feat(hslm): [D] Ternary Activations — STE quantizer + integer matmul

[gHashTag]

Parent: #329 — Fully Ternary Transformer

Level 1 (needs B for int embed) | ~200 LOC | New: src/hslm/ternary_activations.zig

Description

Quantize activations to ternary {-1, 0, +1} with STE (Straight-Through Estimator) for gradients. Pure integer ternary matmul: Vec32i8 = 32 ops/cycle with widening i8→i16 multiply.

Implementation

• TernaryQuantizer: forward quantize (threshold-based), backward STE pass-through
• integerTernaryMatmul() — pure integer version (i8 activations × i2 weights)
• simdIntegerTernaryMatmul() — SIMD vectorized (Vec32i8 = 32 parallel multiply-accumulates)
• Widening multiply: i8 × i8 → i16, accumulate in i32

tri CLI

tri hslm quantize-act

Tests

• Quantizer roundtrip: quantize(dequantize(x)) preserves sign
• Integer matmul == float matmul (exact for ternary)
• SIMD result == scalar result
• Benchmark: integer vs float matmul throughput

Dependencies

• Needs B (Trit Encoding) for integer embedding input

Blocks

• Issue E (Ternary Attention)
• Issue J (Full Ternary Inference)

[gHashTag]

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