mur loc extract treats Segoe Fluent icon glyphs as translatable strings

[azchohfi]

Summary

mur loc extract treats Segoe Fluent icon glyphs as user-visible text and extracts them into .resw. With --rewrite they are routed through t.Message(...), making a decorative icon depend on a resource lookup — and making it eligible for AI translation.

Repro

samples/TodoApp/Program.cs uses Private Use Area glyphs from the Segoe Fluent icon font as button/label content:

Button("\uE74D", () => structural(new DeleteItem(item.Id)))   // trash can
    .FontFamily((FontFamily)Application.Current.Resources["SymbolThemeFontFamily"])

TextBlock("\uE73A")                                           // decorative checkbox

mur loc extract --source samples/TodoApp --output <out> --rewrite

Produces:

<data name="Text"  xml:space="preserve"><value>&#xE74D;</value></data>
<data name="Text2" xml:space="preserve"><value>&#xE73A;</value></data>

Verified code points (they are not empty strings — they render as blanks in most terminals):

Text   len=1  U+E74D
Text2  len=1  U+E73A

And rewrites the source to:

Button(t.Message(Loc.TodoApp.Text), ...)
TextBlock(t.Message(Loc.TodoApp.Text2))

Why this is wrong

1. A PUA glyph is not linguistic content. There is nothing to translate. mur loc translate would hand U+E74D to an LLM and ask for a French version.
2. It's a translation-corruption risk. A machine or human translation pass over a .resw containing a lone PUA character can easily mangle or drop it, silently breaking the icon.
3. It converts a constant into a lookup. After --rewrite the glyph resolves through IStringResourceProvider; a missing key in some locale degrades or breaks the icon rather than falling back to a glyph that never should have varied.
4. Spec 005 §10.3 says it shouldn't happen. The extractor is documented to ignore "strings that are clearly not user-visible."
5. The generated keys are meaningless. Text / Text2 carry no information about what the glyph is.

Suggested fix

Skip string literals that consist wholly of Private Use Area code points (U+E000–U+F8FF, plus the supplementary PUA planes) — optionally only when the call site also sets a symbol font, though the PUA test alone should be safe and is much simpler.

A stronger version of the same idea: skip any literal containing no letter or digit in any script, which also covers punctuation-only and separator strings.

Notes

Not a systematic false-positive: the samples/ReactorGallery run produced 561 keys with 0 empty or glyph-only values. This is specific to the icon-glyph idiom, which the Gallery happens not to use in extractable positions — but the idiom is common in real WinUI apps and is used in the repo's own sample.

Reproduced with mur built from 88ef6db4.