Disagg KV transfer hardening (rebased onto v1.3.0rc13)

[yifjiang]

Summary

Rebase of #13495 onto the v1.3.0rc13 release tag (b9ce4b69d, 2026-04-26). Same 4 commits as the original PR:

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

This rebase pairs with ai-dynamo/dynamo#8870 (which bumps dynamo's stock pin from rc11 → rc13 — first published rc tag with #12976's admission fix). Together they enable a dynamo image that has both PR #12976's admission fix (in rc13) and PR #13495's NIXL release cancellation hook (this branch), without needing to chase TRT-LLM main HEAD.

A separate version of this PR rebased onto current TRT-LLM main (3b7af1c21) is at #13655.

Conflict resolution

Only 1 file conflicted vs rc13: tensorrt_llm/_torch/pyexecutor/py_executor.py. Two conflict regions, both took --ours (rc13 base):

• rc13 already includes the is_disagg_context_complete_state short-circuit + requests_finished_by_transfer stats tracking + force_terminate_for_partial_reuse flag (these landed independently between rc11 and rc13). The PR's pass-only handling of the same state would lose the stats tracking, so taking rc13 is strictly more.
• The error-budget / drain-queue fatal-error path on rc13 supersedes the PR's _has_inflight_generation_transfers / _fail_closed_for_inflight_generation_transfer helpers, which were not folded in.

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 line) is preserved verbatim — verified by extracting from the built wheel.

Net diff vs rc13: +321 / -34 across 11 files.

Validation (in progress)

• Build TRT-LLM wheel from this branch (ARM in progress; x86 in progress)
• Build dynamo container image with the wheel + rc13 dynamo source
• Multihost disagg e2e (1P:1D, Qwen3-Coder-30B-A3B, dlcluster GB200)

🤖 Generated with Claude Code