Fix quick-xml 0.42 build breakage + ParseKit 1.0 / anydoc plan

[cpetersen]

Two independent commits. The first unbreaks main; the second is planning only.

1. main does not compile (e82bc41)

PR #51 (Bump quick-xml from 0.41.0 to 0.42.0) was merged 2026-09-05 with a failing build check. bundle exec rake compile has failed on main ever since, with 5 errors from two quick-xml 0.42 breaking changes:

• LocalName now implements AsRef<str> rather than AsRef<[u8]>, so the std::str::from_utf8(...) wrapper no longer type-checks.
• BytesText now holds an already-decoded Cow<'i, str>; decode() is gone, replaced by xml10_content() / xml11_content() / html_content().

7 insertions, 8 deletions. 334 examples, 0 failures.

Behavior is preserved — xml10_content() additionally unescapes entities, but that path is unreachable here (see the entity bug below).

The other seven recent dependabot PRs (#44–#50) all merged green, so this looks like an isolated slip rather than a systemic gap. Worth confirming branch protection requires build to pass.

This commit can be fast-tracked ahead of reviewing the docs.

2. Planning docs (478ca8e)

Evaluation of anydoc and the resulting 1.0 plan. No code changes.

Doc	What it covers
ANYDOC_EVALUATION.md	What anydoc is, measured head-to-head against our own fixtures, and where it has no answer
PARSEKIT_1_0_PLAN.md	Target architecture, ParseKit::Document API, 7-phase migration, spec-coverage plan, settled decisions
PARSEKIT_BIN_GEM.md	What the parsekit-bin gem on RubyGems actually is

Headline findings

anydoc is materially better on office formats. Measured on spec/fixtures/:

Fixture	parsekit today	anydoc
sample.docx	tables silently dropped	full GFM pipe table
sample.xls	RuntimeError	parses correctly
sample.pptx	one run-on line	headings + tables per slide

Speed is a wash (~1.4ms vs ~1.0ms per doc), so correctness and structure are the argument, not performance.

Two more pre-existing bugs surfaced while planning:

1. XML/HTML entities are silently dropped. <a>Tom &amp;amp; Jerry</a> yields "Tom Jerry". quick-xml emits entity references as a separate Event::GeneralRef (in 0.41 and 0.42) and our _ => {} arm swallows them. This matters because parse_xml is code the migration keeps — anydoc has no HTML/XML support.
2. Two specs are load-bearing on bugs. integration_spec.rb:61 asserts the .xls failure via raise_error(/Failed to parse Excel file/); integration_spec.rb:24 checks for "Table example" (the paragraph label) but never the table's contents, which is exactly how the dropped-table bug stayed green. Both must be inverted during migration, not deleted.

Recommendation

Adopt anydoc as the office/PDF engine; keep Tesseract for images and the text/JSON/XML/HTML handlers. Drops MuPDF, docx-rs, calamine, zip, and regex; deletes ~450 of parser.rs's 630 lines; adds .doc, .rtf, .odt, .ods, .odp, .epub, .xlsb for free.

parsekit-bin

Not ours and not yankable — owned by lucas__domeij, origin ssh://github.com/Teamtailor/parsekit-bin. Verified benign: source is byte-identical to our 0.1.2 except a version string and a calamine bump, extconf.rb is unmodified, and the precompiled binary links only libSystem / libiconv / libc++ with no networking. It exists because they needed precompiled gems in Jan 2026 and we did not ship them until 0.2.0 in June. Decision: no action against them; instead add aarch64-linux to our release matrix, the one platform they ship that we do not.

🤖 Generated with Claude Code

https://claude.ai/code/session_01U2V9AiGqoaLwUQzqJAzE1q