fix(scan): normalize separators before prompt-injection matching #110

Problem

The prompt-injection rules in the content scanner (content-scanner.ts, used by release scanning) match on raw text. Separators that readers render as spacing or drop entirely slip past them:

• U+0085 (NEL) is not matched by JS \s.
• U+00AD (soft hyphen) is not in INVISIBLE_UNICODE, so Ig\u00ADnore all previous instructions in a valid UTF-8 SKILL.md is published.
• In lossy UTF-8 decodes, single invalid bytes (0x85, 0x93, 0x96, 0x9D, 0xFF after \t) become U+FFFD, which is neither whitespace nor a word character.

Suggested fix

Normalize text before running the injection regexes: strip U+00AD and other default-ignorable code points, and treat U+0085 and U+FFFD as whitespace. Add tests for each separator.

Pre-existing (not a regression). Found during isolated review of #96.