Add unknown text BM25 fallback

[Mathews-Tom]

Stack: C3 language coverage expansion

Position: 4/4
Base: feat/lang-extraction-fixtures
Head: feat/unknown-file-fallback

Changes:

• Add deterministic 80-line windows for unknown UTF-8 text files.
• Keep binary files skipped.
• Keep unknown chunks BM25-only by excluding them from vector, rerank, and SPLADE embedding paths.
• Reduce cold parse/index cost by extracting symbols and imports from one tree-sitter parse per file.
• Reduce chunking/index cost by caching chunk candidate text/token counts and skipping surrogate generation for BM25-only indexes.

Validation:

• uv run ruff check && uv run ruff format --check . && uv run pyright passed.
• uv run pytest tests/parse/ tests/pipeline/ tests/index/test_chunker.py tests/test_performance.py::TestQueryCacheSkipsParse -q reported 510 passed, but exits non-zero because coverage is configured for the whole source tree.
• uv run pytest -q reported 2240 passed, 4 deselected, 89.33% coverage.
• Operator run before the optimization completed 35 benchmarks and archex benchmark gate printed Quality gate passed.
• Existing-language recall regression check: 0 regressions across archex_query, archex_query_fusion, and archex_query_fusion_rerank.
• Local micro-timing on the archex_project_index query improved from pre-fix mean parse/index/total 490.6/1632.9/2316.6 ms to 325.0/1092.4/1560.8 ms including first cold import, and approximately 264/1079/1499 ms after warm import. Rerun the operator benchmark gate to confirm the 10% benchmark-wide index-time ceiling with full task coverage.