The disk tier writes 4-6 full-KV boundary records per turn (measured 12 GB/90s in a live tool chain; ladder rungs move every turn so dedup never fires; 5 hits/312 misses acceptance) and bakes ~300 MB of SSM state into every record. The Python engine's design — 64-token chain-hashed delta blocks, separated SSM companion store (measured 20x on a 43.7k-token doc), partial-boundary rehydration across restarts, leaf-first forest eviction — is fully mapped with file:line in the speed-parity plan (vmlx-swift#383 head comment + local SPEED-PARITY-PLAN.md §A/§D-P1). Staging: (a) delta store + chain-walk restore behind a flag, (b) restart boundary index, (c) eviction, (d) flip default after the restart-resume matrix beats the snapshot tier everywhere. This also removes the main pressure source behind the allocator collapse fixed in #2572.
The disk tier writes 4-6 full-KV boundary records per turn (measured 12 GB/90s in a live tool chain; ladder rungs move every turn so dedup never fires; 5 hits/312 misses acceptance) and bakes ~300 MB of SSM state into every record. The Python engine's design — 64-token chain-hashed delta blocks, separated SSM companion store (measured 20x on a 43.7k-token doc), partial-boundary rehydration across restarts, leaf-first forest eviction — is fully mapped with file:line in the speed-parity plan (vmlx-swift#383 head comment + local SPEED-PARITY-PLAN.md §A/§D-P1). Staging: (a) delta store + chain-walk restore behind a flag, (b) restart boundary index, (c) eviction, (d) flip default after the restart-resume matrix beats the snapshot tier everywhere. This also removes the main pressure source behind the allocator collapse fixed in #2572.