Upstream contribution · morluto/leantoken
leantoken · Keeping LeanToken benchmark evidence bounded and interpretable
Problem
A blocking read from an MCP child could outlast the readiness budget. Nonfinite values in the paired benchmark threshold could also allow a significant slowdown through the regression gate. Either condition undermines repeatable comparisons.
Engineering response
The benchmark client now shares one absolute readiness deadline across retries and notifications, invalidates unusable streams, bounds frame size and wakes blocked callers on close. The paired evaluator rejects thresholds outside the finite float range. The npm launcher also forwards subsequent signals while its child remains alive.
Evidence and scope
The PRs describe executable peers that reproduce incomplete frames and notification floods, a CLI threshold regression with finite-value controls, and a child fixture that receives two signals through the launcher. The npm fixture used a Node child, and the checks establish behavior in these harnesses.