[adamant-pwn]

Bulk bit-vector construction repeatedly appends packed words. Share the growth path for aligned and overflowing appends, remove the cached last-word pointer, and use unsigned shift counts. Assigning the packed value after emplace_back() avoids passing it by reference into vector growth, removing spills in the measured GCC loop. Existing builder calls require no changes.

In SSHash's encode_strings, median CPU-cycle ratios versus the original append implementation were:

Compiler	DNA, 2047 chars	Protein, 2047 chars	DNA, 63 chars	Protein, 63 chars
GCC 15.3	0.813	0.869	0.848	0.892
Clang 23.1	0.907	0.944	0.815	0.965

Measured on a Ryzen 9 7945HX with -O3 -DNDEBUG -mavx2 -mbmi2, using seven randomized paired blocks per workload and at least 0.55 seconds per sample. Long inputs contained 1,800 records; short inputs contained 60,000. Timing includes stream parsing and encoding, with file I/O and packed-output checks outside the measured section.