Subprocess-brokered polyglot vs in-process — tradeoffs I hit building a 10-runtime host

[RicKanjilal]

I built Sym, a polyglot host that takes the opposite tradeoff from GraalVM: instead of embedding runtimes in-process, it launches each as a worker and brokers a neutral object model between them. Live objects stay put and travel as handles with distributed GC.

The obvious costs are per-call latency and no shared memory. What I got is ten runtimes (JVM, Node, CPython, PHP, Ruby, R, Perl, plus compiled Go/Rust blocks and C via FFI) in one protocol small enough to read in an evening — 80-cell test matrix, green in CI.

Two questions for people who've gone deeper than me: (1) where does the subprocess model break down that I haven't hit yet at my scale, and (2) how does Truffle handle object identity across languages — my handle table feels naive next to what you'd need for real interop?
https://github.com/RicKanjilal/sym.git