May 2026 focused on strengthening the performance and maintainability of the attention path in the flash-linear-attention project. Delivered a substantial refactor of attention residuals handling that replaced the int64 pointer table with a padded tensor tuple, aligning with modern residual handling styles and enabling more efficient kernel data access patterns. End-to-end benchmarks show ~5% faster runtimes on typical shapes, with ~10% per-call latency reductions on small-N shapes, while preserving functional parity within a tight 0.005 tolerance on tests. Major improvements included cleanup of internal API naming and docs, restoration of symmetry between caller and kernel parameters (res_ptrs/dres_ptrs), renaming helper _pad_block_ptrs to _build_ptr_table, and removal of internal BL symbols from the fused_attnres docstring. These changes reduce technical debt and improve future extensibility. Overall impact: Higher performance, more reliable attention residuals path, and clearer API/documentation, enabling faster development cycles and more stable training/inference. Collaboration was strengthened through co-authorship with Yu Zhang. Commit includes e91af3c61de0ed8fd5604c9a1a40cd3c56ef7911.