[PD] Read nixl TransferInfo.is_dummy as a field in unit tests

[ShangmingCai]

Motivation

Important

Merge this only after #36612. TransferInfo.is_dummy is still a method on main until that PR lands, so this PR is red on its own and turns green the moment #36612 is in. Merging it first would break main in the opposite direction.

#36612 turns TransferInfo.is_dummy from a method into a bool dataclass field computed once in from_zmq(). It converts the two call sites that existed when it was branched (2026-08-27):

test/registered/unit/disaggregation/test_nixl_backend_basic.py:168,185

test_nixl_backend_basic.py has since grown five more, added by #34977 (2018c64e22, 2026-09-11) — after #36612's last push on 2026-09-09:

test/registered/unit/disaggregation/test_nixl_backend_basic.py:205,226,246,264,284

These are not textual conflicts, so git merge succeeds silently and base-a-test-cpu then goes red with:

TypeError: 'bool' object is not callable

Verified with a local trial merge of #36612 into current main: Automatic merge went well, then five is_dummy() call sites remain.

Modifications

Convert all seven info.is_dummy() assertions in test_nixl_backend_basic.py to field reads. No other change — the production is_dummy() call sites in nixl/conn.py are #36612's to convert.

All seven (not just the five #36612 misses) so the file is self-consistent whichever order the two land in; the two that overlap change to identical content, so the merge stays clean either way.

Scope check

I audited the rest of the disaggregation unit tests against #36612's other behavioral changes (the rewritten CommonKVManager.update_status policy, the deleted NIXL update_status override, and the new settle-then-raise barrier in transfer_worker). Nothing else breaks:

• TestNixlUpdateStatus — both cases (Failed is terminal; a missing room is not resurrected by Failed) are exactly what the new shared policy implements.
• TestNixlTransferWorker — builds TransferInfo by keyword without is_dummy, which picks up the new = False default and matches the old computed value; its check_xfer_state stub returns DONE, so _await_handles settles on the first pass and the worker never enters the new failure path.
• TestNixlNodeFailure, TestNixlNotifications, TestNixlReceiverPoll — unaffected.
• test_nixl_deferred_kv_release.py, test_receiver_connection_pool.py, test_decode_queue_cleanup.py, test_kv_transfer_replica_metric.py — all mock or stub update_status / failure_exception.
• test_nixl_sender_failure_cleanup.py — covers NixlKVSender.failure_exception; [PD] Share the prefill->decode failure notification across backends #36612 only changes the receiver's.
• test_disaggregation_wire.py — its is_dummy assertions are mooncake's, already a field.

Checklist

• Format check passes (whitespace-neutral: only () removed; every line stays well under the line limit)
• No production code touched

🤖 Generated with Claude Code

---

CI States

Latest PR Test (Base): ❌ Run #34615153783
Latest PR Test (Extra): ❌ Run #34615153400
Latest PR Test (AMD ROCm 10): ❌ Run #34615153703

[ShangmingCai]

/rerun-test test/registered/unit/disaggregation/test_nixl_backend_basic.py

[github-actions]

Results for /rerun-test test/registered/unit/disaggregation/test_nixl_backend_basic.py:

🚀 ubuntu-latest (1 test): ✅ View workflow run

cd test/ && python3 registered/unit/disaggregation/test_nixl_backend_basic.py