feat(svg): recognise an svg by reading it

[andiwand]

🤖 Generated with Claude Code

FileType::scalable_vector_graphics was a label the generic image wrapper would put on any bytes at all — nothing read the file to see whether it was one. Detection now reads it, and it costs no parse of its own.

What is in here

• internal/svg/svg_file.{hpp,cpp} — SvgFile, an abstract::ImageFile over an xml::XmlFile. The xml module parses, rejects what is not well formed and resolves the encoding from the declaration; svg answers the one question left, is_svg_file, against XmlFile::root_name(). So open(file, scalable_vector_graphics) throws NoSvgFile on anything that is not an svg, and the probe reads the xml parse instead of repeating it.
• internal/svg/svg_util.{hpp,cpp} is gone. It held the old check_svg_file(std::istream &), which parsed the bytes a second time to look at one element name. is_svg_file(const xml::XmlFile &) replaces it — a predicate rather than a throwing check, because open_strategy asks the question without wanting the file: an xml that is not an svg is handed on as the XmlFile already built.
• Each layer reaches the one below — SvgFile::xml_file(), text_file(), file(), with text() and document() forwarded, and XmlFile::text_file() under that. A reader downstream needs neither a second parse nor a second decode. One trap is documented on the accessor: text_file()->text() decodes with the encoding detected over the bytes, text() with the one the declaration names.
• NoSvgFile in exceptions.hpp, and the file_type_table row.
• util::string gains to_lower, equals_ignore_case, starts_with_ignore_case and find_ignore_case (second commit). static_cast<char>(std::tolower(static_cast<unsigned char>(c))) was written out in five places; all five now go through it, and is_safe_uri loses the mutation of the scheme it had just built.

Rendering is unchanged

An svg still goes into the page as <img src="data:image/svg+xml;base64,…">, the same path as png and jpeg, and the same path an svg inside a document already took.

The commit list will look odd, so: this branch first inlined the markup — so the drawing would scale to the viewport and its text be selectable — and that is reverted in refactor(svg): keep the detection, drop the drawing. Inside an <img> a browser renders svg in secure static mode: no scripts, no external fetches, animation frozen. Inlined, the markup is live, and it is the only path in the library where the input file authors our output DOM — in a WebView with a bridge to native. Earning secure static mode back by hand meant a scrub (script and embedding elements, event handlers, references that leave the document, SMIL aimed at any of those, css reaching outside) plus a script-src 'none' page — a lot of machinery, and it interferes with the file: a legitimate external image or a webfont went too. Not worth it until something actually needs a scalable, selectable svg; if that comes, isolation in a sandboxed iframe beats modifying the file.

That reasoning is recorded in internal/svg/AGENTS.md rather than only in the git history, together with the rule it follows from, now in the root AGENTS.md: the input file never authors the output markup. Which is where the one place we are inconsistent is written down too — html/pdf_file.cpp filters a PDF /URI action to an allowlist of navigable schemes, while a document hyperlink is only escape_attributed, so a javascript: href in an odt reaches the page. Not touched here; worth deciding separately.

Test plan

• ./test/odr_test --gtest_filter='SvgFile*' — 6 tests: an svg opens as an image that knows it is one, only an svg opens as one (NoSvgFile / NoXmlFile / UnknownFileType), a prefixed root element is still an svg, every layer it was read through is reachable, the declared encoding is what it is decoded with, and it renders as an <img> data url.
• ./test/odr_test --gtest_filter='string_util*' — 8 tests, four of them new, covering the char and string overloads, the non-ascii passthrough, a prefix longer than the string, and from past the end.
• Full suite: 861 passed, 0 failed, 8 skipped (the same 8 as on main — missing private test data).

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: e20c94feb5

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".