AGG-E2E: end-to-end forged-aggregate exercise against the fixed rVM (before enabling aggregation)

[dendisuhubdy]

Required before any genesis enables aggregation (docs/aggregation.md, 'Before enabling aggregation'); needs a ≥64 GB machine. Also a Production-profile constraint differential.

[dendisuhubdy]

From the #62 review pass on rVM cross-iteration behaviour at Production with N ≥ 2 (circuits b9ffc39, read-only): no finding. Every per-proof quantity is bound per iteration: the challenger is re-zeroed, each iteration gets its own PoW, queries and challenges, and its public values are absorbed at slot 13+34·i. B is absorbed once at words 5–12. N is hinted once and serves as trip count, sponge word 4 and length. The only difference between Test and Production is num_queries (16/80) and pow_bits (4/20); the emitted code has no profile-dependent branch.

One test gap belongs here: the Production N ≥ 2 aggregate program has never been executed, not even in the emulator. Production N=1 is #[ignore]d, N=2 and N=3 exist only at Test, and their proving runs are #[ignore]d too. The loop body is emitted once, so the instruction stream is the same for every N. What stays unexercised is register pressure from 80 unrolled queries inside the loop, and the memory and timestamp bounds at tier 22/23. A Production N=2 emulator run (about 4M rows, no proving) closes this cheaply. Do that before aggregation is enabled, together with the forged-aggregate exercise. The latent loop-allocator hazard that the same pass found is #63.