First, "I regularly have 10+ instances of my harness running or idling on my laptop" implicitly asserts that you need that many instances running at once.
You don't. You can have a single instance running that handles your sessions in parallel.
Second, "I sometimes run my harness on machines that are low-power, low-RAM single-board computers." presumes the whole agent has to run on the SBC itself. It doesn't. There are agents (including Pi, which makes this even sillier) that can run their tool calls in a separate process/host, meaning that the agent can live on your development machine while the tiny tool-call daemon runs on the SBC.
Finally, "If my harness was written in JS, this would be both annoying in terms of responsiveness and limiting in terms of how much work I can do in parallel." could only be written out of a complete lack of understanding about how harnesses are actually implemented.
Agent harnesses don't do very much. The limiting factor is not the harness itself, but in the filesystem (reads, writes, file searches, greps), the computational tool calls (compilers, linters, evaluating the program), and the LLM API itself.
Your arguments against Javascript appear to be based on fallacies, lack of understanding of technology, and strong emotions against JS rather than actual fact.