test: assert corpus round-trip convergence for supported syntax

[Azganoth]

Summary

Delivers the spike outcome from #135: an automated round-trip guard over the corpus subtrees whose syntax Leafdown supports.

• Adds src/features/editor/tests/corpusRoundTrip.test.tsx, asserting that serialize(parse(source)) and a second pass over that result produce the same string for 16 files across corpus/commonmark/, corpus/gfm/, and corpus/isolated/end-of-file/. The editor may normalize on first open; it must not keep changing the document afterwards.
• Notes in CONTRIBUTING.md that those three directories now carry automated convergence coverage, so the manual corpus pass over them is for rendering, interaction, and navigator behavior. The remaining subtrees stay fully manual.

Why convergence, and why it is interim

Byte identity is the property worth wanting here: open a file, save it, the file is unchanged. It is not assertable today. Measured over the scoped set, 12 of the 16 files are rewritten on first open, several by around 100 lines. Three causes:

1. Escaping applied out of context, the dominant one by volume. The serializer escapes characters that would not parse as syntax where they sit — garden_sensor_name → garden\_sensor\_name (intraword _ is never emphasis), *opening-only → \*opening-only (an unpaired * is already literal), name@example → name\@example. This is not forced by the document model, and it is the open question in How a Markdown escape should be represented in the editor #245.
2. Delimiter form the model does not retain: - bullets → *, ~strike~ → ~~strike~~, _em_ → *em*. Fixable per construct, as feat: preserve the authored autolink form on save #243 did for autolinks.
3. Structural spelling the model does not retain: lazy blockquote continuation gains an explicit >, tight lists are written loose.

The 4 files that already round-trip byte-identically do so because they contain almost no inline constructs to escape, not because anything preserves authored form. So there is no meaningful identity-holding subset to pin, and no useful hybrid between the two assertions.

Convergence is therefore chosen as an interim: it holds the editor to a stable document without blessing any particular normalization, needs no baseline or deviation list, and is what catches the #247 class of defect. It should tighten toward identity as #245 settles. A green run here does not mean the corpus saves unchanged, and this file should not be read as endorsing the current normalization.

Nothing new was built to support this. setupMilkdownEditorMount, createMarkdownReferenceContext, and mockTauriApiCommand already cover it, and reading corpus files from disk follows editorPresentation.test.tsx.

Related Issue

Closes #135
Refs #247

Verification

corpusRoundTrip.test.tsx proves the property directly: all 16 scoped files converge. The same test at the spike's baseline failed on three of them (commonmark/lists-and-blockquotes.md, commonmark/text-and-breaks.md, gfm/tables.md), which is what surfaced #247; those three pass here on top of its fix, so the file is a real guard rather than a tautology.

Ran pnpm check:frontend for this frontend-only change.

Not verified: rendering, interaction, and navigator behavior, which convergence says nothing about. This test never types, undoes, or copies.

Notes

Scope, so a green run is not over-read:

• Convergence catches a normalization that keeps changing the document. It cannot catch one that is stable and wrong. **one followed by ** on its own line normalizes to escaped literal text and converges cleanly — correct per CommonMark, since the closing delimiter run follows whitespace, but the point stands that "stops changing" is not "changes correctly".
• The guard asserts the first two passes only.
• It does not cover the open editing-surface defects (Editing at the closing edge of a source projection destroys the construct #231, Typing nested emphasis delimiter runs leaves literal text or duplicates runs #232, Typing a strikethrough inserts extra tildes and swallows the text after it #233, Undo and Redo produce inline states that were never authored #235, Copying a text-only selection writes its Markdown unescaped #244).

Deliberately excluded, per the analysis in #135:

• corpus/boundaries/bytes/, which exists to pin CR, CRLF, BOM, NUL, and a missing final newline — precisely what reading and serializing normalizes away.
• corpus/environment/, whose subject needs the Rust backend and is already covered by the createArticleTree and createFolderContext factories.
• corpus/extensions/, which mostly asserts that unsupported syntax survives as escaped literal text, already covered in markdownCompatibility.test.tsx.
• corpus/practical/, whose local image and link references would turn this into per-path resolution mocks.

Three counts in the #135 analysis were off and are corrected by this branch and its Outcome section:

• isolated/end-of-file/ holds 4 files rather than 3, so the scoped total is 16 rather than 15.
• Only commonmark/links-and-images.md needs the image resolver mock, not two files.
• "Nine of the fifteen candidate files contain a construct the serializer normalizes" understates it substantially. Measured, 12 of 16 differ from source, and the listed constructs are not the main driver — out-of-context escaping is.

Observed while measuring, not investigated and not claimed as a defect: tight lists are serialized loose (- A\n- B → * A\n\n* B), which changes rendered output (<li>A</li> versus <li><p>A</p></li>). Whether the schema distinguishes tight from loose at all is unchecked. Recorded in #135 for whoever picks up #245.