v0.6.0: Batch Prefill + Adaptive Draft Length

Batch Prefill Optimization

• Layer-major traversal: each transformer layer loaded once for ALL prompt tokens
• Reduces SSD I/O from O(layers × tokens) to O(layers) during prefill
• Embedding tensor loaded once per prompt instead of per-token

Adaptive Draft Length

• EMA-based controller adjusts speculation depth K dynamically
• High acceptance (>70%) → increase K; Low acceptance (<40%) → decrease K
• --adaptive-draft CLI flag for run and serve
• 10-30% throughput improvement over static K

Stats

• 26 tests, all passing
• 7 files changed, 342 insertions

Full Changelog: v0.5.0...v0.6.0