perf(sdk): O(1) size-class guest allocator replacing first-fit linked list

[0xAndoroid]

Problem

The no_std guest allocator (linked_list_allocator first-fit, selected via ZeroOS's alloc-linked-list feature) walks its free list on every alloc and dealloc. In a zkVM every executed instruction is a proving-cost cycle, so this walk is pure trace-length overhead — and it grows with the number of live blocks.

Measured on a real workload (full stateless Ethereum mainnet block validation — reth/revm + MPT witness reveal over a 1.5 GiB heap, jeth): PC-sampled instruction profile attributed 85.7% of all executed instructions to the allocator:

58.62%  <linked_list_allocator::hole::HoleList>::allocate_first_fit
27.10%  <linked_list_allocator::hole::HoleList>::deallocate
 5.57%  <k256::arithmetic::projective::ProjectivePoint>::add   (actual workload…)

Swapping in this allocator cut total trace length ~6× (29.76B → 4.94B cycles for a 58M-gas block; 10.79B → 2.30B for a 31.8M-gas block), with identical guest outputs.

In-repo evidence (examples/btreemap, execute-only row counts): −8.9% rows at n=6000, −21% at n=20000 — the win scales with live-block count exactly as expected from the O(free-list) walk.

Design — jolt_platform::size_class_alloc

• Allocations round up to a power-of-two size class (min 8 B); each class keeps a singly-linked free list (freed blocks store next in word 0). Alloc pops the head or bumps the arena cursor; dealloc pushes. Every operation is a few dozen instructions; nothing searches.
• Bump blocks are class-size aligned, so any align ≤ size is satisfied; larger aligns round the class up. Free-list reuse re-checks the requested alignment (in-place-grown blocks only guarantee their original class alignment) and falls through to the bump path on mismatch.
• realloc grows the newest bump block in place (catches Vec doubling), and treats same-class or shrinking requests as pointer-stable no-ops.
• No splitting/coalescing: bounded internal fragmentation (<2× per live block) + per-class high-water free lists, deliberately traded for O(1) — appropriate for single-run guest heaps.
• Registered in __platform_bootstrap before kinit, for target_os = "none" guests only. std/musl guests keep the ZeroOS allocator (musl's malloc manages its own arenas over mmap/brk).

Testing

• Host-runnable unit tests (cargo nextest run -p jolt-platform -E 'test(size_class)'): class rounding, alignment-guarded reuse, LIFO recycling, in-place growth, copy-growth preserving contents, shrink stability, oversize/exhaustion nulls, alloc/free churn.
• btreemap-guest builds and executes to completion with matching (non-panicking) results on the emulator; the jeth workload additionally self-checks via its post-state-root assertion (any allocator bug ⇒ wrong state root ⇒ panic).

Note: GitHub Actions may be slow to schedule (ongoing incident) — happy to rebase/retrigger as needed.