ADF quoting: inline content inside a blockquote-illegal node is flattened to text

[Hinne1]

Found while verifying the --reply-to ADF quoting fix (branch claude/reply-to-adf-quoting).

Behaviour

QuoteADF copies blockquote-legal nodes verbatim. A node the content model rejects (heading, table, panel, expand) is flattened with ADFToText. For a node whose children are inline, that flattening is lossy in three ways at once.

Input — a heading with mixed inline runs:

heading[ text("Fixture heading for "), mention(@Hinne Stolzenberg), text(" with inline") ]

Quoted output:

paragraph[ text("## Fixture heading for@Hinne Stolzenbergwith inline") ]

Three losses: the mention node becomes plain text (no notification, no link), the ## markdown prefix leaks into the visible text, and the spaces between adjacent inline runs are eaten.

Not a regression

ADFToText returns byte-identical output on the same input, and the pre-fix code applied that flattening to the entire comment. The current code applies it only to illegal nodes, so this is strictly less lossy than before — but it is the last remaining fidelity gap in quoting.

Suggested fix

For an illegal node whose content is inline, rebuild its children into a paragraph rather than calling ADFToText: keep mention/text nodes as nodes, preserving marks and spacing. paragraph is blockquote-legal, so the result is valid; only the heading level is lost, which a blockquote cannot express anyway.

Reproduction (drop into internal/api, no network needed):

func TestHeadingInlineFlatten(t *testing.T) {
	doc := &ADF{Type: "doc", Version: 1, Content: []ADFContent{{
		Type:  "heading",
		Attrs: &ADFAttrs{Level: 2},
		Content: []ADFContent{
			{Type: "text", Text: "Fixture heading for "},
			{Type: "mention", Attrs: &ADFAttrs{ID: "acc-1", Text: "@Hinne Stolzenberg"}},
			{Type: "text", Text: " with inline"},
		},
	}}}
	q := QuoteADF(doc)
	t.Logf("%q", q.Content[0].Content[0].Text)
	// got: "## Fixture heading for@Hinne Stolzenbergwith inline"
}