feat(hslm): [G] 3-Level Adaptive Sparsity — magnitude pruning

[gHashTag]

Parent: #329 — Fully Ternary Transformer

Level 0 (no dependencies) | ~100 LOC | New: src/hslm/adaptive_sparsity.zig

Description

Three sparsity levels: dense (0%), sparse (33%), ultra_sparse (66%). Per-layer sensitivity analysis for automatic level selection. Magnitude-based pruning.

Implementation

• SparsityLevel enum: dense (0%), sparse (33%), ultra_sparse (66%)
• applyMask(weights: []i2, level: SparsityLevel) → []i2 — magnitude pruning to target density
• analyzeSensitivity(layer: Layer) → SparsityLevel — auto per-layer: measure output variance change
• Mask is binary: keep or zero out

tri CLI

tri hslm sparsity

Tests

• Sparsity percentages: correct % of zeros after pruning
• Pruning: smallest magnitudes removed first
• Sensitivity: attention layers get less pruning than FFN

Blocks

• Issue L (Integration — trainer.zig applies sparsity after optimizer)

[gHashTag]

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