Skip code spans anywhere in a table cell, not just at its start

[youdie006]

tableRow runs its code-span check once, before the scan loop (parser/block_table.go:32-37), so it only fires when a cell starts with a backtick. tableFooter (:97-103) and tableHeader (:140-146) run the identical check, comment and all, inside their scan loops.

So a cell whose code span starts anywhere but the first character is split at the pipe inside it, and the last cell is dropped because the column count came from the header:

| h1 | h2 |
| --- | --- |
| x`|`y | z |

<td>x`</td>
<td>`y</td>     <!-- and z is gone -->

tableHeader only counts columns; it hands the header line to tableRow to split (:278), so header rows are affected the same way.

This moves the check into the loop. The i < n guard goes away because the loop condition already asserts it.

The behaviour was added in #217, whose body says "Check for a codespan in the table (header, row or footer) and simply skip to the end of it and resume scanning" - header and footer resume, the row never did. Every golden that PR added puts the code span at the start of a cell, which is why this survived. #195 is the original request.

Verification

go test -count=1 -run 'TestTable$' ., Go 1.26.3. md5sum parser/block_table.go on every row.

parser/block_table.go	md5	result
this PR	2389f3e8	pass
pristine HEAD	f2994f1a	3 failures, one per new golden
check moved back outside the loop	b8f6a10e	3 failures
i += isCode instead of i += isCode - 1	9a956f46	1 failure
drop the data[i] == ''` fast path	fe3a27d8	pass, equivalent
'`' to '~'	363b04be	7 failures

The i += isCode row only dies on the third golden, | x`|`| z |. With the first two the byte after the code span is ordinary text, so both forms stop at the same pipe; the two disagree only when that byte is the delimiter.

The fast-path row is a genuine equivalent mutant. For a non-backtick byte codeSpan sets nb = 0 (parser/inline.go:135), the delimiter loop never runs, and fBegin == fEnd == 0 returns 0 at parser/inline.go:180-181, so isCode > 0 is false either way. The guard saves a call per byte, and tableHeader/tableFooter both keep it.

Goldens go in testdata/Table.tests, which is where #217 put its own.

Gates as .github/workflows/go.yml runs them: go test . ./ast ./parser ./html all ok, go test -bench=BenchmarkReference -benchmem passes. Not run: CI pins Go 1.18.x and this box has 1.26.3 only.

Disclosure: this change was written with AI assistance (Claude). The measurements above were reproduced and checked by me before opening this PR.

[kjk]

Thanks!