Binary notation: self-referencing links take 3 bytes, the spec says 2 #51

Summary

docs/BINARY-NOTATION-SPEC.md says a self-referencing link is encoded with its id written once:

Self-Referencing Link (5: 5)

Text notation: (5: 5 5) (10 bytes)
Binary encoding:

Type: 0x0F (SOURCE_IS_ID | TARGET_IS_ID | SELF_REF | HAS_ID)
ID: 0x05 (varint 5)
Total: 2 bytes (80% reduction)

Both encoders write 3 bytes instead: the type, the id, and then the shared source/target again.

Reproduction (at bdc7631)

import { BinaryNotation, HEADER_SIZE } from './js/src/protocol/binary-notation.js';
const body = BinaryNotation.encode({ id: 5, source: 5, target: 5 }).slice(HEADER_SIZE + 1);
console.log(Buffer.from(body).toString('hex')); // 0f0505, the spec says 0f05

The Rust encoder does the same: encode_link_body in rust/src/backends/binary_notation.rs writes the id when link.id.is_some(), then, for is_self_ref, writes source_id again.

Expected

Either the encoders and decoders follow the spec (when SELF_REF and HAS_ID are both set and the id equals the source, write the id only, saving one varint per point link — a store of 1000 points (i: i i) takes 4759 bytes today), or the spec's example is corrected to 0F 05 05, 3 bytes. Changing the encoding is a wire-format change, so the decoder and the spec version would need to move together.

A smaller point in the same example: (5: 5 5) is 8 bytes of text, not 10.

Context

Found while comparing this notation with the binary links notation of link-cli (link-foundation/link-cli#106, spec). The comparison and the scripts used are in research-notes.md §8.