Surface silent PDF rendering loss: gate dropped glyphs, report overflow, fix \setminus

[asluk]

Two changes to stop the PDF build from silently dropping content from the rendered output. The PDF opens fine — but individual glyphs go missing from the math with no error surfaced, so a downstream spec build shipped with missing symbols unnoticed.

1. PDF quality gate (#100)

doc_builder.py's stderr filter previously suppressed every Overfull \hbox warning and let Missing character warnings scroll past without failing the build. So right-margin overflow and silently-dropped glyphs were both invisible in CI.

New behavior in the stderr_processor + a new _report_pdf_diagnostics helper:

• Missing-character / unrepresentable-glyph → FAILS the build by default. A character silently vanishing from the PDF is meaning-changing, so it should hard-fail. --no-check-glyphs downgrades to a warning.
• Overfull \hbox (Xpt too wide) → collected and REPORTED, worst-first, non-fatal by default. --check-overflow makes them fatal; --overflow-threshold-pt sets the cutoff (default 1.0).
• Underfull boxes still ignored.

Sample output:

[doc_build] Right-margin overflow: 185 line(s) exceed the text width by >= 1.0pt (worst 172.3pt):
[doc_build] PDF diagnostics reported above (non-fatal).

# silently-dropped glyph -> build FAILS:
[doc_build] Missing glyphs (silently dropped from the PDF): 6 warning(s) [U+29F5 x6]
[doc_build] PDF quality gate failed (aousd/doc_build#100):
  - 6 missing-glyph warning(s) -- characters were silently dropped from the PDF

2. \setminus remap (template/after-header-includes.latex)

unicode-math maps \setminus → U+29F5 (reverse solidus operator), which no bundled math font provides — DejaVuMathTeXGyre and latinmodern-math both lack it, and tectonic's bundle ships only latinmodern-math.otf as an OTF math font. Result: \setminus renders as a blank/tofu box. Remapped to U+2216 (standard set-minus), which DejaVuMathTeXGyre has:

\renewcommand{\setminus}{\mathbin{\char"2216}}

Minimal repro (blank before, set-minus after):

$A \setminus B$

This is purely an engine/font-coverage gap, so it belongs here. (A separate, source-level issue — a text-mode \^ accent misused in math mode, which drops the accent — is fixed in the downstream spec source, not here.)

Validation

Built a downstream spec against this branch (local Windows + Linux CI, byte-identical PDFs): 0 missing-glyph warnings after the fix; set-minus renders. Validated in CI both directions: a clean build passes; a deliberately-reintroduced missing glyph fails the build (gate message + exit 1) with overflow staying non-fatal — so the build fails only on the dropped glyph, as designed. Before/after renders available on request.

Decisions for maintainers (@OleksiyPuzikov @dgovil)

This touches every spec's CI, so the defaults are a policy call:

1. Hard-fail on dropped glyphs for all specs — reasonable as the default? (My take: yes — a vanished character is never intended.)
2. Overflow threshold / default-non-fatal — currently report-only at >= 1.0pt. Honest but noisy (~185 lines), largely from RaggedRight + wide math. Keep non-fatal default, or pick a higher threshold?
3. Interacts with [ISO] Change from ragged-right margin to right-justified #11 (ragged-right → justified): switching to justified changes overflow behavior, so the threshold here should be coordinated with that work.
4. Re Debug the LaTeX handling of Windows paths in the font setup #23 (Windows font paths in font setup): hit exactly this during validation — absolute C:/ font paths break fontspec; the existing relative posix dejavufontpath workaround is what makes it build. Confirms Debug the LaTeX handling of Windows paths in the font setup #23's root cause; nothing here depends on changing it.