feat(pdf): decode jbig2 images

[andiwand]

🤖 Generated with Claude Code

Stacked on #702.

Three scanned certificate pages of Fahrradtraeger_Uebler_Primavelo_P2_P3.pdf
came out blank. Their page body is a /JBIG2Decode image mask: the filter
framework stopped at the codec and Do skipped the image. There is no library
to defer to the way pdf_jpx defers to openjpeg, so pdf_jbig2 decodes it in
house — ~900 lines, no new dependency.

What it covers

The coding a scanner emits, and nothing beyond it:

• the MQ arithmetic decoder (T.88 Annex E, software conventions)
• the arithmetic integer and symbol-id decoders (Annex A)
• generic regions, all four templates, with typical prediction (TPGDON)
• symbol dictionaries and text regions, every reference corner and transposed
• page information and composition through the external combination operators
• the PDF embedded-stream organization, /JBIG2Globals included

MMR/Huffman coding, refinement and halftone regions are refused. Refusing skips
the image exactly as today rather than painting a page missing its content.

Notes

The generic-region template is sorted by row and column rather than laid out in
the spec's bit order. That is a relabeling of the arithmetic contexts, and the
decoder adapts per context, so any bijection over the same pixel set decodes
identically — the same trick pdf.js uses, and the ordering it produces is the
spec's for nominal AT pixels, which is what the TPGDON context constants
assume.

/JBIG2Globals names its shared symbol dictionary by reference. Only the
parser can follow one, so the bytes reach the filter through a new
DecodeOptions rather than the filter growing document knowledge.

Test

• 10 unit tests over the segment-header walk (long referred-to form, four-byte
  page association), the page and packing path, and every refusal.
• Full suite green. Exactly one reference-output file changes — the pdf that
  prompted this — and its three scanned pages now match a ghostscript render of
  the same file, at the mask's own 2352×3328 rather than the raster's.
• The correctness of the arithmetic decoding is guarded by that snapshot; there
  is no encoder to round-trip against and I did not want to invent a test
  vector.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: b64b1de9dd

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".