, structurally, with no tag of its own to key off. */
.doc-root .table-wrap table tr:first-child > * {
opacity: 0.6;
font-size: 92%;
}
/* 3: rules, not zebra. A hairline under the header row, and one above any
Panel A/B/C row, both low-opacity, no fills.
Panel rows can't be matched on their text ("Panel A. ...") -- CSS has no
:contains() -- so this is matched on shape instead. Measured against the
built page (10 tables, 11 rows whose first cell reads "Panel..."): two
shapes account for 9 of the 11 -- (a) a lone | spanning the
whole row (table 6's Panel A/B), and (b) a label | followed by exactly
one empty | (tables 1/2/3/9's Panel A/B). (a) has zero false
positives anywhere in the document. (b) has 4, all inside table 1's
worst-converted stretch, rows structurally identical to a real Panel row
(bare label + one empty spanning cell) with no signal left for CSS to
tell them apart -- e.g. a lone "Number of observations" row picks up a
hairline it shouldn't. The 2 remaining real Panel rows are missed
entirely: docling fused the Panel label into the same cell as the next
row's data text (table 6 Panel C, table 9 Panel B). Not chased further --
the source is what's broken here, not this selector; full counts in the
task report. */
.doc-root .table-wrap table tr:first-child > * {
border-bottom: 1px solid;
}
.doc-root .table-wrap table tr:has(> th[colspan]:only-child) > *,
.doc-root .table-wrap table tr:has(> th:first-child:nth-last-child(2) + td:empty[colspan]) > * {
border-top: 1px solid;
}
/* 5: captions set apart from the data. Docling wraps every caption in
` ` inside ` `; before this they ran straight
into the header row with no more space than an ordinary cell gets. */
.doc-root .table-wrap table caption {
text-align: left;
padding-bottom: 0.9em;
}
.doc-root .table-wrap table .caption {
font-size: 90%;
opacity: 0.65;
}
/* 6: empty cells (32% of table 1's, measured -- matches the task brief's
own count) are conversion artifacts, not blanks the paper's authors
chose. Deliberately NO rule targets them: no border, no fill, no glyph.
Any of those would flag "something is missing here" on a cell whose row
neighbours are just as often real data -- and inserting a placeholder
character would be fabricating content over a cell nobody verified. The
only safe lever left is restraint: the rules above already don't zebra
or box individual cells, so an empty one reads as neutral whitespace,
not as a flagged gap. */
/* Section-level clip above closes almost all of it, but at a true 375px
viewport (measured via a same-origin iframe, since resizing a browser
window does not move window.innerWidth) 7px of sideways scroll remained:
documentElement.scrollWidth 382 against a 375px viewport. Isolated to
main itself, not .doc-root section, not MathML, not the menu button.
`main` is an ancestor of every sticky heading on the page, so `hidden`
here would be more destructive than at the section level — a scroll
container above every h2/h3 would silently kill sticky site-wide. `clip`
contains without creating one; verified sticky still pins at this level. */
main { overflow-x: clip; }
|