Answer architecture round 3: a link alone holding a titled image keeps its caption, and the decision's title names []()

AGENTS.md

@@ -25,7 +25,7 @@ In `docs/decisions.md`:
 - CommonMark is a subset
 - Tables
 - Links
-- Every CommonMark image and empty link reads: the image where it stands alone, else a link
+- Every CommonMark image and empty link reads: the image where it stands alone, else a link, `[]()` as nothing
 - Ids stay site-local
 - Portable task ids come from position
 - A callout title keeps its link targets

README.md

@@ -245,9 +245,10 @@ Serves Goals 1, 4 and 5.
     `[]()` as nothing.
   - An image anywhere but alone in a plain paragraph reads as its alt text linked to its URL,
     `See ![the moon](moon.png).` as `See [the moon](moon.png).`, and inside a link as plain text
-    the link marks.
-  - An image's title reads as its caption where the image stands alone in a plain paragraph, as
-    its link's title where it reads as alt text linked to its URL, and inside a link it drops.
+    the link marks — unless the link's whole text is the image and the link stands alone in a
+    plain paragraph, which reads as the image the link marks.
+  - An image's title reads as its caption where the image reads as an image, as its link's title
+    where it reads as alt text linked to its URL, and inside a link otherwise it drops.
 
   And three more: a list continuing past a marker change stays one list against CommonMark's two,
   a shortcut reference matching its definition only under Unicode case folding stays unresolved,

docs/decisions.md

@@ -202,7 +202,7 @@ link reference definition — and a directive link CommonMark could spell is ref
 link — the bracket form goes literal, the directive form refused — which is CommonMark's prose
 where its reference implementation nests one `<a>` in another.
 
-## Every CommonMark image and empty link reads: the image where it stands alone, else a link
+## Every CommonMark image and empty link reads: the image where it stands alone, else a link, `[]()` as nothing
 
 2026-10-04, the maintainer and a writer panel, 3 of 3; `[]()` 2026-10-09, settled by Goals 4 and 7
 as the maintainer's plan item directed. Goals 4, 5, 6 and 7. Valid while `mediaInline` carries no

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

@@ -1014,6 +1014,8 @@ test('reads a titled image alone in its paragraph as the image captioned with it
 test('reads an image alone in a link as the image the link marks, and the link a title gives as its own', () => {
   const linked = (node: AdfNode, mark: AdfMark): AdfNode => ({ ...node, content: (node.content ?? []).map((media) => ({ ...media, marks: [mark] })) })
   assert.deepEqual(content(losslessMarkdownToAdf('[![a](/u)](/v "t")\n')), [linked(image('/u', 'a'), link('/v', 't'))])
+  const captioned = linked(image('/u', 'a'), link('/v'))
+  assert.deepEqual(content(losslessMarkdownToAdf('[![a](/u "c")](/v)\n')), [{ ...captioned, content: [...(captioned.content ?? []), { content: [text('c')], type: 'caption' }] }])
   assert.deepEqual(content(losslessMarkdownToAdf('[![a][i]][v]\n\n[i]: /u\n[v]: /v\n')), [linked(image('/u', 'a'), link('/v'))])
   assert.deepEqual(content(losslessMarkdownToAdf('[[![a](/u)](/v)](/w)\n')), [{ content: [text('['), marked('a', link('/v')), text('](/w)')], type: 'paragraph' }])
 })