Complete record of the EAGLE3 debugging session:
- 4 bugs found and fixed (truncate+overwrite, cache accumulation,
aux normed vs pre-norm, position off-by-one).
- Final γ sweep numbers: matched=true everywhere, speedup=0.27x-0.95x.
- Per-slot acceptance analysis: d[0]≈10%, d[1..3] worse, d[5..7]
surprisingly recovers (target verify follows EAGLE hallucination).
- Root cause analysis: verify_cost grows ~linearly with γ+1 while
avg_accepted grows sub-linearly, so speedup < 1 across all γ.
Path forward: tree-based drafting (bigger lever) + faster batched
verify (flash-attention-2 with multi-query K/V sharing).
