donislawdev · Oct 15, 2026 · Oct 15, 2026 · Oct 15, 2026 · Oct 15, 2026
@@ -14,6 +14,32 @@ because it turns other people's test suites red.
 
 ## [Unreleased]
 
+### Breaking
+
+- **An HTML page with no label has a title that says Page.** The smallest
+  page and every page written with `--clean` carried an empty title, which
+  the HTML standard does not allow and the Nu HTML Checker reports as an
+  error. They now carry `<title>Page</title>`, a word that says nothing
+  about this tool, since `--clean` is there to keep it out of the file.
+  Pages with a label are unchanged. The smallest HTML page grows from 118 B
+  to 122 B, so 118 B to 121 B are now refused, and `tfg preset eject
+  empty-and-minimal` prints 122 where it printed 118 - a manifest from a run
+  of that preset records a different `recipe_hash`.
+- **The catalogue of every element renames its forms section and its hidden
+  field.** With `elements=every` the forms section is `id="tfg-forms"` and
+  the hidden field is `name="tfg-hidden"`, the way the list of links has
+  always been `tfg-links`. DOMPurify removes an id or a name that shadows a
+  property of the document, so after sanitising, the two links to the forms
+  section led nowhere and the hidden field left a submitted form without its
+  name. The smallest catalogue grows from 6924 B to 6944 B.
+- **A newsletter's page says its language.** With `layout=newsletter` the
+  HTML part opens with `<html lang="en">`, as the pages of the `html` format
+  do, and the line in Polish and Japanese marks each half with its own
+  language, `<span lang="pl">` and `<span lang="ja">`, so a screen reader
+  does not read them as English. The smallest newsletter grows by 10 B in
+  7bit, and by 57 B to 78 B with 8bit, quoted-printable and base64, which
+  carry that line. A message laid out `simple` is unchanged.
+
 ### Added
 
 - **Films: the `webm` format.** A WebM video with an AV1 picture and no
@@ -312,6 +338,13 @@ because it turns other people's test suites red.
 
 ### Changed
 
+- **The `eml` settings say what Thunderbird and MimeKit do with them.**
+  Measured with Thunderbird 157.0.1 and MimeKit 4.18.1. `charset` says that
+  MimeKit with no code page provider tries UTF-8 and then Latin-1, and reads
+  Latin-2 and Windows-1250 right once one is registered. `message_depth` says that Thunderbird shows 32 levels in full,
+  and past that shows only the headers of the 32nd message inside, without a
+  word. `filename_style` says that Thunderbird calls an attachment with no
+  name Part 1.2, Part 1.3 and so on.
 - **`empty-and-minimal` now includes the films, GeoJSON and mail.** The preset builds the
   smallest file of every format this build has, so with `formats` left at
   `all` it gains a `webm`, an `mp4`, a `geojson` and an `eml` target, and the recipe

@@ -47,7 +47,7 @@ func newMessage(s settings, seed uint64, children, pics []child) *message {
 	return &message{
 		s: s, id: id, children: children, pictures: pics, seed: seed,
 		plainLine: s.internationalLine("", ""),
-		htmlLine:  s.internationalLine("<p>", "</p>"),
+		htmlLine:  s.htmlInternationalLine(),
 		images:    imageLines(s, pics),
 		frame:     newFrame(s, id),
 	}
@@ -189,7 +189,7 @@ func (m *message) plain(o *out) {
 // the pictures before them, in its frame.
 func (m *message) html(o *out) {
 	eol := o.eol
-	o.put(m.frame.htmlBefore, "<html>")
+	o.put(m.frame.htmlBefore, m.frame.htmlTag)
 	if m.parity {
 		o.put(" ")
 	}

@@ -44,15 +44,15 @@ type frame struct {
 	headers string
 	// plainOpen and plainClose stand before and after the plain text.
 	plainOpen, plainClose string
-	// htmlBefore stands before <html>, htmlOpen after it and htmlClose after
-	// the paragraph of words.
-	htmlBefore, htmlOpen, htmlClose string
+	// htmlBefore stands before the html element's start tag, htmlTag is that
+	// tag, htmlOpen stands after it and htmlClose after the paragraph of words.
+	htmlBefore, htmlTag, htmlOpen, htmlClose string
 }
 
 // simpleFrame is the frame of a message laid out simple: the body element and
 // nothing else, the bytes this format has always written.
 func simpleFrame(s settings) frame {
-	return frame{htmlOpen: "<body>" + s.textEOL(), htmlClose: "</body></html>"}
+	return frame{htmlTag: "<html>", htmlOpen: "<body>" + s.textEOL(), htmlClose: "</body></html>"}
 }
 
 // newsletterFrame is the frame of a newsletter, unsubscribe being the https
@@ -67,7 +67,13 @@ func newsletterFrame(s settings, unsubscribe string) frame {
 			"You receive this because you subscribed to "+newsletterName+".") +
 			frameLine(s, "Unsubscribe ("+unsubscribe+")"),
 		htmlBefore: frameLines(s, "<!DOCTYPE html>"),
-		htmlOpen:   frameLines(s, newsletterHead...),
+		// The page says its language, as the pages of the html format do. The
+		// Nu HTML Checker asks for it, and a screen reader reads the page in
+		// the voice it names (until 2026-10-15 the newsletter had none). In the
+		// frame rather than in the writer, so quoted-printable writes its = as
+		// =3D like every other attribute.
+		htmlTag:  frameLine(s, `<html lang="en">`),
+		htmlOpen: frameLines(s, newsletterHead...),
 		htmlClose: frameLines(s, newsletterFoot(unsubscribe)...) +
 			frameLine(s, "</html>"),
 	}

@@ -282,7 +282,7 @@ func charsetProperty() format.Property {
 		Default: CharsetUTF8,
 		Detail: "The character set of the text when text_encoding is not 7bit. iso-8859-2 and windows-1250 hold Polish and no Japanese, " +
 			"and they write ą, ś and ź as different bytes, so a program that takes one for the other shows the wrong letters. " +
-			"MimeKit with no code page provider registered reads both as Latin-1, and Go's standard library converts no character set at all. " +
+			"MimeKit with no code page provider registered tries UTF-8 and then Latin-1, which shows the Polish letters of both wrong, and reads them right once one is. Go's standard library converts no character set at all. " +
 			"With 7bit the text is plain ASCII and says us-ascii, whatever this is set to.",
 	}
 }
@@ -341,7 +341,7 @@ func filenameStyle() format.Property {
 		Detail: "How the names of attached files and pictures are written. rfc2231 is the standard way. " +
 			"rfc2047 is the encoded form RFC 2047 forbids inside a name, and many readers decode it anyway. " +
 			"both writes the two, and content-type puts the name only where old mail programs looked for it. " +
-			"none writes no name at all, and a program that needs one has to make one up. " +
+			"none writes no name at all, and a program that needs one has to make one up - Thunderbird 157 calls them Part 1.2, Part 1.3 and so on. " +
 			"Go's standard library reads no name from content-type and the encoded text itself from rfc2047. " +
 			"Unless headers is encoded a name is written as it is and only where it goes changes, " +
 			"so rfc2047 and both write the same bytes.",
@@ -357,7 +357,8 @@ func depthProperty() format.Property {
 		Default: "1",
 		Detail: "How many messages deep the message goes. 1 is a message alone. 2 carries a forwarded message as an attachment, " +
 			"and each step adds one more inside the last, the smallest message each can be. " +
-			"Python 3.14's email package runs out of recursion at around 490, and MimeKit stops taking them apart past 513 without a word.",
+			"Python 3.14's email package runs out of recursion at around 490, and MimeKit stops taking them apart past 513 without a word. " +
+			"Thunderbird 157 shows up to 32 in full. Past that the 32nd message inside shows only its headers and nothing under it appears, without a word.",
 		Group: attachmentsGroup,
 	}
 }
@@ -30,7 +30,8 @@ const (
 // Windows-1250 write as different bytes.
 const (
 	polish        = "Zażółć gęślą jaźń."
-	international = polish + " 日本語のテキストです。"
+	japanese      = "日本語のテキストです。"
+	international = polish + " " + japanese
 )
 
 // singleByte is how ISO-8859-2 and Windows-1250 write the letters of the
@@ -97,7 +98,36 @@ func (s settings) internationalLine(open, close string) string {
 	if !s.international() {
 		return ""
 	}
-	line := open + s.encodedLine() + close
+	return s.textLine(open + s.encodedLine() + close)
+}
+
+// htmlInternationalLine is the line as the HTML part writes it. A newsletter's
+// page says it is English (<html lang="en">), so there each half says which
+// language it is in a span of its own - the Polish half, and the Japanese one
+// where the character set holds it - or a screen reader reads both as English
+// (review of #190, as the html format marks its paragraphs in other scripts).
+// A message laid out simple says no language anywhere and keeps the bytes it
+// always wrote.
+func (s settings) htmlInternationalLine() string {
+	if s.layout != Newsletter {
+		return s.internationalLine("<p>", "</p>")
+	}
+	if !s.international() {
+		return ""
+	}
+	// Two lines rather than one where the Japanese half comes: the spans make
+	// the line 83 characters, past the 78 RFC 5322 asks for, and 8bit has no
+	// soft line break to keep it under (measured by the structural checker).
+	// A line break inside a paragraph is white space, as the space it replaces.
+	if _, single := singleByte[s.charset]; single {
+		return s.textLine(`<p><span lang="pl">` + s.encodedLine() + `</span></p>`)
+	}
+	return s.textLine(`<p><span lang="pl">`+polish+`</span>`) + s.textLine(`<span lang="ja">`+japanese+`</span></p>`)
+}
+
+// textLine is one line of the text in the encoding the text is written in,
+// with the text's line ending.
+func (s settings) textLine(line string) string {
 	if s.encoding == QuotedPrintable {
 		line = quotedPrintable(line, s.eol)
 	}

@@ -42,15 +42,22 @@ const (
 		`<meta name="description" content="Every element of the HTML standard, each one at least once">` + "\n" +
 		"<style>\nmark { background: #ffeb3b }\n</style>\n"
 
+	// catalogueBody is the elements themselves. The forms article is
+	// tfg-forms rather than forms for the reason the links list is tfg-links:
+	// an id that shadows a document property is taken out by DOMPurify 3.4.17,
+	// and the two links to it then lead nowhere (O282, until 2026-10-15). The
+	// hidden field is named tfg-hidden for the same reason - a name hidden
+	// shadows document.hidden, DOMPurify took it out, and the field left a
+	// submitted form without one (measured 2026-10-15).
 	catalogueBody = `<header>
 <hgroup><h1>Every element</h1><p>Every element of the HTML standard, each one at least once.</p></hgroup>
-<nav><ul><li><a href="#text">Text</a></li><li><a href="#tables">Tables</a></li><li><a href="#forms">Forms</a></li><li><a href="#media">Media</a></li></ul></nav>
+<nav><ul><li><a href="#text">Text</a></li><li><a href="#tables">Tables</a></li><li><a href="#tfg-forms">Forms</a></li><li><a href="#media">Media</a></li></ul></nav>
 <search><form><label>Search <input type="search" name="q"></label> <button type="submit">Find</button></form></search>
 </header>
 <main>
 <article id="text">
 <h2>Text</h2>
-<p>Plain <a href="#forms">link</a>, <abbr title="HyperText Markup Language">HTML</abbr>, <b>bold</b>, <bdi>name</bdi>, <bdo dir="rtl">reversed</bdo>, <cite>a title</cite>, <code>code()</code>, <data value="42">forty two</data>, <del>removed</del>, <dfn>term</dfn>, <em>stressed</em>, <i>alternate</i>, <ins>added</ins>, <kbd>Ctrl</kbd>, <mark>marked</mark>, <q>quoted</q>, <s>no longer accurate</s>, <samp>output</samp>, <small>fine print</small>, <span>span</span>, <strong>important</strong>, H<sub>2</sub>O, x<sup>2</sup>, <time datetime="2026-10-14">14 October 2026</time>, <u>annotated</u>, <var>n</var>, long<wbr>word, a line<br>break.</p>
+<p>Plain <a href="#tfg-forms">link</a>, <abbr title="HyperText Markup Language">HTML</abbr>, <b>bold</b>, <bdi>name</bdi>, <bdo dir="rtl">reversed</bdo>, <cite>a title</cite>, <code>code()</code>, <data value="42">forty two</data>, <del>removed</del>, <dfn>term</dfn>, <em>stressed</em>, <i>alternate</i>, <ins>added</ins>, <kbd>Ctrl</kbd>, <mark>marked</mark>, <q>quoted</q>, <s>no longer accurate</s>, <samp>output</samp>, <small>fine print</small>, <span>span</span>, <strong>important</strong>, H<sub>2</sub>O, x<sup>2</sup>, <time datetime="2026-10-14">14 October 2026</time>, <u>annotated</u>, <var>n</var>, long<wbr>word, a line<br>break.</p>
 <p><ruby>base<rp>(</rp><rt>annotation</rt><rp>)</rp></ruby> and <math><mi>x</mi><mo>=</mo><mn>2</mn></math> and <slot name="note">slot fallback</slot> and <tfg-note>a custom element</tfg-note>.</p>
 <h3>Third level</h3>
 <h4>Fourth level</h4>
@@ -78,12 +85,12 @@ const (
 <tfoot><tr><td colspan="3">footer row</td></tr></tfoot>
 </table>
 </article>
-<article id="forms">
+<article id="tfg-forms">
 <h2>Forms</h2>
 <form>
 <fieldset>
 <legend>Every type of input</legend>
-<label>hidden <input type="hidden" name="hidden" value="h"></label>
+<label>hidden <input type="hidden" name="tfg-hidden" value="h"></label>
 <label>text <input type="text" name="text"></label>
 <label>search <input type="search" name="search"></label>
 <label>tel <input type="tel" name="tel"></label>

@@ -195,7 +195,7 @@ func labelledHead(s settings, r format.Request, min int64) (head, heading string
 		// does - one for the tab and one for the reader. A fragment has no head
 		// to put a title in, so it carries only the heading, and the label is
 		// still visible because a heading is.
-		head = strings.Replace(plain, emptyTitle, "<title>"+label+"</title>", 1)
+		head = strings.Replace(plain, placeholderTitle, "<title>"+label+"</title>", 1)
 		extra = int64(len(head) - len(plain))
 	}
 	if extra+int64(len(heading))+min <= r.Bytes {

@@ -78,14 +78,15 @@ const (
 
 // plain is how a page has been spelled since 2026-08-01, the golden hashes
 // holding it there. The skeleton is in two halves so the catalogue and the
-// style sheet can go into the head between them, and the title is empty until a
-// label fills it - the empty pair is what the smallest document carries.
+// style sheet can go into the head between them, and the title says Page until
+// a label fills it - what the smallest document and every --clean one carry
+// (an empty title until 2026-10-15, O277).
 var plain = markup{
 	headOpen: `<!DOCTYPE html>` + "\n" +
 		`<html lang="en">` + "\n" +
 		`<head>` + "\n" +
 		`<meta charset="utf-8">` + "\n" +
-		emptyTitle + "\n",
+		placeholderTitle + "\n",
 	headClose: `</head>` + "\n" +
 		`<body>` + "\n",
 	bodyClose:  "</body>\n</html>\n",

@@ -9,8 +9,13 @@ package htmlfile
 // The markup itself - the skeleton, the label, the levels and the blocks - is
 // spelled in markup.go, in every layout and every way of carrying styles.
 
-// emptyTitle is the title a page carries until a label fills it in.
-const emptyTitle = `<title></title>`
+// placeholderTitle is the title a page carries when no label fills it in: at
+// the smallest sizes, and in every page under --clean. Not empty, because the
+// standard asks a title for text and the Nu HTML Checker calls an empty one an
+// error (O277). A word that says nothing about this tool, because --clean is
+// there to keep the tool's description out of the file
+// (docs/HTML-6-2026-10-15.md).
+const placeholderTitle = `<title>Page</title>`
 
 // byteOrderMark is the UTF-8 mark, the only one a page may carry: the standard
 // requires UTF-8 (docs/HTML-STRUCTURE-2026-09-08.md section 1). Written as its

@@ -135,6 +135,16 @@ func TestTheEMLCheckRefusesANewsletterBrokenOnPurpose(t *testing.T) {
 			".frame { width: 600px; background: #ffffff; }\r\n"+
 			".button { background: #1a73e8; color: #ffffff; text-decoration: none; }\r\n", "<style>\r\n"),
 			"layout=newsletter and the HTML lacks: style sheet in its head"},
+		// The page's language, which the frame writes since 2026-10-15 and
+		// nothing read before (docs/HTML-6-2026-10-15.md).
+		{newsletter, newsletter, replaced(`<html lang="en">`, "<html>"),
+			"layout=newsletter and the HTML lacks: language of the page"},
+		// Each half of the line in Polish and Japanese says its language, since
+		// the page says English (review of #190).
+		{newsletter, newsletter, replaced(`<span lang="pl">`, "<span>"),
+			"layout=newsletter and the HTML lacks: the Polish line marked as Polish"},
+		{newsletter, newsletter, replaced(`<span lang="ja">`, "<span>"),
+			"layout=newsletter and the HTML lacks: the Japanese line marked as Japanese"},
 	}
 	dir := t.TempDir()
 	for i, c := range cases {

@@ -458,6 +458,13 @@ func pageGoldenCases() map[string]engine.Target {
 			Properties: map[string]string{"structure": "fragment", "bom": "true", "nesting": "2"}},
 		"html_64kib_everything": {ID: "g", Format: "html", Sizes: engine.Uniform(1, 65536), Label: true,
 			Properties: map[string]string{"nesting": "600", "characters": "international", "elements": "every", "bom": "true"}},
+		// The two pages no label fills the title of, which carry the
+		// placeholder: the smallest document, where the label has no room, and
+		// a page under --clean. No case held either until 2026-10-15, when the
+		// placeholder went from empty to Page and moved bytes nothing pinned
+		// (O277, docs/HTML-6-2026-10-15.md).
+		"html_smallest_page": {ID: "g", Format: "html", Sizes: engine.Uniform(1, 122), Label: true},
+		"html_4kib_clean":    {ID: "g", Format: "html", Sizes: engine.Uniform(1, 4096), Label: false},
 		// Pictures inside the page: the default format, and an odd size in a
 		// fragment inside levels, with a format whose own bytes are text. Each
 		// picture is the file its format writes alone, so these also pin the

@@ -143,7 +143,8 @@ func TestEachHTMLShapeAnswersForItsOwnMinimum(t *testing.T) {
 // clothes of a feature. Saying nothing and saying "document" out loud are two
 // different routes through the parser and have to meet.
 func TestTheDefaultHTMLShapeIsTheBytesItAlwaysWrote(t *testing.T) {
-	for _, size := range []int64{118, 119, 400, 4096} {
+	smallest := htmlMinimum(t)
+	for _, size := range []int64{smallest, smallest + 1, 400, 4096} {
 		silent := writeEncoded(t, "html", size, nil)
 		spoken := writeEncoded(t, "html", size, shapeProps(shapeDocument))
 		if !bytes.Equal(silent, spoken) {
@@ -215,3 +216,16 @@ func TestTheStructuralCheckIsToldWhichHTMLShapeToExpect(t *testing.T) {
 		}
 	}
 }
+
+// htmlMinimum is the smallest page the registry declares, asked rather than
+// written down. The two tests that ask for it were written with 118 B and 119 B
+// in them, and went red on 2026-10-15 when the minimum moved to 122 B for a
+// reason that had nothing to do with what they hold (O277).
+func htmlMinimum(t *testing.T) int64 {
+	t.Helper()
+	d, err := format.Get("html")
+	if err != nil {
+		t.Fatal(err)
+	}
+	return d.MinBytes
+}
@@ -229,7 +229,8 @@ func TestEveryHTMLSettingIsRecordedInTheManifest(t *testing.T) {
 func TestTheDefaultHTMLSettingsAreTheBytesItAlwaysWrote(t *testing.T) {
 	spoken := map[string]string{"nesting": "0", "characters": "ascii", "elements": "basic", "bom": "false",
 		"formatting": "lines", "styles": "none", "links": "0", "inline": "0", "inline_format": "png", "inline_size": "8kb"}
-	for _, size := range []int64{118, 119, 4096} {
+	smallest := htmlMinimum(t)
+	for _, size := range []int64{smallest, smallest + 1, 4096} {
 		if !bytes.Equal(writeEncoded(t, "html", size, nil), writeEncoded(t, "html", size, spoken)) {
 			t.Errorf("at %d B: saying nothing and saying every default produce different bytes", size)
 		}