fix(index): normalize invalid UTF-8 source bytes

[simianhacker]

What

Normalize undecodable source bytes to Unicode replacement characters before constructing Elasticsearch file and line documents. This preserves valid UTF-8 exactly while preventing lone surrogate code points from causing bulk indexing failures.

The issue was reproduced while indexing apache/grails-core at commit fa1e14756daba70da13bde862d94056489171080. Legacy-encoded bytes in grails-forge/grails-forge-api/src/test/resources/features_es.properties were decoded with surrogateescape, producing lone surrogates such as \udcf3, \udcae, and \udce1. Elasticsearch rejected those line documents with Invalid UTF-8: Illegal surrogate character, leaving the ref incomplete.

How to Verify

1. Run uv run --locked pytest tests/test_documents.py -v.
2. Run uv run --locked pytest tests/ -v.
3. With local Elasticsearch credentials configured, index the exact repository revision that exposed the issue:

   uv run sourcerer index github/apache/grails-core \
     --commit fa1e14756daba70da13bde862d94056489171080 \
     --force \
     --retry-window 0s \
     --url "$ELASTICSEARCH_URL" \
     --api-key "$ELASTICSEARCH_API_KEY"

4. Confirm indexing completes without Illegal surrogate character bulk errors. The verified local run completed with 14,026 files and 1,324,487 lines.

Breaking Changes

None