feat(web): stop repeating the comparison table, and give touch users the affordances the cursor was carrying - #337

…the affordances the cursor was carrying

Three of the top-ranked items from docs/SITE-IMPROVEMENTS.md.

THE COMPARISON TABLE WAS ON /features TOO, all eight rows, 9 of 9 character-
identical to the primary table on /comparison. The cost was authority rather
than bytes: a reader who studied it here and met it again there learned that the
page named for the comparison had nothing further to offer. Keeping two copies
honest also needed its own build check, which existed because they had already
drifted twice.

What replaces it is the one row that is the product's thesis, the three verdicts
for it, and a door to the page holding the evidence. The claim still appears
where a reader deciding between products will meet it; the argument now lives in
exactly one place.

THE LIGHTBOX DID NOT EXIST ON A PHONE. Its only affordance was `cursor:
zoom-in`, and a touch device has no pointer to change shape. So on the viewport
where these images are SMALLEST -- measured at 390px, each 2880x1800 capture
renders at 335x210, an 8.6x reduction -- nothing on screen suggested they could
be opened. The reader who most needed to expand the picture was the one never
told they could.

A glyph in the corner now says it, always present rather than hover-revealed,
because a hover-revealed hint has precisely the same problem as the cursor.

AND THE WIDE TABLES GAVE NO SIGN THEY SCROLL. macOS and iOS both hide overlay
scrollbars until a scroll is already underway, so the affordance appeared only
after performing the action it exists to suggest. A fade at the right edge says
it without instructions: content that dissolves rather than ending continues.

THE STICKY COLUMN HAD BEEN APPLIED TO ONE TABLE AND NOT THE OTHER. .cmp got it;
.patch -- the interactive routing matrix on the home page -- did not, and it is
the worse case: 403px of 736 hidden at 390px, 55%, on a table you scroll
sideways in order to CLICK a crosspoint. Losing the track label there means
choosing a cell without knowing which of six tracks it belongs to.

The background is a variable now because the two sit on different grounds -- the
comparison tables on the page, the matrix inside a surface-filled card. The
wrong opaque colour is as visible as no colour at all.

GUARDS, and the first version of two of them was wrong in a way worth keeping.
They searched only dist/_astro/*.css and matched the source spelling `::after`.
Astro inlines a component's scoped styles into the pages that use it, so
`.shot-open` is in features.html and in no CSS file; Lightning CSS then rewrites
`::after` to the CSS2 `:after`. The check reported a glyph missing that a
browser had already shown working. Both now read the HTML as well and accept
either spelling.

The sticky guard names BOTH classes rather than the concept, because naming only
the class that was fixed is exactly what let .patch stay broken.

Mutation-tested: dropping .patch from the rule, removing the fade, and removing
the glyph each fired their own named check, and the baseline is clean.

Also removes index-1440.png -- 314 KB of agent screenshot that a `git add -A`
swept into the previous merge.