fix: default indexing to crash-free sequential extraction

[WarGloom]

Summary

• default production indexing to the deterministic sequential extraction path
• retain CBM_WORKERS as an explicit opt-in for parallel regression tests and diagnostics
• make the back-pressure regression test explicitly request parallel extraction

Why

Parallel extraction corrupts heap state on a real mixed-language repository. The failure was reproduced on Serena with the default worker count and with CBM_WORKERS=2 and CBM_WORKERS=4; glibc aborted with allocator-corruption errors. The same 770-file workload completes cleanly with one worker.

This is a stability containment, not a claim that the underlying shared parser/allocator race has been repaired. It trades default indexing throughput for correctness while preserving an explicit route for debugging and testing the parallel implementation.

Verification

• sanitizer sqlite_writer: 12 passed
• sanitizer mcp: passed
• sanitizer pipeline: 226 passed
• make -f Makefile.cbm lint-format CLANG_FORMAT=clang-format
• git diff --check
• clean upstream production build: make -f Makefile.cbm cbm
• Serena clean-cache index: 10,171 nodes, 50,959 edges, zero skipped files