Why this is hard

• Vector execution is harder than storage: algorithms, randomness, and approximation choices happen at runtime, not at ingest.
• ANN breaks naïve determinism: traversal order, pruning, and sampling introduce entropy that must be declared and bounded.
• Replay is non-trivial: deterministic replay demands bit-stability; ND replay demands statistical envelopes and provenance strong enough to prove consistency.
• Backend drift is silent: without comparison and divergence detection, approximate systems rot unnoticed.
• Budgets are hard constraints: latency/memory/probe limits must halt execution, not merely warn.