fix(templates): head every slot with the section's own title

[DemchaAV]

Why

ModularCvTemplate promises that a section assembled at runtime reaches the page under the
heading its author wrote. Three of the presets that are about to declare it did the
opposite: Sidebar Portrait headed all six of its slots with a label chosen when the
preset was written, Monogram Sidebar its sidebar skills block, Mint Editorial its
profile and its skill bars. A CV whose section is titled "Employment History" reached the
page as EXPERIENCE; "Education & Certifications" as EDUCATION.

The rename is invisible in a CV whose headings happen to match the preset's vocabulary,
and wrong in every other one — and a CV assembled from data carries headings nobody in the
library chose, which is the reason its author wrote them.

What

Every slot that draws a section prints that section's title. Two blocks keep a label,
because neither draws the section it was handed:

• Sidebar Portrait's language list picks the language rows out of an "Additional
  Information" section. Three of that fixture's four rows reach no page at all, so the
  section's title there would head content the reader cannot see. The block takes the
  title whenever every row it was given was drawn — the normal case for a section routed
  there by its role rather than by a keyword.
• Mint Editorial draws one skills section as two adjacent blocks: an index of the group
  names, then a bar per skill. They cannot both carry one title, so the author's heads the
  index — the block a reader reaches first — and the section opens in their words even when
  the column flow puts the bars on the next page. The bars carry the preset's own "Skills",
  and nothing at all when the section is itself titled that.

Monogram Sidebar's four main slots already read the title, through a
title().isBlank() ? "…" : title() whose blank branch could not run: every CvSection
rejects a blank title in its constructor. Those branches are gone.

Wrapping. A longer heading also has to fit a narrower column. Letter-spacing inserts a
space between every pair of letters and wrapping breaks on whitespace, so
"EDUCATION & CERTIFICATIONS" came out as "EDUCATION & CERTI / FICATIONS".
TextOrnaments.spacedUpper now uses U+00A0 for the gap inside a word:
Character.isWhitespace is false for it, so a break falls between words wherever a
boundary exists, and a word wider than its column is still split. Sidebar Portrait's
character-for-character copy of that helper is deleted — the duplicate is what would have
let the fix reach one preset and not the two others this change gives long author-written
headings to.

Tests

SlotHeadingFidelityTest renders each preset exactly the sections its slots claim plus a
"Publications" section the tail is guaranteed to draw, and asserts every slot heading
appears above that marker. The leftover tail prints titles verbatim too, so without the
ordering a title found anywhere in the document would prove nothing about the slot.

Two existing tests changed for reasons worth naming:

• LeftoverSectionTest read placement off the heading — the slot's label against the
  tail's title — which is exactly the signal this change removes. It reads placement off
  the order now, against a section that can only be in the tail: that asserts the position
  directly rather than by proxy.
• The continuation-page control in SidebarPortraitContentFidelityTest pinned the unruled
  page to a literal read off one rendering (< 6.0), and the moved content made it 8.0. It
  measures the difference the page-margin rule makes instead, so it attributes the inset to
  the rule exactly and does not need loosening the next time the fixture shifts.

Every behavioural site was reverted one at a time to check the guard goes red — 14 of them,
including both directions of the language rule and the wrapping fix. All 14 red.

Measured, not assumed. The pixel gate stayed green throughout: the change moves 12,652
pixels on the Sidebar Portrait page against a 50,000 budget sized for cross-platform font
drift, so it would have shipped unnoticed. Setting the budget to 0 showed the three presets
had zero mismatch before this change while thirteen untouched ones differ by 69–47,959,
which is what makes the numbers above attributable. The U+00A0 gap on its own renders
0 changed pixels, and the cover letters — which share the ornament — report identical
numbers with and without it.

Five baselines re-recorded (sidebar_portrait ×2, monogram_sidebar, mint_editorial ×2);
the twelve visual.approve rewrote as collateral were reverted, which mattered here because
committing them would have baked this machine's drift in as the reference. Four committed
previews refreshed. Full reactor gate green.

composedText also stood verbatim in three test classes; it moves to CvComposedText
alongside the normaliser the assertions share, and the copies are gone.