Recurring SIGILL/segfault in JSC GC (CodeBlockSet::clearCurrentlyExecutingAndRemoveDeadCodeBlocks) — bun 1.3.14

Summary

Claude Code crashes recurrently with a null-pointer segfault inside JavaScriptCore's garbage collector (bundled bun 1.3.14). The crash happens both mid-session and on claude --resume. Bun's crash handler catches the SIGSEGV, prints the bun.report URL, then finalizes with ud2, so the OS records it as SIGILL.

Decoded crash (from bun.report)

Crash reason:  Segmentation fault at address 0x00000000
Runtime:       Bun v1.3.14 on linux x86_64_baseline [StandaloneExecutable]

#0  JSC::CodeBlockSet::clearCurrentlyExecutingAndRemoveDeadCodeBlocks()   CodeBlockSet.cpp:56
#1  JSC::Heap::runEndPhase()                                              Heap.cpp:1763
#2  JSC::Heap::runCurrentPhase()                                          Heap.cpp:1441
#3  JSC::Heap::collectInMutatorThread()                                   Heap.cpp:2073
#4  JSC::Heap::stopIfNecessarySlow()                                      Heap.cpp:2054

Full report: https://bun.report/1.3.14/B_1521eeddmgggkEuhogC4wvo4E+ypR41gjqF0mphqF4v6nqF63hwqFyt5nqF6zjynE+hph0Cgxgh1Cw1lqxCmg+2tEsk04wCy5/qpC6z0qpCy8wK0oxK6iyqpCA2AgigC

The fault is in the concurrent GC path. Some cores show the same finalizer trap on the HeapHelper (GC collector) thread instead of the main thread.

Environment

• Claude Code: 2.1.168
• Bundled bun: 1.3.14 (x86_64_baseline, StandaloneExecutable)
• OS: Ubuntu 24.04, kernel 6.8.0-124-generic
• CPU: AMD Ryzen 5 5600X (AVX2/FMA present — not a baseline/ISA issue)
• Install: ~/.local/lib/node_modules/@anthropic-ai/claude-code

Frequency

Recurring across many releases (2.1.101 → 2.1.168) over ~7 weeks — dozens of coredumps, every one SIGILL at the same finalizer address (0x2b05212, the ud2 after bun resets fatal signal handlers). Deterministic crash site; timing-dependent trigger (GC race), more likely on long/large sessions.

Workaround

export BUN_JSC_useConcurrentGC=0 (disables the concurrent collector that faults).

Ask

Looks like an upstream JSC GC bug in bun 1.3.14 (CodeBlockSet::clearCurrentlyExecutingAndRemoveDeadCodeBlocks during Heap::runEndPhase). Could a future Claude Code bundle a bun with this fixed, or set the concurrent-GC workaround by default on Linux?