fix(pdf): word spacing, page rotation and glyph names

[andiwand]

🤖 Generated with Claude Code

Four rendering fixes, found by reading six new pdf test files into the private corpus.

Word spacing (Tw)

html/pdf_file.cpp emitted CSS word-spacing only on the substitute-font
path. A run painted with an embedded font carries PUA-re-encoded glyphs, whose
space is not a real U+0020, so the property was skipped outright — and every
justified line rendered ~9% short, the shortfall pooling into gaps mid-word:

Lithiumbatterien gelten als gefährlich und müssen laut V⎵⎵⎵⎵orschrift …

Byte 32 now keeps a literal U+0020 in the visual layer and write_font_face
maps U+0020 onto the very glyph the font puts at that code, so the painted
shape is unchanged and the browser can space it. The visual run is escaped with
escape_markup rather than escape_text, whose   substitution is a
different character that word-spacing does not move.

/Rotate

pdf_document_parser.cpp resolved and normalized page.rotate, and nothing
ever read it. It now composes onto begin_page's box transform, and a quarter
turn swaps the reported page dimensions. svg_background_with_page_rotation_issue402.pdf
— the file added for issue 402 — renders landscape for the first time.

Glyph names through the font's charset

abstract::Font gains glyph_for_name, implemented on CffFont against the
charset; Font::glyph_for_code consults it before the glyph-list detour, as
9.6.6.2 prescribes for simple Type1/CFF fonts. Subset producers name glyphs
gidNNNNN, which no glyph list translates, so those files fell all the way
through to the code-as-GID last resort and painted unrelated glyphs
(Gutenbergstr, 21 → 6cm87G8utemun▯▯2). It also picks the exact named
variant a producer asks for — one file had been drawing a comma where the pdf
asked for hyphen.case.

Flate streams without an ADLER32 trailer

A producer that ends its streams at the final deflate block and omits the
4-byte checksum made the whole file fail to decode. zlib_inflate drops the
check (ghostscript does the same, loudly) and clears the input queue, which
also stops stray bytes trailing a stream from being appended to what it
decodes.

Test

• Full suite green; reference output regenerated for both repos and all four
  data pins advanced.
• New unit tests for the name lookup at both layers: CffFontTest.SelectsGlyphByName
  and PdfFont.simple_font_glyph_for_code_via_font_charset.
• Each fix was checked page by page against a ghostscript raster of the same
  file.

Known gap left for follow-ups: JBIG2 image masks and filled form-field
appearance streams.