Profiling the post-memo stream put 59% in artifact.verify and 38% in reconstruction, so the parallel target is verification, not reconstruction: reconstruction is what the plaintext memo made sequential-friendly, and fanning it out would contend on that cache.

Reconstruction stays single-threaded; a bounded batch is hashed by workers walking disjoint strides, then emitted in order. Workers touch no store, no cache and no allocator, so there is no queue, lock or atomic — join supplies the ordering. libfossil 2.35s -> 1.37s wall, byte-identical output. closed check-in: 4b2fceeacd user: dmestas tags: frames-parallel