fix(html2text): keep a pipe in a table cell from adding a column

[L4XB]

The bug

A GFM table row is split on every unescaped pipe, and html2text writes a cell's text through unchanged. A cell that holds a pipe therefore adds a column.

Measured on main (862f6bc):

input:   <table><tr><th>Name</th><th>Modes</th></tr>
                <tr><td>codec</td><td>a|b|c</td></tr></table>

output:  | Name | Modes |
         | --- | --- |
         | codec | a|b|c |

Read back, that last row has four cells against a two-column header. Everything after it is one column out of step, and a reader assigns b and c to columns that do not exist.

It is not an exotic cell value for a crawler. A pipe turns up in regex alternations, shell commands (grep -E 'a|b'), part numbers, "either/or" notes, and anything pasted out of a terminal — which is to say, in exactly the documentation and reference pages people crawl.

The fix

escape_table_cell_pipes escapes the pipes while the converter is inside a <td>/<th>, tracked with a single flag set in the existing table branch of handle_tag.

Two details worth calling out:

• The run of backslashes in front of the pipe is matched, not the pipe alone. escape_md_section leaves a backslash before a pipe untouched, because | is not in RE_SLASH_CHARS. A cell holding a\|b would otherwise come out as a\\|b, where the cell's own backslash escapes the one we added and the pipe splits the row anyway. Doubling the run first gives a\\\|b, which reads back as the literal a\|b.
• CustomHTML2Text calls it too. Its handle_data returns early for inside_pre and inside_code and writes straight to self.o, so a pipe inside a code span in a cell never reached the base escaping. A code span does not protect a pipe — GFM splits the row before it looks at backticks — so both early returns escape as well.

ignore_tables and bypass_tables return from the branch above, so the flag never gets set for them and their output is byte-identical. A pipe outside a table is untouched.

The table padder had to agree. pad_tables reformats the finished table by splitting each row on every "|", which turned the escaped pipe straight back into a column boundary:

| codec | a\    |b\ |c |

reformat_table now splits on the unescaped pipes, through one split_table_row helper used at all three of its split sites. PAD_TABLES defaults to False, so this only shows up for callers that turn it on — but leaving it would have traded one broken table for another.

Tests

tests/unit/test_html2text_table_cell_pipe.py, 18 cases — every one parameterised over both HTML2Text and CustomHTML2Text. The helper reads the produced table back the way a reader does (split on unescaped pipes, then resolve the CommonMark escapes) and asserts on the cells, not on a fixed string, so the test pins the property rather than the formatting.

Covered: a|b|c, a part number, a cell with its own a\|b, a header cell, a pipe inside inline code, the same table under pad_tables=True, plus guards for C:\path, a plain cell, a pipe in a paragraph outside any table, bypass_tables, and a plain table under pad_tables=True.

Measured:

	result
with the change	22 passed
source changes reverted to main, tests kept	12 failed, 10 passed

tests/unit + the markdown suites: 111 passed with the change against 89 passed without it, and the same 28 failed on both sides — those are pre-existing failures in test_sitemap_namespace_parsing.py and friends, unrelated to this change.

[Ryan-cw]

ok