fix(core): correct get_neighbor_code +boundary detection and level-0 UB

[nikbott]

What (from the parallel adversarial audit)

The get_neighbor_code +-direction overflow guard (current | ~dim_mask) == UINT64_MAX only fires at full resolution, so a coarser cell at the +-domain boundary got a wrapped neighbour code instead of the UINT64_MAX sentinel:

• +x, or any axis at small max_level → a code sorting above every valid leaf (benign in balance, but wrong contract), and
• +y/+z at max_level=21 (3D) → wraps to the origin (the carry runs off the top of the 64-bit word), which can make balance() spuriously flag the wrapped-location cell.

Two audit agents disagreed on severity (one called it benign, one high); working the bit-arithmetic shows both are partly right — benign at DIC scales, a real wrap at the envelope edge. Fixed regardless, plus a companion UB.

Fix

• +dir guard → an at-maximum-aligned-coordinate test: (current & dim_mask) == (dim_mask & domain_mask) & ~(one_dilated-1). Returns the sentinel exactly at the boundary; unchanged for interior cells.
• Early level == 0 → sentinel (a root cell spans the whole domain): also removes a 1<<64 shift UB in one_dilated at the maximum level (CORE-2).
• Applied identically in omp / mpi / cuda.

Validation

• New [boundary] regression at max_level=21 (far-corner +x/+y/+z and level-0 → sentinel; interior still resolves). Fails on the old code.
• Full omp (18 cases) + mpi (np-driven) suites pass; CUDA byte-identical balance parity re-run on the A2000 — all fixtures pass (the fix touches balance and balance_ref identically, so parity is preserved).
• clang-format + full pre-commit clean.