Fail fast when receiver topology is empty

[koriyoshi2041]

Summary

• fail fast when receiver assignment has non-empty buckets but no local receiver ranks
• add a CPU regression for the zero-receiver topology

Problem

_gen_h2d_buckets() intersects targeted ranks with each participant's local topology before calling _assign_receiver_ranks(). A participant with no selected local rank can therefore reach receiver assignment with non-empty global buckets and an empty local topology.

In that state, receiver_list is empty while flattened_buckets is not. The outer assignment loop never advances assigned_cnt, so the update hangs indefinitely.

An isolated CPU reproduction calling _assign_receiver_ranks([(7, "bucket")], {}, {"mlx5_0": {7}}) remained alive after 750 ms and had to be terminated.

Fix

Raise a clear ValueError before entering the assignment loop when no local receiver can be selected. Returning an empty assignment would risk turning the collective mismatch into a false-success update.

Testing

pytest tests/test_assign_receiver_ranks.py -q
7 passed in 0.93s

ruff check checkpoint_engine/ps.py tests/test_assign_receiver_ranks.py
All checks passed!

ruff format --check checkpoint_engine/ps.py tests/test_assign_receiver_ranks.py
2 files already formatted

git diff --check

[koriyoshi2041]

Closing after a deeper reachability audit. The update path returns non-target ranks before bucket generation, while gathered local topology includes each valid world rank, so an empty selected topology is not reachable for a valid non-empty target list. The helper can loop on a malformed direct input, but that does not justify the core-path claim in this PR. Thanks for the review bandwidth.