[yhay81]

Closes #18
Closes #224

pyropdf moves to hayro 0.9.0 with LaurenzV/hayro#1734 applied, for release in 0.2.2, by the owner's decision of 2026-11-03 (recorded on #18). Awaiting the owner's review. The owner chose all three below on 2026-11-04 (「推奨どおりで」, as recommended), and the tag of 3 is pushed.

The patch

• Until hayro releases hayro-interpret: Expose the font dictionary of a glyph run LaurenzV/hayro#1734 (GlyphRun::font_dict(), hayro: expose the font dictionary of a glyph run #584), [patch.crates-io] takes every hayro crate from yhay81/hayro 394e4a0d: hayro main 4f7d16fa (0.9.0 plus hayro-font hayro-font: Read Unicode from a ToUnicode map written with cidranges LaurenzV/hayro#1730 and hayro-ccitt hayro-ccitt: Allocate fallibly LaurenzV/hayro#1733) with #1734 on top. The hayro crates name each other by path, so patching hayro-interpret alone would bring a second hayro-syntax whose types are not the store's; from one commit, Cargo.lock lists each hayro crate once.
• fuzz/Cargo.toml repeats the table, and fuzz/Cargo.lock follows (the fuzz targets type-check against it). Build against released hayro again: remove the [patch.crates-io] of LaurenzV/hayro#1734 #603, blocked by hayro: expose the font dictionary of a glyph run #584, removes both.
• Releases build with it. maturin sdist (1.15.0) keeps the table and the git sources of Cargo.lock, so a build from the sdist fetches 394e4a0d from GitHub.

Engine (the branch's earlier commits)

• Object numbers are u32 at every hayro call site (I8), and Reference loses its i32 caveat.
• Glyph runs are grouped by the font dictionary of the run (GlyphRun::font_dict()): the key is Dict::cache_key() of the font dictionary, as 0.8's font_cache_key was, so fonts over one program stay apart. Glyph use for subsetting, the fonts table, and text spans read it from GlyphStyle::font.
• RenderContent replaces render_annotations, begin_marked_content reads the MCID from the property list, and hayro-jpeg2000 is 0.5.0, hayro's.

The failures of #18's last handoff

Failure	Cause	Settled by
content tail (Rust, Python), test_unreadable_content_leaves_the_document_unchanged ×2	hayro-syntax reads up to 33 operands (LaurenzV/hayro#1599)	tests use 34; docs say 34th
page_content_joins_streams_as_hayro_does	hayro skips /Contents items that are not streams (LaurenzV/hayro#1600)	PageContent::read skips them too, instead of stopping; tests show the skip
test_text_render_modes_follow_the_painting	text in mode 7 reaches the device as invisible text (LaurenzV/hayro#1677)	the test expects an INVISIBLE run; docstrings say so; an extraction test shows the text is extracted
value::tests: [/AB#FG]	hayro-syntax 0.9 reads the name as written	the test checks the name and key order; the panic guard stays for panics not known yet (#19)
value::tests: -1 0 R	object numbers are u32	two integers and a keyword read as null
value::tests: 65 levels	hayro-syntax 0.9 parses 512 levels	the store keeps 64 as its own bound, below
subset renders ×8, and test_font_subset_poppler.py in CI	hayro hints text by default (LaurenzV/hayro#1705)	pyropdf renders unhinted, as with 0.8, until the owner chooses (below)

Nesting (F6). Cloning, comparing, dropping, copying for edit, and decrypting a value recurse into it, as does the binding's conversion of an operand, all bounded by hayro-syntax 0.8's 64 levels. The store's conversion now refuses deeper values itself (value::MAX_NESTING), so objects nested 65 to 512 levels stay malformed input as in 0.2.1, and the comments that named hayro's bound name the store's.

#224

A stream whose /F names a file parses with its dictionary and bytes (hayro-syntax 0.9, LaurenzV/hayro#1681), and the store refuses to decode it, as hayro-syntax does, instead of returning the bytes to ignore; without a /Filter the store returned them unread. The round-trip test of every typed view writes /F on every stream type, and the spec map's engine rule is gone (read: 634 done; no engine status left).

Tests added

• A file whose page tree, font, and content are numbered 2^31, 3,000,000,000, u32::MAX − 2, and u32::MAX − 1: they resolve, render, extract, take an edit, and save incrementally and fully (hayro-syntax: object numbers as u32 (I8) #18's acceptance).
• The 64-level bound in the store, the engine, and Python; /F streams in the store and in Python; clip-only text in extraction; a luminosity soft mask; the deprecation warning of SoftMask.xref, and that nothing else of a mask warns.
• The writer's property tests generate references to any u32 number and streams with /F, which they left out for hayro-syntax 0.8.

Owner's choices (#18, 2026-11-04)

Evidence: evidence/hinted-subsets. Rendered hinted, mhlw-doc.pdf after subset_fonts differs from the original in 5,907 and 1,349 pixels of its two pages, by up to 239 of 255, and unhinted in none; canvas text in a CID-keyed CFF face differs in 873 of 36,000 pixels after the save. The auto-hinter of TrueType outlines reads 32 Latin and 150 CJK reference characters through the cmap, which a subset empties; a compacted CFF program loses its hints.

1. Hinting: pyropdf renders unhinted in 0.2.2, as this PR does, so that a patch release changes no render. Keeping the auto-hinter's reference characters in subsets is rejected: about 150 CJK glyphs more in each subset. Follow-ups: Render text hinted on request: a hinting option of page.pixmap and Document.render #612, a hinting=True option of page.pixmap and Document.render, which only adds to the API; Keep a CFF face's hints when a save compacts it #611, compacted CFF programs that lose their hints, which matters to other viewers too.
2. SoftMask.xref: kept as 0, and reading it warns with DeprecationWarning. A private field holds the value, so building (SoftMask(subtype=...) and SoftMask(xref=..., subtype=...)), comparing, hashing, printing, and pickling a mask do not warn, and it stays a frozen dataclass. Not restored through hayro's Debug hash, which is no API. Remove the deprecated SoftMask.xref in the next minor release #610 removes it in the next minor release.
3. The tag pyropdf/glyph-run-font-dict on 394e4a0d goes to yhay81/hayro, so that released sdists keep fetching the commit if the branch of hayro-interpret: Expose the font dictionary of a glyph run LaurenzV/hayro#1734 is force-pushed. Pushed: a signed annotated tag, 972ea320.

For 0.2.2's changelog

• pyropdf builds against hayro 0.9.0 with hayro-interpret: Expose the font dictionary of a glyph run LaurenzV/hayro#1734 applied, which hayro has not released: every hayro crate comes from yhay81/hayro commit 394e4a0d, which also holds two fixes after 0.9.0 (ToUnicode maps written with cidrange, hayro-font: Read Unicode from a ToUnicode map written with cidranges LaurenzV/hayro#1730; fallible allocation in CCITT decoding, hayro-ccitt: Allocate fallibly LaurenzV/hayro#1733), and a build from the sdist fetches it from GitHub (hayro-syntax: object numbers as u32 (I8) #18).
• Objects numbered above 2,147,483,647 resolve in rendering and text extraction (hayro-syntax: object numbers as u32 (I8) #18).
• A stream whose /F names an external file reads, with its dictionary and its bytes, and decoding its data raises MalformedPdfError instead of returning the bytes to ignore (hayro-syntax: parse a stream whose dictionary names an external file #224).
• Content operators with up to 33 operands, such as scn with up to 32 DeviceN components, are read, rendered, and extracted; reading stopped at the 11th operand, which also left the rest of the page unrendered.
• An item of a page's /Contents array that is not a stream is skipped, and the streams after it are read; reading stopped at it.
• Text in render mode 7, which only clips, is extracted, and page.content() reports it as an INVISIBLE text run.
• A name whose # is not followed by two hexadecimal digits reads as written instead of raising MalformedPdfError.
• SoftMask.xref is deprecated: reading it warns with DeprecationWarning, and it is 0, since hayro 0.9 no longer identifies the group of a soft mask. The next minor release removes it.
• Operations.parse and page.operations raise MalformedPdfError for an operand nested deeper than 64 levels, where reading stopped before it.
• hayro 0.9.0's fixes and performance work reach rendering and extraction, as its changelogs list them (patterns, soft masks, clipping, embedded fonts, the Unicode of glyph names, malformed files); text still renders unhinted.

Checks

On Windows, after rebasing onto v0.2.1:

• cargo fmt --all --check, cargo clippy --workspace --all-targets --all-features --locked -- -D warnings, cargo test --workspace --all-features --locked: pass (engine 184, store 327 and 2 ignored, corpus 2)
• uv run pytest: 6,544 passed, 689 skipped
• uv run ruff check ., uv run ruff format --check ., uv run ty check: pass.
• cargo check of fuzz/ against the patch: pass.
• The sdist of maturin 1.15.0 installs from source into a fresh Python 3.14 environment (12 min 48 s, with Cargo's git cache warm), and the result opens, extracts, and renders mhlw-doc.pdf. GitHub serves 394e4a0d by its hash (git fetch --depth 1), as Cargo fetches a full rev.
• CI on d30ae65: green, the six fuzz targets included.

Follow-ups: #603 (remove the patch), #608 (TilingPattern.opacity, which hayro 0.9 leaves to devices), #610 (remove SoftMask.xref in the next minor release), #611 (keep CFF hints in compacted programs), #612 (a hinting option). After this lands, #584 no longer blocks #19, #365, and #437, which the patched commit already serves.

🤖 Generated with Claude Code