nx_fhe.nx

fhe.nx -- fully homomorphic encryption scaffold. EFFICIENCY_ROADMAP §6.1. Compute on ciphertext; the party holding the key never sees plaintext, yet gets the right answer. Targets privacy-preserving analytics, untrusted compute substrates (cloud as "hostile arithmetic factory"). Schemes available today: BFV / BGV -- exact integer arithmetic, SIMD-friendly CKKS -- approximate real arithmetic, batching TFHE -- binary / bit-level, fast bootstrapping Each has its own complexity tradeoffs. Our scaffold exposes a uniform API (`fhe_add`, `fhe_mul`, `fhe_rotate`) + a scheme-selector enum; dispatchable backends land later. Full implementations are ENORMOUS: OpenFHE is ~200k LoC C++, SEAL is ~100k. Phase A here is API-locking + noise-budget accounting so library code that wants to "maybe run FHE someday" can be written in terms of the right abstractions now. Invariants: FHE1 All ciphertext operations are CONSTANT time in the plaintext (leaking neither value nor sign). FHE2 Every op consumes "noise budget"; fhe_budget_left reports remaining. Bootstrap refreshes (expensive). FHE3 Decrypt with the wrong key produces random garbage, NOT an error indication -- semantic security.