Qwen3.6-27B hits 218K, stabilizes tool calls

This is less about chasing one more benchmark number and more about crossing the line from "impressive demo" to "actually usable local agent runtime." On a 24GB consumer GPU, stability under long prefill and tool output matters as much as raw decode speed.

• –218K context on a single 3090 is the main headline; the throughput tradeoff is real, but the usable window is much larger than the earlier ~125K setup
• –The no-OOM 25K-token tool-output path is the practical win for coding agents, since failures there usually matter more than a few extra tokens/sec
• –The PN12 anchor-drift bug shows how brittle local inference stacks can be: patch application correctness can dominate model choice
• –The second memory cliff around 50-60K for single-prompt workloads means this is still a tuned edge case, not a universally smooth default
• –The result reinforces that consumer-GPU AI is now an optimization problem across model, quantization, KV layout, and serving stack, not just a model benchmark