Severe performance regression after October 10 update: duplicate Context7/Draw.io MCP startup processes saturate CPU

Finding: A single external report (https://github.com/openai/codex/issues/52930) says that on macOS, after the Codex App update of October 10, 2026, starting Codex with Context7 and Draw.io MCP servers configured produced duplicate startup processes and sustained CPU saturation. The exact app build and macOS version are UNKNOWN. Reported symptoms: the reporter observed 3-4 concurrent Context7 MCP startup instances and 3-4 concurrent Draw.io MCP startup instances, roughly 6-8 concurrent npm exec processes in total. Each reportedly used about 100% of one CPU core, and overall machine CPU approached 100%. Some processes later exited, but the remaining ones reportedly kept consuming a full core each. No exact error message was provided. Restart behavior: the reporter says that fully quitting and restarting Codex immediately reproduced the problem, and that the new processes belonged to the new Codex parent process. The reporter takes this to rule out leftovers from the earlier Codex instance as the sole explanation. This is the reporter's inference from local process sampling, not independently verified. Cause status: the report states the root cause has not been confirmed. It does not establish whether the defect is in Codex, an MCP server, npm, or their interaction. The reporter suggests investigating MCP startup, retry behavior, overlapping startup attempts, and cleanup of superseded or failed attempts. This is a hypothesis for investigation, not a confirmed diagnosis. Maintainer, workaround and fix status: no maintainer response, workaround, fix commit, or release was present in the supplied data, so all are UNKNOWN. Closure or release status should not be inferred. Limitations: this is an external report only and was not reproduced by KFA. Raw sampling output, logs, exact commands and configuration, and version details are not attached. Only the initial head/tail issue timeline was supplied, so middle comments may be missing. Useful follow-up evidence would include the exact Codex build, the MCP configuration, process command lines, and whether the duplicates occur with only one MCP configured.