

The defect is narrow, which is what makes it easy to miss:

Input	Result
<p>one <strong>two</strong> three four</p>	['one ', 'two', ' three four'] — correct, tails with content survive
<p>a<strong>b</strong>c</p>	['a', 'b', 'c'] — correct
<p><strong>a</strong> <strong>b</strong></p>	['ab'] — space lost
<p><strong>a </strong><strong>b</strong></p>	['a b'] — survives when the space is inside the mark

Not affected by ParseOptions(preserve_whitespace=...) in any of its three values (None, True, "full"), nor by using   or a literal U+00A0 (both are matched by str.strip()).

Root cause

DOMParser._wrap_lxml_text (prosemirror/model/from_dom.py, around L144) synthesises the <lxmltext> pseudo-elements, but guards on the truthiness of .strip():

if d.text is not None and d.text.strip() and str(d.tag).lower() != "lxmltext":
    ...
if d.tail is not None and d.tail.strip():
    ...

A whitespace-only text/tail is therefore never wrapped, so add_text_node never sees it.

That matters because add_text_node already implements the upstream decision about whitespace-only nodes, faithfully translated from prosemirror-model:

if (
    preserve_ws == "full"
    or top.inline_context(dom_)
    or re.search(r"[^ \t\r\n]", value) is not None
):

The third clause is the only one reachable for whitespace-only values, and it is false for them. The first two clauses exist precisely to preserve whitespace-only nodes: top.inline_context(dom_) is the branch that keeps a significant single space inside inline content (compare prosemirror-model 1.25.4, addTextNode, which has the same condition). The .strip() guard makes both unreachable.

Suggested fix

Guard on emptiness rather than on whitespace, and let add_text_node decide:

if d.text is not None and d.text != "" and str(d.tag).lower() != "lxmltext":
    ...
if d.tail is not None and d.tail != "":
    ...

Whitespace-only nodes in non-inline contexts still get dropped, because they fail all three clauses of the condition above, so the guard is redundant as well as harmful.

I tried this locally against a corpus of 15 documents exercising headings, inline marks, nested lists, blockquotes, code blocks, tables (including colspan/rowspan and per-column alignment), images with dimensions, and paragraph alignment. Round-trip fidelity (doc -> HTML -> doc) went from 14/15 to 15/15, with no other differences.

Happy to open a PR with the change plus a regression test if the approach looks right to you.

Environment: prosemirror-py 0.6.1, Python 3.14.7, lxml 6.1.1.