docs build does not catch nested inline RST markup (renders as literal backticks, sphinx exits 0) #30

Summary

Nested inline markup is invalid reStructuredText, renders as literal
backticks in the built HTML, and nothing in the docs build catches it.
sphinx-build exits 0 with no warning, so the only thing standing between this
class and a published page is whoever happens to look at the rendered output.

It occurred three times in a single session while writing #29, each time in
prose that read perfectly in the source:

**Supplying ``sample_col`` multiplies the cost by the gene count.**
**The ``dask`` path wins on every instrument**
**The ``dask``-less path is a different bargain ...**

All three render with the backticks visible:

<p><strong>The ``dask`` path wins on every instrument</strong>, and that is ...

Three instances of one shape is a missing check rather than three mistakes.

Why it is easy to write and hard to see

The rule — inline markup cannot nest — is not one most people hold while
writing, and the source reads exactly as intended. A literal inside a bold span
is the natural way to write "the dask path wins" when the emphasis
belongs to the whole clause. Docutils does not warn: it simply stops treating
the inner markers as markup.

It is also invisible to every check we already run. sphinx-build -b html is
clean, the page builds, links resolve, and the only symptom is two stray
backticks in a paragraph nobody re-reads after the first review.

How it was caught, and the suggestion

By scanning the built HTML for markup that should never survive rendering —
unrendered literals, directive lines, unresolved roles:

for page in glob.glob(f"{build}/*.html"):
    h = open(page, encoding="utf-8").read()
    leaks = (h.count("``") + h.count(".. warning::") + h.count(".. note::")
             + h.count(":doc:`") + h.count(":ref:`"))

Zero is the only acceptable count, and it found all three instances plus a
:ref: accidentally written inside a code-block comment, where roles are not
interpreted.

Worth making a docs-build check rather than an agent's habit — a few lines in
docs/ run after sphinx-build, failing on a non-zero count. It needs no
knowledge of the RST rules; it just asserts that nothing that is supposed to be
markup survived as text.

A source-side linter (doc8, rstcheck) is the other option, but the HTML scan
is strictly better here: it catches the code-block-comment case and anything
else that fails to render, without needing a rule per construct.

Scope

Documentation infrastructure only; no runtime effect. Filed separately from #29
rather than folded into it, since it is about the build rather than about that
change.