[deaneeth]

Summary

Phase 5 of the TinyAttention engine (Phase 0-4 are on main via PR #1).

• NAIVE schedule (naive_ctrl): the standard 3-pass safe softmax, run block by block. Z and P tiles round-trip through HBM (HBM grows to 2048 words). The four passes reuse the tile core through new operations: OP_NZ, OP_NM, OP_NP, OP_NPV.
• K/V overlap (cfg_overlap, FLASH only): double-buffered K/V with a prefetch sequencer that loads tile kj+1 while tile kj computes.
• Performance baseline: sim/perf_baseline.json pins busy cycles for FLASH (with and without overlap) and NAIVE over N, causal, cfg_qrows and latency. The sequential FLASH values match the closed-form model exactly; overlap is checked never to be slower.

Verification

• bash sim/lint.sh: -Wall clean.
• bash sim/run_unit.sh: 15 suites pass.
• SEEDS=200 bash sim/run_e2e.sh: 3 tests pass. Bit-exact O for FLASH and NAIVE, with and without overlap, against the golden model, for every N, causal setting, block size and latency. HBM bytes, tile counts and K/V loads match traffic.py. The perf grid matches the baseline.
• Mutation check: a wrong alpha in the NAIVE O update fails the end-to-end test.

Notes

• Overlap is FLASH only. NAIVE runs sequentially.
• NAIVE's 2.7x slowdown versus FLASH at N = 32 is recorded but not yet broken down by pass.
• See docs/ARCHITECTURE.md section 10 for the design and measured numbers.