fix: keep the form a code block was authored with

[Azganoth]

Summary

A save rewrote every code block into one shape. Indented code became a fenced block, a tilde fence became a backtick fence, the space between an opening fence and its info string was dropped, a fence indented up to three spaces moved to column zero, and a fence the author left unclosed gained a closing run. The tilde rewrite is the damaging one: an info string may hold a backtick only when the fence is spelled with tildes, so changing the fence forced the info string to change too and ~~~ language + backtick saved as a backtick fence carrying `. The block still read back as the same code, but the info string another tool reads was no longer the one the author wrote.

Both forms parse into one code node carrying a value and an info string, and mdast-util-to-markdown picks one form and one fence spelling for the whole document, so nothing on the node answered for either. The form survives only in the slice of the file the node was built from, and the head of that slice names it: an indented block's position opens on the line its indentation is written on, while a fence's opens at the fence itself, past whatever indentation the file gave it, so indented code can never stand on a fence run — its own four spaces stand there first. That holds inside a container too, where the position opens past the container's prefix rather than at the line, which is what #321 asked to settle before the fence axes were worth attempting.

Both choices reach the serializer through mdast-util-to-markdown's own options rather than being written directly, because its indented branch also holds the conditions CommonMark puts on that form. A block carrying an info string, opening or closing on a blank line, or holding nothing but whitespace cannot be written indented and gives way to a fence, which is what keeps an edit inside a block from writing a file that reads back as something else. Setting the fence option to the recorded character is also what puts the literal backtick back, because the serializer marks a backtick unsafe only inside a grave-accent fence — no unescaping pass is involved.

The length is kept as the surplus over the shortest run that can hold the content rather than as the run the file spelled. A fence has to outrun anything inside it, so the length is a floor the content can raise at any time, and recording the number itself made the record shift on the save that raised it. corpus/commonmark/code.md reaches that without an edit: a backtick in a backtick info string opens no fence, so the rest of that file parses as one fence holding backtick runs of its own and the run has to widen from three to four. Measured against the same floor on the way back in, the surplus survives it.

An open fence is ended by whatever ends the block it stands in rather than only by the end of the file, so one left open at the end of a blockquote, a list item, or a footnote definition is written open and the blocks after that container still read outside the code. It is recorded open only where it stands last in its container, which is the only place one can be written open — recorded anywhere else it names a form the file can never hold and the record flips on the save that closes it.

Indentation is kept only at the document root, where the column the fence opens at is the indentation and nothing else. Inside a container that column also counts the prefix the container wrote and mdast names neither separately, so a fence inside a container keeps no indentation of its own. Deriving the prefix from the block's own lines was implemented, measured, and withdrawn: it assumes the container writes the same prefix on the line the block opens on as on the lines under it, which a footnote definition does not — it writes its label on the first line and indents the rest by four, so the difference read as indentation the file never wrote and was written into the code the block holds. A form axis that can reach the content is narrower than one that cannot.

Related Issue

Closes #320
Closes #321

Verification

markdownCompatibility.test.ts gains a Code block form block of 34 cases. Seven cover the indented form — at the top level, holding a blank line, holding a fence of its own, and inside a list item, a blockquote, and a list item holding a blockquote — and each was confirmed to fail without the change, reproducing #321's reported output including the ```` cascade. Ten cover the fence axes the same way: both fence characters, a run longer than the content needs, both spacings before an info string, a tilde info string holding a literal backtick, and one- and three-space indentation.

The rest are guards. A block made by insert.codeBlock and by format.codeBlock writes the default fence; an indented block whose content grows a leading blank line gives way to a fence; a recorded run rises when content outgrows it and a recorded surplus stays above the run the content needs; a fence left open at the end of a blockquote, a list item, or a list item holding a blockquote stays open; an unclosed fence is closed once a block follows it inside its container; a fence standing alone with no final newline is not read as its own closing run; and a fence inside a footnote definition keeps clear of the label's own width, which is the case that withdrew the container prefix derivation. The container cases assert the reopened document alongside the written bytes, because byte equality cannot see a form record that flips or indentation pushed into content.

corpusRoundTrip.test.ts moves isolated/end-of-file/unclosed-code-fence.md to byteIdenticalFiles, completing that directory at four of four, which is what #320 named as the reason to separate the unclosed case.

Two existing cases in markdownCompatibility.test.ts asserted the old behavior — an indented-code first child inside a list item written back as a fence. They are folded into the list of non-paragraph first children that now round-trip unchanged, which keeps the coverage they existed for (the block stays a first child inside its item) under a stricter assertion.

Measured across all 82 Markdown files under corpus/, mounting each, saving, and remounting: against origin/main the only status change is isolated/end-of-file/unclosed-code-fence.md moving from rewritten to byte-identical. Every other file converges and preserves its document exactly as it did before, including the hand-verified subtrees the automated guard does not cover. extensions/frontmatter/yaml.md and practical/extension-survey.md fail document preservation on origin/main as well and are untouched by this change.

Compared against Typora 1.14.9, the editor Preserve the form a file was written in cites, on three fixtures covering twelve form cases. Manually verified on Windows 11:

1. Opened each fixture in Typora and saved it, then diffed Typora's output, Leafdown's output, and the original against each other.
2. A fence left unclosed at end of file: both write it back byte-identically.
3. A fence left unclosed at the end of a blockquote: both leave it open. This is what the last commit changed — Leafdown closed it before, which was the only unclosed-fence difference between the two.
4. Indented code at the top level, inside a list item and inside a blockquote; both fence characters; a backtick in a tilde info string; info-string spacing; an over-long run; a fence indented three spaces at the top level: identical in both.

pnpm check:frontend passes.

Not verified: no manual pass in the assembled application, and pnpm check:backend was not run. The change is parse-time and serializer behavior with no rendered surface of its own, and it touches no Rust or Tauri code.

Notes

• Three differences from Typora remain, all outside what this pull request records. A fence indented inside a blockquote is written flush, which is the axis withdrawn above. A closing fence indented differently from the fence that opened the block is written at the opening fence's indentation; CommonMark lets the two differ, but a second indentation is a further axis rather than part of the one A code fence's authored form is rewritten on save #320 names. An info string keeps its spacing but not its escapes or character references, so language+one© ``` is written back decoded — that is the character-reference class under A character reference is decoded and saved as the literal character #262.
• Four spaces do not interrupt an existing paragraph. loses its indentation, which Indented code is rewritten as a fenced block on save #321 puts out of scope as a read-time difference no serializer change reaches. Typora keeps it, because its model holds the raw line rather than the parsed paragraph; that is a parser difference rather than a serializer one.
• corpus/commonmark/lists-and-blockquotes.md loses its code-block differences entirely; what remains is one lazy continuation line gaining its block marker, which is A lazy continuation line gains its block marker on save #322. corpus/commonmark/code.md stays converging on the classes above plus the escaped backtick runs on its A backtick in a backtick info string does not open a fence line. Saving inserts a blank line between blocks the author wrote adjacent #324 is already merged, so it is no longer what either file is waiting on.
• The content lines of an indented fence are written at that fence's indentation. That is what makes a uniformly indented root-level block byte-exact, and it adds bytes only where content sat at column zero under an indented fence — the parse strips up to the fence's indentation from each content line, so how much each line originally carried is not recoverable.
• blockSeparatorMarkdown.ts now reads a code fence under up to three spaces. A fence never carried its own indentation into that guard before, and without the widening the blank line before an indented fence could no longer be dropped, which shifted a recorded separator in corpus/commonmark/code.md.
• withCodeBlockSeparator is replaced by withCodeForm, so the code block carries the block separator alongside its own form the way every block Leafdown holds a form for already does.
• The form is not carried through parseDOM and toDOM, so a block copied out of one document and pasted into another is written with the default fence. That matches how a heading's authored form already behaves, and the fallback is the safe direction.