feat(hslm): [B] Trit Token Encoding — 729 tokens → 6 balanced trits

[gHashTag]

Parent: #329 — Fully Ternary Transformer

Level 0 (no dependencies) | ~100 LOC | New: src/hslm/trit_encoding.zig

Description

Encode tokens as balanced trits instead of dense embeddings. 729 = 3⁶ tokens → 6 balanced trits per token. TritEmbedding: 6×243 = 1,458 params vs 177K = 121x compression.

Implementation

• tokenToTrits(token: u32) → [6]i2 — base-3 decomposition into balanced {-1, 0, +1}
• tritsToToken(trits: [6]i2) → u32 — inverse
• TritEmbedding struct: 6 sub-embeddings of size 243 (3⁵), compose via bind
• Forward: token → 6 trits → lookup 6 sub-tables → bind together → output vector

tri CLI

tri hslm trit-encode

Tests

• Roundtrip all 729 tokens: tritsToToken(tokenToTrits(t)) == t
• Unique vectors: all 729 embeddings distinct
• Gradient flows through embedding lookup

Blocks

• Issue D (Ternary Activations)
• Issue J (Full Ternary Inference)

[gHashTag]

Consolidated into feat(hslm) epic #329 — Fully Ternary Transformer