fix(templates): stop Sidebar Portrait's language block losing rows

[DemchaAV]

Why

Sidebar Portrait's language block accepts a wider "Additional Information" section and
picks the language rows out of it. It had a fallback for a section where nothing matched
— then it drew every row — so the loss needed a mixture: one row that looked like a
language suppressed the fallback, and every row that did not look like one reached no
page at all. Because the section counted as claimed it never reached the leftover tail
either, so three of the shipped example's four rows were simply gone: work eligibility,
open source, speaking.

A section routed here by SectionRole.LANGUAGES was no safer — English (C1) beside a
plain Deutsch B2 was enough to lose the Deutsch. That is the path the modular contract
advertises, which is what made this worth fixing before the branch goes anywhere.

What

The slot claims in two steps and knows which it got:

• Entirely languages — routed by the role, or titled for the job — is drawn whole,
  whatever its rows look like.
• A wider section is still picked over, and the rows the block does not draw go to the
  main column under the title their author wrote, walked in document order so the rest of
  a section keeps the place it had among the sections no slot claimed.

The key list is plural-only. Keys match as a substring of the stripped title, so a
singular "language" would claim "Natural Language Processing" and "Sign Language
Interpreting" for this block and draw them upper-cased in a third of a page's width —
"languages" still reaches "Language Skills", whose stripped title starts with it.

The category row is matched exactly rather than by containing the word, so a row labelled
"Sign Language" keeps its name instead of being expanded into its body; a row whose split
yields nothing falls through instead of being consumed by a branch that drew nothing; and
a level written without brackets is drawn as a level, because the block now hands the
renderer the pair already separated instead of one joined string it had to re-split on a
bracket it may not contain.

Tests

Placement is read off the identity name — the main column draws it, the sidebar does not —
because the words alone cannot say which column drew them. That boundary is asserted to be
drawn exactly once, since it is only a boundary while that holds.

All three language tests go red against the previous commit. That is worth stating
plainly: the first version of the role test used a fixture with no bracket anywhere, which
the old code's all-or-nothing fallback drew whole — so it passed against the bug it named
and guarded nothing. It now puts a bracket on one item and leaves the other two plain,
which is the mixture that actually lost rows.

New coverage: a wider section with no language in it at all (the whole thing goes to the
main column, and the sidebar draws no heading over an empty list), and drawn-once counts
so a row cannot satisfy the assertions by appearing in both columns.

Full reactor gate green, 936 tests. One baseline re-recorded
(sidebar_portrait-page-1 — the rescued rows land on page two; the sidebar is unchanged
and page one does not move), the seventeen visual.approve rewrote as collateral reverted,
and the committed preview refreshed. Measured at budget 0 after the review pass, no CV
preset moves a pixel beyond that one page.