Markdown Conformance

Lanexio™ Parser’s Markdown parser passes all 652 CommonMark 0.31.2 examples and all 648 GFM examples.

Curated CommonMark gate (CI)

CommonMark 0.31.2 is continuously validated in CI through two surfaces:

• Package harness (pnpm test:packages): every one of the 652 spec examples has its own it() block, rendered through renderMarkdownHtml and compared byte-for-byte to the spec HTML, honoring the package ledger (packages/parser-grammar-markdown/src/spec/expectations.json).
• Repo-level aggregate (pnpm test:curated-suites, see ADR 0034): one aggregate assertion runs the same oracle over all 652 examples so a single pnpm test:repo-gate validates toml-test, xmlts, and CommonMark together. Any example whose outcome contradicts the ledger fails the aggregate.

The ledger currently carries 0 expect-fail and 0 skip entries: 652/652 pass.

Serializer roundtrip

The Markdown serializer (serializeMarkdown) was delivered in phases MD-S1 through MD-S4. All CommonMark and GFM features roundtrip correctly: parse → serialize → re-parse produces AST-identical trees.

Phase	Features	Tests
MD-S1	Leaf blocks (thematic breaks, code blocks, HTML)	1,841
MD-S2	Headings, paragraphs, inline formatting	1,864
MD-S3	Links, images, blockquotes, lists	1,880
MD-S4	GFM tables, character escaping, golden roundtrip	1,898

Test suite results

Suite	Tests	Passed	Failed
CommonMark 0.31.2	652	652	0
GitHub Flavored Markdown	648	648	0
Total	1,300	1,300	0

The docs/MARKDOWN_GAPS.md ledger in the repository is a frozen Phase 1 pre-flight triage (snapshot 2026-05-01) that predates the completed CommonMark and GFM work; it does not reflect current results.

CommonMark spec coverage

All 652 CommonMark 0.31.2 spec examples pass. Covered sections include:

Section	Examples
Thematic breaks	19
ATX headings	16
Setext headings	26
Indented code blocks	12
Fenced code blocks	29
HTML blocks	44
Link reference definitions	27
Paragraphs	8
Block quotes	25
List items	48
Lists	26
Backslash escapes	13
Entity references	12
Code spans	17
Emphasis and strong emphasis	131
Links	68
Images	22
Autolinks	19
Raw HTML	21
Hard line breaks	15
Soft line breaks	2
Textual content	1

GFM spec coverage

All 648 GFM-specific examples pass. GFM extensions covered:

Extension	Status
Tables	Implemented
Task list items	Implemented
Strikethrough	Implemented
Autolinks (extended)	Implemented
Disallowed raw HTML	Implemented

GFM is default

GFM extensions are enabled by default in parseMarkdown. Pass { gfm: false } for strict CommonMark-only parsing.

Large-file performance

The Markdown grammar is linear-time on large inputs. Four quadratic pathologies were mitigated in a single performance run, all behavior-identical:

• Extended-autolink email fallback: the inline loop skips the email fallback entirely when the inline text has no @, and resolves the residual with-@ case through a once-per-inline-text nextAt table plus a deadAt memo (amortized O(1) per boundary).
• Multi-line paragraph byte offsets: the per-lookup segment-slice scan is replaced by a precomputed byteAt table plus a breakByteOverrides map (one O(n) build, 4 bytes per char, multi-line paragraphs only).
• Document gap fill: fillDocumentGaps builds the node array in a single forward pass instead of splicing O(gaps x n).
• Container close: closeContainersTo rescans only the fresh region between a container and the previously closed inner container, carrying the inner max end forward.

The regression guard packages/parser-grammar-markdown/src/large-file-robustness.test.ts runs four adversarial fixtures (one per pathology) asserting a < 2000 ms 1x parse, a < 3.5 2x/1x scaling ratio, emitSource byte round-trip, and no parse errors.

Spec references

CommonMark 0.31.2 specification
RAG: commonmark_docs::spec@0.31.2

GitHub Flavored Markdown specification
RAG: gfm_docs::spec@gfm-0.29.0

Corpus conformance

pnpm conformance:suite measured this family over the test_files/md corpus on 2026-10-06 at Lanexio™ Parser version 1.0.0:

family	valid	invalid	raw	contract-subset	REAL-DEFECT
markdown	10800	8	99.9%	100.0% (10800/10800)	0

The raw figure is below 100.0% because the directory rule classifies 8 inputs as invalid; the contract-subset over the 10,800 valid inputs is 100.0%.