Initialize Standard JSON device padding and sentinels

[ashkanvg]

Fixes #3

Root cause

The reported initcheck finding and the returned-result gap are two distinct uninitialized regions:

• depth_init_MathAPI reads packed 32-bit words from out_string_open_close_8_GPU. When the open/close count is not divisible by four, the allocation contains 1-3 tail bytes that removeCopy never writes.
• The returned pinned host layout reserves one overlapping slot for structural[last] and pair_pos[0], but that synthetic root/closing sentinel was not initialized until iterator construction.

The host result buffer is allocated only after depth_init_MathAPI and CUB sorting complete, so initializing that host buffer alone cannot resolve the earlier device read.

Fix

• Zero only the 1-3 device padding bytes before removeCopy and depth_init_MathAPI.
• Initialize structural[0] and the overlapping structural[last] / pair_pos[0] sentinel before returning.
• Preserve all allocations, copy destinations, pair-array offsets, kernels, and public layout.
• Avoid a full pinned-buffer memset.

Validation

Environment: NVIDIA A100 80GB PCIe (sm_80), driver 580.173.02, CUDA 13.3, Compute Sanitizer 2026.2.1.

• depth_init_MathAPI initcheck findings by tail padding: padding 0: 0 -> 0; padding 1: 1 -> 0; padding 2: 1 -> 0; padding 3: 1 -> 0.
• CUB internal tuple<uint32_t,uint8_t> padding diagnostics remain unchanged (3 x open_close_count) and are independent of the application tail/sentinel storage.
• Pre-iterator sample sentinels: baseline SLAST=0, P0=0; fixed SLAST=22389, P0=22389; overlap verified.
• Valid inputs: 80/80 successful; malformed padding-1/3 inputs: 40/40 rejected with zero false passes.
• Sample output unchanged: 22,390 elements and lang=en.
• Full twitter_large_record.json output unchanged: 61,072,730 elements and lang=en.
• Full real-data hashes unchanged: structural hash 2211520110405220824; 4,676,467 opening pairs; opening-pair hash 8222789647060382053.
• GPU SASS and resource hashes are identical.
• Performance, 5 warmups + 30 interleaved pairs on the 804 MiB dataset: baseline median 337.144 ms; fixed median 339.667 ms (+0.748%, within run-to-run noise).

The change is limited to five lines in cujson/parse_standard_json.cu.