Reuse freed GC block IDs without restarting a full-table search for every block #208

BlockDataInfo's constructor in src/hx/gc/Immix.cpp checks gBlockInfoEmptySlots, then searches gBlockInfo from index zero for the first null slot. destroy() nulls a slot and increments the count, but does not record the freed ID. Refilling many holes repeats a growing prefix scan for every block.

After a large heap is compacted and released, regrowing into N empty slots can examine N*(N+1)/2 entries. gBlockInfo retains its historical size, so this affects later allocation bursts even after the managed heap shrinks.

A metadata-only probe at checkout 03a4e8d copied the constructor's first-null-slot search over a cleared hx::QuickVec<void*>. Linux x86_64 through WSL, g++ -O2 -std=c++17:

This isolates the search pattern; it does not instantiate BlockDataInfo or measure an actual application allocation pause. It establishes the repeated scan cost, while a real grow/shrink benchmark is still needed to determine its contribution.

Track freed block IDs directly, or maintain a suitable next-free cursor, so reuse does not restart at zero each time. Preserve existing IDs for live blocks and verify the BlockIdType limits under default and big blocks. Keep native allocation-failure handling and the locking/collection lifecycle intact.

Test holes at the start, middle and end, mixed surviving groups, and repeated release/regrowth with verifier builds. Measure allocation throughput and tail latency after a temporary large heap. Related: #207 covers a separate quadratic metadata operation during release, and #151 covers benchmarking.