[Force67]

Closes the field crash observed in long-running multiplayer sessions; the container fix lives in Force67/equilibrium#17 (submodule bump only here).

The crash

Thread 13 "recreation" received signal SIGSEGV
#0 tx::network::ZPacketQueue::ProcessOutgoingPackets
   at znet/z_packet_queues.cc:314
   const i32 packet_age = static_cast<i32>(now - packet.last_send_time);

The reliable-transport retry scan range-fors awaiting_ack_packets_ every
retry interval. Under minutes of gameplay traffic (constant reliable packets +
ack churn) that traversal faulted dereferencing a near-null chain node —
confirmed in core dumps from three separate crashed sessions, with the same
signature each time.

Root cause

LockFreeOrderedHashMap::begin() walked the bucket chains looking for the
first live node before the iterator's ebr::Guard announced. The skipped
deleted-but-not-yet-swept nodes are exactly what a concurrent
collect_garbage() pass unlinks, retires and frees, so the preamble could
step into freed memory and start the traversal on a garbage node.
order_begin() had the same gap on a smaller window.

The bump

vendor/equilibrium e68c2b8 → efb03b7 (Force67/equilibrium#17): announce the
epoch before the first dereference in begin()/order_begin() and keep the
guard pin continuous into the iterator. The container's existing test suite
passes under ASan with the change, and a begin()-under-churn stress
reproducer that faults in the preamble on the parent commit no longer does.

Follow-up

Force67/equilibrium#18 documents a separate, deeper iterator-protocol flaw
(reproducible under duplicate-key insert churn, structural rather than
lifetime-related) with reproducers; it needs its own root-cause pass and does
not block this bump, which fixes the unguarded preamble the production crash
traversed.

---

Updated 2026-09-19: what this bump now carries

Force67/equilibrium#17 was rebased onto devel5 to clear a conflict, so its
commit changed and this pointer moved with it. The pin here was e68c2b8, which
turned out to be 25 commits behind devel5, so this is no longer a one-commit
bump: it advances vendor/equilibrium to devel5 plus the iteration fix, 26
commits. The notable ones beyond the fix:

• 74e85ca Remove the STL and the C runtime from base
• 363b036 / dc1a10c / 3223dd4 security bug bash 1-3
• 52d4bea container value semantics, string search and std view interop
• 3c336fe compile the windows sources with a mingw toolchain
• f961dcb an Option<const char*> owns the string it is given

Verified locally against this exact pointer, in the configuration that consumes
equilibrium (-DZNET_USE_STL=OFF):

• the library and the samples build clean;
• handshake_integration and mp_integration pass;
• the stressed chaos-soak profile CI runs (drop/reorder/jitter 5/5/5, seeds 1-3)
  passes.

Two notes on merging:

• Land fix(base): announce the EBR epoch before begin()'s first node dereference equilibrium#17 first, then re-point this at the resulting devel5
  commit. Right now it names the PR branch's head, which a submodule fetch can
  only reach while that branch exists.
• The red checks here are not this PR. Every zetanet CI run has failed since
  2026-09-01 (including the two that were merged): the actions/checkout steps
  in all three workflows lack submodules: recursive, so no vendored dependency
  is present and the build dies on xxhash.h (which lives in vendor/lz4/lib).