[initial-d]

Embedding export previously materialized a whole-table CPU copy and Python bytes object, while native XOR prediction looped over every byte in Python. Stage weights in bounded row slices with incremental SHA-256, and compute each block's identical XOR residual through C-backed Python integer operations. Format, source IDs/dtype, compression candidates and C queries are unchanged.

Validation:

• 31 tests pass with PyTorch, including noncontiguous float32/float16/bfloat16 sources, block-sized byte staging, streamed hashes, packed-file identity for Python/native backends, and XOR zeros/odd-width/special-bit cases. System Python passes with 17 optional cases skipped.
• The pinned official Engram module and draft factory each pass all nine full-output cases. Tiny random weights are interface fixtures, not trained-model evidence.
• All 27 baseline/optimized file pairs on nine existing trained SGNS tables are byte-identical and reconstruct every row through C. Median XOR export improves about 1.9×.
• A separate 64 MiB fixture repeats trained rows for process-memory measurement: median export drops from 4.121 s to 2.207 s; additional process peak RSS drops from 67.19 MiB to 2.70 MiB. This fixture is not representative compression evidence. Native no-dedup descriptors, dedup dictionaries, Python-builder memory and temporary disk-space limits remain documented.

Exact baseline/code commits, commands, source/library/data/artifact hashes, all timing samples and limitations are retained in docs/export_evidence_20261021/ and summarized in docs/streaming_export_20261021.md. No new data download or retraining. CPU only; no trained Engram, GPU, task-quality or query-speed claim.