[cneuralnetwork]

Summary

• Add a 1,023-byte self-unpacking Linux x86-64 SmolLM2-135M engine.
• Supply the original tokenizer and safetensors as mmap-able regular files on fd 0 and fd 3, removing pathname opens from the ELF.
• Include a 1,149-byte descriptor-input variant that retains the checked tokenizer behavior, plus sources, packed binaries, and reproducible build/verification scripts.

Accuracy tradeoff

The 1,023-byte engine uses greedy longest-vocabulary-match tokenization. It matches all 300 reference generated IDs across the three answer-key prompts, but only 13/20 focused tokenizer cases and 236/1,480 texts in a separate deterministic corpus. It always generates 100 tokens and omits the final newline. This is an experimental loose variant, not a general exact replacement.

The 1,149-byte variant passes 20/20 focused tokenizer cases, 300/300 reference generated IDs, all three reference text outputs, and exits cleanly.

Verification

On relaydeck, from a clean checkout of current upstream main:

cd x86
./build_fd.sh
./verify_fd.sh ~/smol

The 1,023 and 1,149 byte figures count the complete packed ELFs. The model, tokenizer, and optional shell wrapper are outside the binary.