Unwrap the boundary-conditions note, and an editing pass

[lmoresi]

The prose is joined onto one line per paragraph, so a one-word change no longer
reflows a block and shows a dozen touched lines in the diff. 721 lines to 415.

Everything whose line breaks carry meaning is untouched: front matter, fenced
blocks and directives, display maths, tables, HTML, headings, and list
structure, with each item on its own line and continuations folded in.

Verified, not assumed. The markdown word stream was byte-identical across
the unwrap, as were the fenced blocks, table rows, headings, display-maths
blocks and links. The archival PDF was then built from both versions and the
rendered text compared: the same 6261 words in the same multiset, over the same
13 pages. Two accidental trailing-double-space hard breaks went, one of them
mid-sentence after a comma.

On top of that, an editing pass — lighter phrasing through the opening and the
classical-strategy paragraph — and four typos: a doubled "time", a capitalised
"Imposing" mid-sentence, and double spaces around two equation references.

Two claims corrected

The locking figure's caption said the node-normal panel "converges towards the
exact solution", which one figure at one resolution cannot show; before that it
said the panel was the exact solution, which overstates a numerical solve.
It now says it matches to plotting accuracy at this resolution. The left panel
is genuinely the analytic velocity substituted at the nodes —
truth.data[:] = _exact.evaluate("velocity", truth.coords) — so "exact
solution" is the right name for that one.

The finding about the facet normal is restored in plainer terms: the velocity
error does not approach zero under refinement, with a pointer to the
stress.py locking sweep that measures it across five cell sizes.

Checked

pixi run validate clean over 58 articles; the archival PDF builds;
pixi run check-style reports one hint (six not-X-but-Y constructions at 1.1
per 1000 words, down from nine before the editing pass).

Two things left deliberately, both the author's call: line 63 is the one
paragraph that leans on its heading ("### 1. A direct penalty" → "This
could not be more simple"), and the stress.py locking sweep also shows the
leak reading ~1e-5 while the answer is wrong — a sharper version of the note's
own argument, held back as paper material rather than note material.

Underworld development team with AI support from Claude Code

[github-actions]

Preview

• Boundary conditions on non-planar boundaries — PDF

Or the whole site.

Built from 1a8cbd0. Shows notes at draft and review, which the published site withholds. Not indexed, no comments, and not the citable version. Confirmed serving this commit before this was posted.