fix: Form XObject text garbled when /Resources is an indirect reference

[MADENIYOU]

Fixes #398.

Root cause

FontCMaps::walk_xobject_fonts (src/tounicode.rs) resolved a Form XObject's /Resources entry with:

if let Ok(form_resources) = stream.dict.get(b"Resources").and_then(Object::as_dict) {

This only handles /Resources as an inline dictionary. Per spec, it may equally be an indirect reference to a dictionary — and when it is, .as_dict() on the Object::Reference fails, so the if let silently doesn't run and the Form's fonts (including any /ToUnicode CMap) are never collected. That can produce garbled/empty Markdown and incorrectly route an otherwise-text-based page to OCR.

Interestingly, /XObject resolution a few lines above this already handled both shapes correctly (Ok(Object::Reference(id)) => ..., Ok(Object::Dictionary(dict)) => ...) — /Resources and /Font just hadn't been brought in line with that.

Fix

Added a shared resolve_dict() helper and used it for every dictionary-or-reference lookup in this function (/Resources, /Font, and the font dict's own entries), matching the pattern already used for /XObject.

Also fixed a related gap: the fast/page-selected CMap path (from_doc_pages_fast) skipped Form XObject font collection entirely (collect_cmaps_from_xobjects was gated behind !skip_truetype_fallback), rather than still collecting cheap primary ToUnicode CMaps while only skipping the expensive embedded-font fallback parsing. Threaded skip_truetype_fallback through collect_cmaps_from_xobjects/walk_xobject_fonts so fast mode still discovers primary Form XObject CMaps — matching how the page-level font path already worked.

Removed collect_cmaps_from_fonts, a thin wrapper that became dead code once walk_xobject_fonts calls collect_cmaps_from_fonts_inner directly to pass the propagated flag through.

Testing

• New unit test form_xobject_with_indirect_resources_discovers_tounicode_cmap: builds a minimal in-memory PDF (page → Form XObject → indirect /Resources → Type0 font with /ToUnicode) and confirms FontCMaps::from_doc discovers the CMap. Verified it fails without the fix.
• New unit test form_xobject_indirect_resources_discovered_in_fast_mode_too: same fixture via from_doc_pages_fast. Verified it fails without the second part of the fix (the collect_cmaps_from_xobjects gating).
• cargo fmt, cargo test (948 unit tests passing, 2 new), cargo clippy --all-targets -- -D warnings clean.

🤖 Generated with Claude Code

---

Summary by cubic

Fixes garbled/empty text when a Form XObject’s /Resources is an indirect reference by resolving it correctly, and ensures fast mode still discovers primary ToUnicode CMaps on Form fonts. Previously, indirect /Resources caused font CMap collection to be skipped; now both inline and indirect dictionaries are handled without changing the fast mode’s skip of expensive TrueType fallback parsing.

Review notes

• Adds resolve_dict() and uses it for /Resources, /Font, and nested font dict lookups; aligns with existing /XObject resolution.
• Calls collect_cmaps_from_xobjects in both normal and fast paths and threads skip_truetype_fallback through collect_cmaps_from_xobjects and walk_xobject_fonts.
• Removes the thin collect_cmaps_from_fonts wrapper; collect_cmaps_from_fonts_inner is called directly.
• Adds two unit tests covering indirect /Resources in normal and fast modes.

^{Written for commit 17e55dc. Summary will update on new commits.}

[cubic-dev-ai]

No issues found across 1 file

<sub>Shadow auto-approve: would auto-approve. Focused bug fix: resolves indirect /Resources dictionaries so Form fonts' ToUnicode CMaps are collected, and fast mode now still discovers primary CMaps while skipping only expensive fallback. Tests pin both paths; no rollout, contract, or ops change.

Re-trigger cubic</sub>