Non-English text is blocked and fully redacted as invisible_text #121

The SDK targets English text for now, but it fails on non-English input in the worst possible way: silently, and by destroying the message.

Guard().scan("今天北京的天气很好，我想去公园散步。") returns block, risk 0.85, injection/invisible_text, and redacted_text is [BLOCKED:injection]. Same for Russian. The English translation returns allow, 0.00.

scanners/injection/scanner.py:52 fires invisible_text whenever the normalizer's homoglyph or fullwidth stage changed anything, without checking that the change looks adversarial. Cyrillic а/е/о/р map to Latin and CJK punctuation goes through the fullwidth stage, so ordinary text trips it. The finding spans 0..len(text), so redaction eats the whole message.

Two things to do, in order:

1. Narrow the invisible_text trigger to actual mixed-script or zero-width abuse rather than any normalizer change, and scope the finding span to the offending characters instead of the whole input.
2. Say plainly in the README that detection is tuned for English today, and track multi-language support as its own piece of work.

Not a first issue. Taking this one.