Segfault in JSC MicrotaskQueue::drain (generator microtasks) on rapid PTY shrink — macOS aarch64, Bun 1.3.14 standalone #35010

What happened
opencode 1.18.4 (a Bun 1.3.14 standalone executable, NAPI module loaded) segfaults when its
host terminal is rapidly resized down to extreme sizes (~2 cols × 1 row). The crash is inside
the VM's microtask machinery, not application JS — the VM should surface a catchable error
(e.g. RangeError on stack exhaustion) rather than segfault.

Crash reports (two independent hits, same stack family)

• https://bun.report/1.3.14/Ma10d9b296mwGugogC________urh5kDunr21Cusw11C+/y34C+qy34C+5sy6Cu657H__urh5kD27p5tC+xt2yCA2K4/vshtD
• https://bun.report/1.3.14/Ma10d9b296mwGugogC_______2lkjlDurh5kDunr21Cusw11C+/y34C+qy34C+5sy6Cu657H__urh5kD27p5tC+xt2yCA2K4/ny3zD

Decoded (first report): Segmentation fault at 0x5084EEFFC —
llint_call_javascript → MicrotaskCall::tryCallWithArguments<JSGenerator…> →
callMicrotask (JSMicrotask.cpp:143) → runInternalMicrotask → MicrotaskQueue::drainImpl<true>
→ VM::drainMicrotasks → Bun::jsFunctionDrainMicrotaskQueue (BunProcess.cpp:3806).
Features: spawn, standalone_executable, workers_spawned, napi_module_register, process_dlopen.

Environment
macOS 15 (Darwin 25.4.0), Apple Silicon. The terminal host is xterm.js inside a Tauri app
(PTY via portable-pty); the process group receives SIGWINCH + TIOCSWINSZ per resize step,
~10–120 ms apart, down to 2×1. Process RSS was ~1.0 GB at the time (this instance's normal
baseline in that workspace — not a leak ramp).

Reproduction honesty
Hit twice in the real app (both reports above). We could NOT isolate a standalone repro:
headless PTY harnesses (single jump to 1×1, 1-col staircase drags at 20–100 ms cadence with
bounces, floor jitter, DSR/XTWINOPS query replies, SGR mouse-motion floods with out-of-range
coordinates) all survive. Whatever the missing ingredient is, it's timing/state-dependent in
the real xterm.js byte stream. Happy to run instrumented builds or gather more data.

Downstream mitigation (for anyone else hitting this)
Host-side: never report a grid below 20×5 to the PTY (real terminals have window minimums;
after flooring, the crash class is unreachable). Implemented in our terminal host.