feat(jit): poll the async abort flag at native loop back-edges (#410)

[InauguralPhysicist]

Hard timeouts (eigs_set_abort_flag) now hold at full speed — no EIGS_JIT_OFF required.

Root cause (narrower than the issue assumed)

The gap was only the from-zero thunk: a call-hot function's loop closes natively via the backward patch and never crosses an interpreted back-edge. The OSR tier already polled — its thunk stops before the loop's own JUMP_BACK and re-enters through the interpreted back-edge every iteration, which is where the #379 poll lives. Root-caused during this work; the EMBEDDING.md caveat overstated the gap.

The fix

• g_vm_abort_flag is never NULL now: unregistered points at a static always-zero int (eigs_set_abort_flag(NULL) maps back to it), so both tiers poll with a single unconditional deref. The interpreter's check simplifies accordingly.
• The emitter plants movabs &g_vm_abort_flag / mov (%rax),%rax / cmpl $0,(%rax) / jne →epilogue before every native back-edge. On set it bails to the interpreter at the JUMP_BACK (r13d still holds the op's offset, the standard pre-effect bail shape) — the interpreter re-runs the back-edge, consumes the flag, and raises the ordinary catchable aborted error. No error machinery in native code.
• Unset cost: one predicted-not-taken load+test per back-edge, same shape as the interpreter's.

Planted-fault validation

The new embed_smoke test warms a function past the call threshold then runs one long call (bounded at 200M adds so a regression fails loudly instead of hanging). It FAILS against pre-fix jit.c (the loop ran to the 100M iteration cap ignoring a 150ms timer) and passes after. It sits before the spawn/channel test deliberately — g_vm_multithreaded permanently gates JIT compiles off for the process, so a later placement would test the interpreter twice (verified: zero compiles happen after that test).

Gates

• make jit-smoke (with the new g_vm_abort_flag stub), make embed-smoke, tools/embed_stack_soak.sh
• Release suite 2584/2584; ASan+UBSan detect_leaks=1 2584/2584
• Bench n=5 medians within noise: dmg_shape 326.3→323.4 ms, idxset 39.6→38.8 ms
• EMBEDDING.md / eigs_embed.h caveats deleted

Closes #410

🤖 Generated with Claude Code