fix: roll up [label](/n/…) chrome on first editor paint

[chasehuh]

Closes #84.

The bug

On first open of a note, in-app Markdown links painted raw — [label](/n/id) with
all chrome visible. One Enter rolled them up into the blue label. Worse than
cosmetic: hrefAtPos reads the same tree, so those links were unclickable until
that first edit.

Root cause — not the race it looks like

The obvious read is "ensureSyntaxTree(…, 50) times out on a long note". It doesn't —
the 50ms budget is almost never the binding constraint. It's a deterministic
stale-snapshot read.

// @codemirror/language
class LanguageState {
  constructor(context) { this.context = context; this.tree = context.tree }   // snapshot, once
  static init(state) {
    let vpTo = Math.min(3000 /* Work.InitViewport */, state.doc.length)
    …                                                                          // 20ms budget
  }
}
function syntaxTree(state)      { return field ? field.tree : Tree.empty }      // the SNAPSHOT
function ensureSyntaxTree(…)    { … return result }                            // the ADVANCED tree

collectMarkdownLinks called ensureSyntaxTree for its side effect on the shared
mutable parse context, discarded the return value, and then iterated
syntaxTree(state) — the snapshot from LanguageState.init, which covers at most the
first 3000 chars. Links outside it have no Link node: no decorations, no click target.
The first edit constructs a LanguageState around the by-then-advanced context, so the
links roll up. Hence "press Enter once".

Measured (jsdom, decorations read from the facet)

doc	doc len	syntaxTree(state).length	ensureSyntaxTree	first paint	after one edit
link at start, short	39	39	ok	✅	✅
link at start, 87k	86930	86930	ok	✅	✅
link at end, 87k	86930	86890	ok	❌ none	✅
link at end, 889k	888930	888890	ok	❌ none	❌ still none

The snapshot ends exactly 40 chars — the final paragraph — short of the doc while
ensureSyntaxTree reports success: the complete tree was available and thrown away.
Deterministic, because an incremental Markdown parse deliberately leaves the last line
open, so once a doc passes the 3000-char init viewport the snapshot systematically
excludes the trailing paragraph.

The 889k row is why docChanged was the other half of the bug: LanguageState.apply
only advances to max(mapPos(treeLen), viewport.to), so "press Enter" is not even a
general workaround.

The fix

1. Iterate the tree ensureSyntaxTree returns (documentTree), falling back to
   syntaxTree when the budget is exhausted — a partial tree still decorates everything
   it covers.
2. Rebuild when the tree changes, not only on docChanged. The background
   parseWorker reports progress with an effect-only transaction
   (Language.setState.of(new LanguageState(field.context))), which
   syntaxTree(tr.startState) !== syntaxTree(tr.state) catches — the canonical CM6 idiom
   for syntax-dependent decorations. This makes the budget-exhausted case self-healing
   instead of permanent.

Reading tr.state inside a StateField.update is safe here: @codemirror/state assigns
tr._state = this before computing any slot, and field() routes through ensureAddr,
which computes a dependency on demand with cycle detection. The language field has no
dependency on these fields, so no cycle — and it does not rely on markdown() being
installed first, though it is.

Tests

New agentnoteLinks first paint block in lib/editor/links.test.ts. Assertions read
decorations out of view.state.facet(EditorView.decorations) rather than the DOM —
jsdom has no layout, so the viewport is a couple of lines and a link further down
never renders regardless of correctness. That is why the existing DOM-based decoration
test never caught this.

• link past the initially parsed region is label-marked and its ](url) replaced, with
  no doc change — fails on main
• hrefAtPos resolves it on first paint — fails on main
• control: a link inside the initially parsed region still decorates — passes both ways,
  so the suite proves the fix rather than the fixture

Verification

• pnpm vitest run — 358 passed / 34 files (was 355; +3)
• pnpm exec tsc --noEmit — clean
• pnpm build — compiled successfully
• eslint lib/editor components — 12 problems before, 12 after (all pre-existing)

Notes / residuals

• Parse-progress transactions now trigger a full-document rebuild. Bounded: it stops once
  the parse completes, and each rebuild's ensureSyntaxTree short-circuits through
  parse.isDone(upto).
• Part 2 has no dedicated test. Isolating it needs a document large enough to
  deterministically blow the 50ms budget, which is machine-dependent and would be flaky in
  CI; part 1 alone already fixes every case measured above, including 889k chars. Part 2 is
  defense-in-depth for slower hardware and bigger notes.
• Decoration layer only — no wire-format change, and image/video widgets build their own
  decorations and are untouched.

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
memo	Ready	Preview	Aug 5, 2026 12:02pm