Fix debugger segfault: VmFrame::ctxChain must not relocate its elements

[revarbat]

The crash

Setting a breakpoint inside a module and clicking Continue crashed instantly:

EXC_BAD_ACCESS (SIGSEGV) KERN_INVALID_ADDRESS at 0x0000000000000010
0  oscadeval::Evaluator::buildDebugFrame(EvalContext const*, bool) const + 120
1  oscadeval::Evaluator::buildDebugFrames(EvalContext const*) const + 224
9  oscadeval::Evaluator::checkDebug(...) + 1540
10 oscadeval::Evaluator::evalChildren(...) + 716
12 oscadeval::runCompiledModuleBody(...) + 1104
13 oscadeval::Evaluator::evalUserModule(...)

Root cause

Evaluator::enterUserCall stores CallStackFrame::bodyCtx as a pointer to frame->ctxChain.back(). That pointer is read much later, by any checkDebug() that walks the whole call stack to build per-frame debugger locals.

ctxChain was a std::vector<EvalContext>, so the next push_back reallocated and left bodyCtx pointing at the moved-from element — whose let_ shared_ptr is null. ctx->let_->items() then reads offset 0x10 from null. buildDebugFrame null-checks ctx but not ctx->let_.

The push that does it is Op::PushBuiltinWrap — every transform inside a compiled body. So the trigger is as ordinary as:

module mid(a) { translate([a,0,0]) inner(a); }

It only bites while debugging because a chunk compiles at all only when fastContinueBreakpoints_ is set (chunkEligibleNow) — exactly the "breakpoint set, now continue" state. The breakpointed callee is forced to interpret while its caller stays compiled, and pausing in the callee walks the caller's now-stale frame. The doc comment at bytecode_vm.cpp:276 describes that scenario for an earlier fix of the same class; this is the hole that fix left.

Fix

std::deque<EvalContext> — never invalidates references to existing elements on push_back/pop_back. lastCtx_ and childrenCallerCtx point into the same storage and are covered by the same change.

Cost

None measurable. 8 runs of eval_perf_benchmark (50,000 translate() calls — the same push path):

	min	median
vector	426ms	467ms
deque	432ms	451ms

Within noise both ways.

Verification

• New regression test drives the user's exact scenario (breakpoint forces the callee to interpret while the caller runs compiled, then reads the caller's locals from the debug frames). It segfaults without the fix.
• Full suite green on both engines: 1079/1079 with OSCAD_BYTECODE_VM=1 and =0.

🤖 Generated with Claude Code