Fixes #1 (bus error on macOS / Apple Silicon): the >1MB model lived on the worker-thread stack — fine on Linux (8MB default), fatal on macOS (512KB). Models are now heap-allocated and worker threads request 8MB stacks.
New: birnpack-macos-arm64 binary, and CI now builds + regression-tests on Apple Silicon (16MB multi-threaded roundtrip).
Binaries below (all CI-built and byte-exact-tested): win64, linux-x64, macos-arm64.
Linux/Windows output is byte-identical to v0.1.0.