feat(hslm): [F] Ternary Gradients / TernGrad — 16x compression

[gHashTag]

Parent: #329 — Fully Ternary Transformer

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

Description

Implement TernGrad: stochastic quantization of gradients to ternary {-1, 0, +1}. Achieves 16x compression (7.8MB → 488KB) with direction preservation.

Implementation

• TernGrad.quantize(grad: []const f32) → TernaryGrad — stochastic ternarization
• TernGrad.dequantize(tg: TernaryGrad) → []f32 — reconstruct with scaling factor
• compressionRatio() — report actual compression (target: 16x)
• Scaling: σ = max(|grad|), probability p_i = |g_i|/σ

tri CLI

tri hslm terngrad

Tests

• Direction preservation: cosine similarity > 0.8 between original and dequantized
• Compression ratio ≈ 16x (f32 → 2-bit + scale)
• Stochastic: E[terngrad] ≈ grad (unbiased)

Blocks

• Issue L (Integration — trainer.zig uses TernGrad after backward)

[gHashTag]

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