SyntaxSuggest 3.0.0 leaks raw invalid bytes into stderr when source has invalid encoding

Since 3.0.0 removed CleanDocument#clean_sweep, sources that contain bytes invalid in their encoding are no longer rejected internally — the raw bytes end up in the annotation. Previously clean_sweep raised ArgumentError and the rescue in core_ext.rb produced a clean fallback.

Under LC_ALL=C, captured error output becomes a US-ASCII string with invalid bytes — any regex match on it raises ArgumentError: invalid byte sequence in US-ASCII. This broke spec/ruby/language/source_encoding_spec.rb on three Ruby CI platforms:

• https://rubyci.s3.amazonaws.com/debian/ruby-master/log/20260420T003003Z.fail.html.gz
• https://rubyci.s3.amazonaws.com/s390x/ruby-master/log/20260419T230004Z.fail.html.gz
• https://rubyci.s3.amazonaws.com/openbsd-current/ruby-master/log/20260419T233005Z.fail.html.gz

I added workaround for this at https://github.com/ruby/ruby/commit/f45bb9ba36cff72eff1bf0c0817a023843a1be34

Start by tracing invalid-encoding handling around CleanDocument#clean_sweep and the rescue in core_ext.rb. Reproduce the issue under LC_ALL=C with a source containing invalid bytes, then inspect spec/ruby/language/source_encoding_spec.rb and the linked Ruby CI failures. Done means raw invalid bytes no longer reach the annotation or captured stderr and the affected encoding behavior is covered.