Repository Layout

synapse/
├── bench/        # micro-benchmarks (kernel throughput / latency)
├── cmake/        # dependency pins, test/warning helpers
├── docs/         # design notes, assessments, plans, session records: the source of this site
├── docs-site/    # the Starlight site generated from docs/, published to GitHub Pages
├── scripts/      # developer and CI helper scripts, golden-fixture generators
├── sdk/          # the header-only library: sdk/include/branes/synapse/
├── tests/        # Catch2 regression tests, one executable per file; data/ holds golden fixtures
└── tools/        # host-side tools and reference applications (llama_decode)

The library

sdk/include/branes/synapse/
├── gguf/         # GGUF parser, memory mapping, tensor types, typed tensor views
├── quant/        # block-quantization schemes and their 8-bit activation blocks
├── kernels/      # fused quantized gemv / gemm, block dots, attention, element-wise operations
├── graph/        # graph builder, liveness planner, session (the one arena)
├── models/       # Llama-family text models and the CLIP / SigLIP image encoder
├── tokenizer/    # SentencePiece and byte-level BPE, and their Unicode tables
└── numeric/      # bit-exact widening of F16 / BF16 to float

Where the documents are

Folder	Contents
getting-started/	This introduction, building and testing, and this layout
arch/	The GGML ecosystem, and what a Llama model computes
assessments/	Analyses and trade studies
sessions/	Dated records of development sessions, on GitHub only
plan/	One plan per issue, published on the architecture site