Measure linear grouping for GC free-list rebuilds instead of comparison sorting #206

GlobalAllocator::createFreeList in src/hx/gc/Immix.cpp scans the blocks, builds mFreeBlocks, sorts the pointer array with std::sort and SmallestFreeFirst, then walks it again to build the size-class starting indices. Collect calls this before mutators resume.

SmallestFreeFirst compares only mMaxHoleSize. The current hole-size assignments derive from line counts shifted by IMMIX_LINE_BITS, so the key space is bounded by the block geometry and much smaller than the number of free blocks in a large fragmented heap. This suggests linear grouping/counting by hole size instead of comparison sorting.

This is a source-level performance candidate at checkout 03a4e8d. I have not measured the free-list phase's share of collection pauses or an alternative implementation. Establish that share before changing the algorithm.

Profile rebuild time separately on small heaps, large mostly-empty heaps and fragmented heaps with long-lived objects, under both default and big blocks. Record free-block count, hole-size distribution, allocation search cost, total pause distribution, retained heap and subsequent allocation throughput.

If this phase is material, compare bounded grouping that retains the current smallest-suitable-hole policy. Verify the key alignment/range for every supported configuration rather than assuming the default geometry. Reuse or bound any native scratch storage so the optimization does not replace sorting with an allocation spike during every collection.

The rebuilt order and mNextFreeBlockOfSize must remain consistent. Check GetFreeBlock ownership, pooled allocator reset and background zeroing, and preserve the policy of filling smaller available ranges first. Equal-sized blocks may also have locality effects worth measuring even though the current sort is not stable.

Require the normal, moving, generational, big-block and verifier stress coverage. Related: #151 for phase-aware performance tracking. This targets free-list rebuild work inside the pause, separate from #191's large-buffer lock scope and #201's remembered-set copying.