feat(hslm): [C] Ternary Positional Encoding — multi-scale trit positions

[gHashTag]

Parent: #329 — Fully Ternary Transformer

Level 0 (no dependencies) | ~80 LOC | New: src/hslm/ternary_position.zig

Description

Replace sinusoidal/RoPE float PE with ternary positional encoding using 4 trit levels × 243 entries, multi-scale frequencies (1, 1/3, 1/9, 1/27).

Implementation

• TernaryPE struct: 4 frequency scales, each 243 entries
• positionToTrits4(pos: u32) → [4][dim]i2 — 4 trit vectors at different frequencies
• Multi-scale composition: bind(scale1, scale2, scale3, scale4)
• Precomputed tables for fast lookup

tri CLI

tri hslm trit-pe

Tests

• 81 unique position encodings (3⁴)
• Frequency scales correctly: scale_k has period 3^k
• Gradient flows through PE lookup

Blocks

• Issue J (Full Ternary Inference)

[gHashTag]

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