feat(hslm): [J] Full Ternary Inference Pipeline — 0 float ops

[gHashTag]

Parent: #329 — Fully Ternary Transformer

Level 2 (needs B, C, D, E) | ~150 LOC | New: src/hslm/ternary_inference.zig

Description

Complete inference pipeline with zero floating-point operations: trit embed → trit PE → quantize → 3 blocks (ternary attn + integer FFN + quantize) → argmax.

Implementation

• TernaryInferencePipeline struct: holds all ternary components
• forward(token: u32, pos: u32) → u32 — full pipeline, returns next token
• quantizeI32ToTernary(accum: []i32) → []i2 — inter-layer requantization
• Pipeline: trit_embed(token) → add trit_pe(pos) → for each block: ternary_attn → integer_ffn → requantize → argmax on final output

tri CLI

tri hslm inference --ternary

Tests

• Output is valid token (0..728)
• Zero float operations (compile-time check: no f32/f64 in pipeline)
• Benchmark: ternary inference vs float inference throughput

Dependencies

• B (Trit Encoding) — trit embedding
• C (Ternary PE) — positional encoding
• D (Ternary Activations) — quantizer + integer matmul
• E (Ternary Attention) — ternary scoring

Blocks

• Issue K (FPGA Verilog Export)

[gHashTag]

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