D2: Parser foundation

orgize 0.10.0-alpha.10 against the core corpus (32 files, 1.39 MB: the Org manual, the Org guide, ORG-NEWS, Org's test examples and the Org Syntax specification), with Emacs 30.1 / Org 9.7.11 as the reference through tests/emacs/dump.el.

Measure	Result
Round-trip (to_org() = input=)	32 of 32 files
Panics on corpus files	0
Mutated inputs (200 per file)	6,400 inputs, 0 panics, 0 round-trip failures
Throughput (release build)	62.6 MB/s
Structural agreement with Emacs (Jaccard over element and object start positions)	72.5%

Agreement by type, selected:

Type	Agreement	Main cause
headline, item, src-block, property-drawer, footnotes	100%	--
code, verbatim, entity, macro	99.5% or better	--
keyword	94.0%	Lower-case affiliated keywords (#+attr_texinfo:) before lists and blocks become standalone keywords
link	84.0%	Bracket links containing a newline are not links (open upstream PR since May 2026); plain links such as mailto: are not recognized
paragraph	64.7%	Paragraph starts differ inside descriptive list items; leading blank lines included
table-row, table-cell	55.3%, 0%	Different range conventions for cells and table.el rows
section	2.8%	Different tree shape: planning and property drawers sit in the headline, not in the section
babel-call	50%	Missing cases

• Coverage gaps (§3.2): no citations (org-cite), no inlinetasks, no diary sexp element.
• Context: in-buffer settings such as #+TODO are not read; the caller must build a ParseConfig.
• Incremental parsing: a headline-level replace_range exists. It is not section granular and does not handle keyword changes (R3, R4).
• Tree shape: differs from org-element in several places (sections, planning, cells). Kalem's model, differential tests and plugin API all benefit from matching org-element's shape.
• Maintenance: last commit July 2024, last release June 2024 (still alpha). Pull requests from November 2024 onward are unmerged. The project appears dormant.
• License: MIT, so ideas and test cases can be reused with attribution.

1. Depend on orgize. Rejected: coverage gaps, structural differences and a dormant upstream mean every fix would wait on an inactive maintainer.
2. Fork orgize. Rejected: reaching 99% agreement requires changing the tree shape, the list parser and several element parsers, which is most of the parser. The nom-combinator structure would then fight a design that follows org-element's mode-driven algorithm.
3. New parser (chosen). Follow org-element.el closely: the same element and object types, the same :begin=/:end=/=:contents-begin=/=:contents-end=/=:post-affiliated=/=:post-blank= ranges, the same mode-driven dispatch (org-element--current-element), the same list structure algorithm (org-list-struct) and the same object lexer and restrictions. Differential testing then compares like with like.

• rowan as the tree library, with fine-grained tokens so the UI can hide markers.
• The idea of a typed AST layer over the untyped tree.
• The mutation-based robustness test, which moves into org-syntax's test suite.

• Phase 0 carries the full parser implementation (T0.3).
• The differential harness (tests/emacs/dump.el and the comparison tool) is the main correctness signal from the first day.
• dump.el loads org-inlinetask, so inlinetasks are part of the reference (Kalem enables them by default).