feat(sparse-inference): TernarySparseFfn -- dual sparsity via ternary weights

[simeon-kepp]

Summary

Builds on the LinearBitNet layer merged in #477 to integrate ternary weight quantization directly into the sparse FFN pipeline.

TernarySparseFfn is a drop-in alternative to SparseFfn that stores W1 and W2 as i8 ternary weights {-1, 0, +1} instead of f32. This adds a second, independent axis of sparsity:

Axis	Mechanism	Typical savings
Activation sparsity (existing)	LowRankPredictor skips inactive neurons	70 % of neurons not visited
Weight sparsity (new)	Zero weights inside each active neuron's dot product are skipped	50–60 % of MACs within active neurons

Combined at BitNet b1.58 realistic numbers: ~15 % of dense f32 multiply-accumulate operations remain.

What changed

• crates/ruvector-sparse-inference/src/sparse/ternary_ffn.rs — new TernarySparseFfn struct implementing the FeedForward trait, with from_f32() constructor, forward_sparse(), forward_dense(), and w1_sparsity()/w2_sparsity() accessors. 9 unit tests including a sparse-vs-dense agreement check.
• src/sparse/mod.rs — mod ternary_ffn; pub use ternary_ffn::TernarySparseFfn;
• src/lib.rs — TernarySparseFfn added to the crate-level re-export

Additive only

• SparseFfn, SparseInferenceEngine, and all existing APIs are completely unchanged
• No new dependencies — LinearBitNet is already in the crate from feat(ops): add LinearBitNet — ternary weight GEMV with zero-skip #477
• Feature-flag-free: TernarySparseFfn is always available once the crate compiles

Usage

use ruvector_sparse_inference::TernarySparseFfn;
use ruvector_sparse_inference::config::ActivationType;

// BitNet b1.58 recipe: threshold = mean(|W|)
let threshold = w1.iter().map(|w| w.abs()).sum::<f32>() / w1.len() as f32;

let ffn = TernarySparseFfn::from_f32(
    512, 2048, 512,
    &w1, &w2,
    threshold,
    None, None,
    ActivationType::Silu,
)?;

// active_neurons from your LowRankPredictor as usual
let output = ffn.forward_sparse(&input, &active_neurons)?;
println!("W1 sparsity: {:.1}%", ffn.w1_sparsity() * 100.0);