Summary

• implement the initial Python-runtime containment slice of the ownership contract defined in [None][docs] define Python native KV transfer ownership contract #16347
• add registry-owned receive lifecycle state with exact writer-cohort reconciliation, publication/cancellation gating, idempotent result routing, and separate logical/physical drain state
• retain sender/receiver request roots, NIXL operation state, descriptor state, bounce leases, registrations, and mappings until terminal evidence permits release
• make cancellation and executor/resource-manager teardown fail closed while a Python-native transfer owner is still live
• harden V1 block-pin retirement with transactional unpinBlocksById validation
• add focused lifecycle, race, cancellation, replay, bounce, and executor teardown coverage

Relationship to #16347

This PR is the first implementation follow-up to #16347 and is logically stacked on it. The #16347 head branch lives in the fork, so this PR targets main; until #16347 merges, GitHub will also show its three design-document changes in this PR.

Applicability

• applies only when transceiver_runtime="PYTHON" selects Python-native KvCacheTransceiverV2
• covers direct transfers and eligible bounce transfers; it is not gated by kv_cache_bounce_size_mb or a second ownership flag
• leaves the default C++ CacheTransceiver, its data path, and its wire protocol unchanged
• keeps generation-first attention-DP broadcast on the compatibility path and outside the ownership rollout
• disables receive bounce for PP fan-in until byte-accurate per-writer extents can be authorized before address publication
• preserves the [TRTLLM-13614][feat] Disaggregated KV-cache bounce transfer #15618 wire format, so mixed-version/protocol hardening is not claimed here

Safety and operational impact

The normal path gains lifecycle bookkeeping and exact-set validation but no additional device arena beyond an already-active bounce configuration. Under ambiguous transport failure, the implementation deliberately retains requests, KV allocations/pins, bounce slots, and transport mappings instead of risking reuse or use-after-free. That converts a potential memory-safety failure into visible capacity pressure or a teardown veto; bounded recovery requires the follow-up lease/protocol work described in #16347.

Validation

• iterative trtllm-review self-review through a final clean pass with no P0-P2 findings
• full changed-file pre-commit run --from-ref FETCH_HEAD --to-ref HEAD
• git diff --check
• Python bytecode compilation for all changed Python files
• focused dependency-isolated lifecycle tests: 29 passed
• focused executor ownership/scope checks: 7 passed
• focused post-rebase C++/Python cancellation interaction checks: 8 passed
• git range-diff confirmed all three patches were unchanged when rebased onto current main (d17ba0e4d)
• test-list stage mapping validation

Not run locally: the full pytest selection (the environment lacks transformers), compiled C++ unit tests, or GPU/NIXL/VMM/PP/ADP/multirank runtime tests. Those remain required before this draft is ready for merge.