Disagg KV transfer hardening (rebased onto v1.3.0rc14)

[yifjiang]

Summary

Rebase of #13495 onto upstream main HEAD 3b7af1c21 (2026-04-29). Same 4 commits as the original PR, conflict-resolved against current main:

1. Harden disagg transceiver request lifetime
2. Document disagg KV transfer hardening follow-ups
3. Harden disagg transfer cleanup paths
4. Add NIXL transfer release cancellation hook

Conflict resolution

Two files conflicted vs current main; in both cases main's version was kept because it already addresses the same concern with newer / more refined logic:

• tensorrt_llm/_torch/pyexecutor/py_executor.py — main already has is_disagg_context_complete_state short-circuit + requests_finished_by_transfer stats tracking + force_terminate_for_partial_reuse flag and the error-budget / drain-queue fatal-error path. Took main's structure; the PR's _has_inflight_generation_transfers, _fail_closed_for_inflight_generation_transfer, and _can_terminate_request_now helpers were not folded in (main's coverage of those concerns is sufficient).
• cpp/tests/unit_tests/multi_gpu/cacheTransceiverTest.cpp — main has the newer addSequenceBatch API and *llmRequest deref. Took main's; the PR's addSequence + raw shared_ptr signature would no longer compile.

The C++ runtime fix in tensorrt_llm/_torch/disaggregation/nixl/_agent_py.py (def release(self): self.handle.release() with the Failed to release NIXL transfer handle log) is preserved.

Validation (in progress)

• Build TRT-LLM rc14 wheel from this branch (ARM + x86)
• Build dynamo-trtllm container image with this wheel
• Multihost disagg e2e (1P:1D, Qwen3-Coder-30B-A3B, dlcluster GB200)

🤖 Generated with Claude Code