CORE-006: bayan-xml — Markup-Compatibility-aware XML with lossless preservation

Context

WordprocessingML relies on namespaces (dozens of Microsoft extension namespaces) and on Markup Compatibility and Extensibility (ECMA-376 Part 3: mc:Ignorable, mc:AlternateContent, mc:ProcessContent, mc:MustUnderstand). Lossless round-trip requires capturing unknown content exactly and writing untouched elements verbatim. quick-xml is the planned base (versions from 0.41 fix two 2026 denial-of-service advisories).

Objective

An XML layer that parses OOXML parts safely, applies Markup Compatibility rules correctly, captures unknown content losslessly (both as verbatim source spans and as re-serializable fragments), and writes XML deterministically.

Scope

In scope

• Streaming reader on quick-xml (pinned, at least 0.41) that rejects document type declarations and external entities, supports only predefined and character entities, and enforces limits on depth, attribute count, name length and total size.
• Namespace resolution with a table of known OOXML namespaces (WordprocessingML and its 2010–2023 extensions, relationships, DrawingML families, VML, Office, Math, Markup Compatibility, packaging), including the mapping between Strict and Transitional namespace URIs.
• Markup Compatibility processing per ECMA-376 Part 3: ignorable namespaces not understood are skipped for interpretation but preserved; AlternateContent selects the first Choice whose Requires namespaces are understood, else Fallback, while preserving the whole element for round-trip; ProcessContent, MustUnderstand (error if not understood), PreserveElements/PreserveAttributes.
• Lossless capture: (a) byte ranges of elements in the source part (source spans) and (b) re-serializable fragments carrying the namespace declarations they need, so a fragment can be emitted into a different context.
• Deterministic writer with namespace management, correct escaping, xml:space="preserve" handling, stable attribute order for generated content, and verbatim emission of source spans.
• Fuzz targets for the reader and for the Markup Compatibility processor.
• Tests: real WordprocessingML snippets, the Markup Compatibility examples from the standard, Strict-namespace documents.

Out of scope

WordprocessingML semantics (CORE-102 onward).

Deliverables

The bayan-xml crate, fuzz targets, tests, documentation.

Acceptance criteria

• AC-1 Parse-then-write of sample parts is semantically identical under a canonical comparison, and verbatim spans are byte-identical.
• AC-2 Markup Compatibility test cases (from the standard’s examples plus our own) pass, including nested AlternateContent.
• AC-3 Documents containing a DTD or external entity are rejected; entity-expansion attacks are impossible by construction (test).
• AC-4 Fuzz targets ran for at least one CPU-hour without crashes.
• AC-5 No unsafe; dependencies justified.

Verification

cargo xtask verify; fuzz durations recorded.

Escalate if

quick-xml cannot provide exact source offsets needed for verbatim spans; propose an alternative.