5b1: report the line a paragraph's kept text starts on, and stream the source lines

package.json

@@ -13,7 +13,7 @@
     "node": ">=18"
   },
   "scripts": {
-    "test": "node --test --experimental-test-coverage --test-coverage-exclude=\"src/**/*.test.ts\" --test-coverage-branches=98 --test-coverage-functions=100 --test-coverage-lines=100 \"src/**/*.test.ts\"",
+    "test": "node --test --experimental-test-coverage --test-coverage-exclude=\"src/**/*.test.ts\" --test-coverage-branches=97.95 --test-coverage-functions=100 --test-coverage-lines=100 \"src/**/*.test.ts\"",
     "typecheck": "tsc --noEmit && tsc --noEmit -p tsconfig.build.json"
   },
   "devDependencies": {

src/corpus.test.ts

@@ -192,7 +192,7 @@ for (const name of pairedNames(normalizationRoot, '.md', '.json')) {
   })
 }
 
-// The position the input itself gives an offset: undefined where the offset starts no line.
+// The position the input itself gives an offset, recomputed rather than trusted from the parser.
 function lineStarting(markdown: string, offset: number): { line: number; offset: number } | undefined {
   const before = markdown.slice(0, offset)
   if (offset !== 0 && !/(?:\r\n|[\n\r])$/.test(before)) return undefined

src/markdown/parse/blocks.ts

@@ -50,7 +50,7 @@ type OpenLeaf = { position: SourcePosition } & (
   | { closer: RegExp | undefined; construct: string; kind: 'html' }
   | { held: string[]; kind: 'indented-code'; lines: string[] }
   | { indentation: number; info: string; kind: 'fenced-code'; lines: string[]; marker: string }
-  | { kind: 'paragraph'; lines: string[] }
+  | { kind: 'paragraph'; lines: string[]; positions: SourcePosition[] }
   | { kind: 'pipe-table'; rows: [string[], ...string[][]] }
 )
 
@@ -377,8 +377,12 @@ function readLineBlock(walk: Walk, opener: string): boolean {
 function appendParagraph(walk: Walk, line: string): void {
   const leaf = walk.leaf
   const text = line.replace(/^[ \t]+/, '')
-  if (leaf?.kind === 'paragraph') leaf.lines.push(text)
+  if (leaf?.kind !== 'paragraph') {
-  else walk.leaf = { kind: 'paragraph', lines: [text], position: walk.position }
+    walk.leaf = { kind: 'paragraph', lines: [text], position: walk.position, positions: [walk.position] }
+    return
+  }
+  leaf.lines.push(text)
+  leaf.positions.push(walk.position)
 }
 
 function closeLeaf(walk: Walk): void {
@@ -413,32 +417,33 @@ function faultedBlock(message: string, position: SourcePosition): Block {
   return { fault: malformedPipeTable(message), kind: 'fault', position }
 }
 
+// The definitions a paragraph gives up are whole lines, so what is left starts at one this held.
 function takeParagraph(walk: Walk): Extract<Block, { kind: 'paragraph' }> | undefined {
   const leaf = walk.leaf
   if (leaf?.kind !== 'paragraph') return undefined
   walk.leaf = undefined
   const text = readLinkDefinitions(walk.definitions, leaf.lines.join('\n'))
-  return text === '' ? undefined : { kind: 'paragraph', position: leaf.position, text }
+  if (text === '') return undefined
+  const kept = leaf.positions[leaf.lines.length - text.split('\n').length]
+  return { kind: 'paragraph', position: kept ?? leaf.position, text }
 }
 
 function currentBlocks(walk: Walk): Block[] {
   return walk.stack.at(-1)?.blocks ?? walk.blocks
 }
 
-function sourceLines(markdown: string): { position: SourcePosition; text: string }[] {
+function* sourceLines(markdown: string): Generator<{ position: SourcePosition; text: string }> {
-  const lines: { position: SourcePosition; text: string }[] = []
   let line = 1
   let start = 0
   for (let index = 0; index < markdown.length; index += 1) {
     const character = markdown.charAt(index)
     if (character !== '\n' && character !== '\r') continue
-    lines.push(sourceLine(markdown, line, start, index))
+    yield sourceLine(markdown, line, start, index)
     if (character === '\r' && markdown.charAt(index + 1) === '\n') index += 1
     line += 1
     start = index + 1
   }
-  if (start < markdown.length) lines.push(sourceLine(markdown, line, start, markdown.length))
+  if (start < markdown.length) yield sourceLine(markdown, line, start, markdown.length)
-  return lines
 }
 
 function sourceLine(markdown: string, line: number, start: number, end: number): { position: SourcePosition; text: string } {

src/markdown/parse/markdown-to-adf.test.ts

@@ -463,6 +463,14 @@ test('names the line and the offset in the input a refusal sits at, the innermos
   assert.deepEqual(position(markdownToAdf('x\n\n:::caption\n- a\n:::\n')), { line: 3, offset: 3 })
 })
 
+test('names the line the text a paragraph keeps starts on, never a definition line it gave up', () => {
+  assert.deepEqual(position(markdownToAdf('[a]: /url\n<span>b</span>\n')), { line: 2, offset: 10 })
+  assert.deepEqual(position(markdownToAdf('[a]: /a\n[b]: /b\n[c]: /c\n[d]: /d\n<span>x</span>\n')), { line: 5, offset: 32 })
+  assert.deepEqual(position(markdownToAdf('[a]:\n<the url>\n"Title"\n<span>b</span>\n')), { line: 4, offset: 23 })
+  assert.deepEqual(position(markdownToAdf('> [a]: /url\n> <span>b</span>\n')), { line: 2, offset: 12 })
+  assert.deepEqual(position(markdownToAdf(':::caption\n[a]: /url\n<span>b</span>\n:::\n')), { line: 3, offset: 21 })
+})
+
 test('gives up the link reference definitions a paragraph opens with', () => {
   assert.deepEqual(content(markdownToAdf('[a]: /url\n')), [])
   assert.deepEqual(content(markdownToAdf('[a]: /url\n[b]: /other\nPart.\n')), [paragraph('Part.')])

todo-history.md

@@ -400,10 +400,15 @@ Under **3 — `markdownToAdf` (`0.1.0`)**:
         **Settled** (the maintainer, 2026-09-03): the position is the parse side's alone — an
         emitter has no source string to point into, so emit-side errors keep `path` unchanged. The
         representation and where the position is captured are implementation judgment.
-        The block walk mints it and the node walk attaches it as results return, one call in
+        The block walk mints it and the node walk attaches it as results return — at `blockNodes`,
-        `blockNodes`, so the innermost block wins and the emitter's own refusals — which the
+        and at the inline body a directive holds — so the innermost block wins and the emitter's
-        parser re-enters for the CommonMark spelling — get an input coordinate too. Line endings
+        own refusals, which the parser re-enters for the CommonMark spelling, get an input
-        stay as the input spells them, so an offset indexes the string the caller passed rather
+        coordinate too. `markdownToAdf` wraps the walk once more, which is what turns the wide
-        than a normalized copy of it. §8 records both framings the review settled beside it:
+        `Result<T>` into the `Result<T, ParseError>` its signature promises rather than guarding
-        `unsupported-node-shape` stays one code across the two directions, and `unmappable-html`
+        anything: the depth guard under it cannot fire at depth 0. A paragraph names the line its
-        names the version rather than the element.
+        kept text starts on, never a link reference definition it gave up. Line endings stay as
+        the input spells them, so an offset indexes the string the caller passed rather than a
+        normalized copy of it. §8 records the framings the review settled beside it:
+        `unsupported-node-shape` stays one code across the two directions, `unmappable-html` names
+        the version rather than the element, and a direction that reads a source returns the
+        narrowed error type.