feat(html): extract HTML meta, Dublin Core and Office tags

[swackhamer]

HTML was already detected — File:FileType said HTML — but had no parser and
no FileFormat variant, so every .html/.htm file fell through to the
plain-text fallback and was parsed as TXT. That produced five TEXT:*
statistics ExifTool never reports for an HTML file, and none of the 57 tags it
does.

This adds src/parsers/text/html.rs, a transcription of ExifTool 13.55's
HTML.pm (ProcessHTML), plus the FileFormat::HTML variant, its dispatch arm
and a detection gate.

Measured

combined-samples/HTML.html, per-file Group1:Name scoring, File/System/
ExifTool groups excluded, ExifTool JSON parsed with parse_float=str:

	MATCHED	VALUE-DIFF	MISSING	EXTRA
before	0	0	57	5
after	57	0	0	0

The five EXTRA were TEXT:Encoding, TEXT:FileSize, TEXT:LineCount,
TEXT:LineEnding and TEXT:WordCount — gone now that HTML is no longer treated
as plain text.

Run over the whole 207-file corpus with the same script and the same branch
built with and without the change, the delta is exactly the HTML file:
MATCHED +57, MISSING −57, EXTRA −5. HTML.html is the only file whose score
changed, and no file lost a matched tag. (One unrelated file, XISF.xisf,
drifted by two in ExifTool's own tag count between the two runs; OxiDex matched
zero of its tags both times, so it is not affected either way.)

What the parser does

Five family-1 groups come out of one file, because a meta name carries its own
namespace and ExifTool routes each to a different tag table: HTML:Title (from
the TITLE element), HTTP-equiv:ContentType, HTML-dc:* (9), HTML-ncc:*
(20), HTML-prod:* (2) and HTML-office:* (24, from the MS-Office XML island
inside the <!--[if gte mso 9]> conditional comment).

The details that make values match rather than merely appear:

• HTML-dc:Creator is a list. Two <meta name="dc:creator"> elements
  collapse into ["Phil Harvey","Another Creator"], while dc:language,
  dc:publisher and dc:subject are also List but hold one value each and so
  stay scalars — which is what ExifTool emits.
• Charset. The file declares iso-8859-1; ExifTool maps that to its Latin
  charset, which is cp1252, not bare ISO 8859-1. HTML-office:Category is
  a catégory from a raw 0xE9. Values seen before the content-type element
  are not recoded — HTML.pm says so explicitly and this reproduces it rather
  than "fixing" it.
• Two different unescapers. The meta path uses the full 253-entity HTML
  table (α → HTML-dc:Subject = Greek: α β γ); the XML-island path
  uses only the five XML entities plus numeric references, so
  HTML-office:Description keeps its 
 as a literal CR.
• Conversions are per-tag, not per-kind. HTML-office:CreateDate goes
  through ConvertXMPDate to 2010:06:28 23:52:00Z, while HTML-dc:Date has
  no ValueConv and stays raw as the malformed 2007-30-01.
  HTML-office:TotalEditTime is ConvertTimeSpan($val, 60) → 1 minute.
• Derived names go through two steps. <o:Checked_x0020_by> is not in the
  Office table, so HTML.pm mangles the name (hex escape → space →
  capitalise) and then AddTagToTable normalises it (strip illegal characters,
  ucfirst, Tag prefix when it is under two characters or does not start with
  a letter). Missing that second step reported unknownNsWeird_thing where
  ExifTool reports UnknownNsWeird_thing.
• o:Revision and o:Version both map to RevisionNumber and neither is a
  list, so the later element wins (12.0), matching ExifTool.
• No declared charset means no recoding, so a stray high byte reaches
  ExifTool's writer raw and its FixUTF8 turns each bad byte into one ?.
  Rust's lossy conversion collapses a whole invalid run into a single U+FFFD,
  which is a different string, so FixUTF8 is reproduced instead.

Verified against the Perl, not by hand

Three checks against ExifTool's own source, because a hand-invented table that
ships with tests asserting its own contents is the failure mode here:

• Tag tables. All seven shipped tables (112 tags) were diffed key by key
  against Image::ExifTool::GetTagTable for HTML::{Main,dc,ncc,prod,vw96, equiv,Office} on tag ID, Name and List flag. Identical.
• Entity table. The 253 entries were extracted from %entityNum, and the
  shipped Rust table was then re-parsed and diffed against a Perl eval of that
  block. Identical.
• ConvertTimeSpan. Compared against the Perl for every integer
  o:TotalTime from 1 to 200000 — covering the seconds, minutes, hours and days
  branches. Byte-identical output, which settles %d truncation and %.1f
  rounding.

Nothing omitted

All 57 tags reproduce ExifTool's exact values, so nothing was dropped as
un-reproducible. One guard is in place for a case this file does not exercise:
ConvertTimeSpan's sub-minute branch would need Perl's %.15g number
stringification for a fractional o:TotalTime, and rather than print an
approximation under a real tag name the tag is omitted (AGENTS.md).

Detection, and why nothing else moved

The gate is ExifTool's own, from ProcessHTML: an XML declaration alone is not
enough — an actual HTML element must appear in the first 256 bytes. It runs
after the existing SVG and XML-plist root checks, so all four <?xml-rooted
formats keep their own detection.
test_xml_rooted_formats_keep_their_own_detection pins that ordering: HTML,
SVG, Plist, XMP and a plain RDF/XML document (which must stay TXT) each assert
their own FileFormat.

Verified unchanged on the corpus: XMP.svg → SVG, XMP.xmp → XMP,
PLIST-xml.plist → PLIST, XMP.xml / Geotag.xml / LNK.url → TXT,
Text1..5.txt → TXT, LNK.lnk → LNK, VCard.ics → ICS.

Beyond the corpus sample, nine hand-built probes were compared tag for tag with
ExifTool and all match with zero value differences: a single-line document,
CR-only line endings, a UTF-8 charset with a title spanning lines, an
undeclared charset with a raw high byte, an Office island with attributes and a
self-closing element, a document with no <head> at all, an HTML-4 doctype with
a three-value dc.subject list using the . namespace separator, a
windows-1252 document exercising the cp1252-only band, and an XHTML file
(which correctly still reports FileType: XHTML).

cargo test --workspace passes, cargo fmt --all and cargo clippy are clean.

🤖 Generated with Claude Code