xvetu/neuroglyph: NeuroGlyph v0.4 — semantic interlingua on NSM primes

Abstract

Semantic interlingua over the 65 NSM primes (NSM Chart v15, Goddard & Wierzbicka), with PropBank roles, a 10-axis diacritic feature mask, and macro crystallisation. Codec — binary wire format and agent transport binary.py — NG-BIN v1, a LEB128 varint format over the glyph graph; order-preserving, no external dependencies transport.py — the ng.message JSON-RPC binding mcp_server.py — stdio Model Context Protocol server, two protocol eras, 7 tools specs: codec/BINARY-FORMAT.md, codec/TRANSPORT.md 146 assertions across test_codec / test_verify / test_binary / test_transport Measured on the 58 golden graphs, NG-BIN is 9.46x smaller than the equivalent schema-message JSON (the bar was >= 5x). Two identities, two questions canonical_hash is the meaning identity; binary_hash = sha256(pack(graph)) is the transport identity. A macro and its expansion share the first and differ in the second, which is correct. The public contour is English Corpus annotations, diagnostics and the encoder's rule labels are English. Russian is kept only as language data: prime exponents, molecule surface forms, and the phrases themselves. Honest limitations coverage 41/48 (85%); mean weighted reconstruction loss 0.002 five-language self-test: 29 phrases render and re-parse, but only 1 (3%) returns to the gold hash — reported as a growth target, not a result the details of the newer MCP revision come from its release notes, not from a first-hand reading of the specification