refactor(diagrams): apply the large-graph layout policy at the render point

[jimmykirk]

Follow-up to your review on #2 — you were right, and this is the version that does it once.

What changed

render_dot_to_svg is already the single place every graphviz object becomes
SVG. Its docstring says so, and it already carries one cross-cutting guarantee
(a render failure must never break the embedding page). The size policy is the
same kind of concern, so it now lives there and the seven per-builder calls are
gone. A new diagram type gets it without knowing it exists.

The exceptions had to become explicit

Deciding centrally means the cases that should not be re-laid out can no
longer be implied by which builders remembered to opt in:

• Already force-directed graphs are left alone — render_calls uses fdp,
  render_heatmap uses sfdp. Checked via the engine, not a name list.
• Ranks that carry meaning opt out via RANKED_BY_MEANING. Currently just
  the Hasse diagram, where the ranks are the partial order — force-directing
  it would destroy the thing it exists to show. This is the one case that
  genuinely needs a name, since "my ranks are semantic" is not observable from
  the graph.

It also fixes a real bug I shipped in #2

The merged version sets overlap="prism" — precisely the unsupported default
render_dot_to_svg's own docstring warns against:

never overlap="prism", which needs a triangulation library many distro
packages (Homebrew, several Linux distros) omit and reliably fails with
"remove_overlap: Graphviz not built with triangulation library"

So on a graphviz build without that library, every large diagram would have
fallen back to the placeholder SVG. It is now overlap="false", matching
render_calls. There is a test asserting prism is never requested.

Tests

Builder-level cases now exercise fit_layout_to_size directly rather than
asserting builders self-modify, plus new cases for both opt-outs and the prism
regression. FakeDot grows the attributes the render path now inspects, with a
comment explaining why.

26 passed against 23 on main for the same selection. The 17 errors in that
run are pre-existing in my environment (fixtures shelling out to cabal).

Note

CI on this will fail in Set up GHC like everything else at the moment — #9 has
the diagnosis, and it is unrelated to any of these changes.