fx_gate.nx

fx_gate.nx -- the gate for the Q16.16 fixed-point math 49 modules import. WHY: nx_gensota gen-3 worklist. This module exists for ONE reason -- bit-identical simulation across machines (rollback netcode); IEEE-754 cannot deliver it. Its header numbers four invariants FX1-FX4, and an invariant nothing checks is a wish. ★SCOPE DECLARED HONESTLY: determinism has an IN-PROCESS half (same input -> same bits, every call) and a CROSS-MACHINE half (same bits on every target). T9 proves the first. The second CANNOT be proven from one host and is NOT claimed here -- it would need the same vectors run on a second architecture and compared. Saying so beats a gate that implies coverage it lacks. ⚠OPEN QUESTION FILED, NOT GUESSED (see debt): FX2 says fx_mul and fx_div "round toward zero (arithmetic shift)" -- but an arithmetic right shift rounds toward NEGATIVE INFINITY on negatives while integer division truncates toward ZERO, so the two may DISAGREE on negative operands. Rather than pin a guessed direction, T5 asserts the property that is true either way and matters to callers: the result is within ONE ULP of the exact product. license_tier: ORIGINAL expect_exit: 0 No hw writes (Rule 26).