[Bugfix][Mooncake] Offload producer partial tails on request finish

[Dao007forever]

Summary

Offload a producer's final partial Mamba align tail when the request finishes before another allocation can trigger copy-on-write.

The core now finalizes the unchanged table source only when the request ends exactly at the registered boundary with no tokens in flight. The handoff is routed through KVConnectorBase_V1 and MultiConnector; Mooncake immediately pins the exact boundary-state block, queues metadata with its store-job ID, and releases the pin after every worker reports completion. The finished request itself does not remain retained.

Root cause

Partial-tail offloads were emitted only when a later allocation triggered CoW and moved the registered boundary state into a durable private block. A prefill-only producer that finished at that boundary never allocated again, so its final partial Mamba state was dropped and could not be persisted for later prefix hits.

Impact

Later requests can reuse a producer's final partial hybrid prefix from Mooncake Store even when the producer ends immediately after prefill. Safety checks consume the marker without emitting a handoff if tokens remain in flight or computation advanced past the registered boundary.

Duplicate-work check

This is not duplicate work:

• [Bugfix][Mooncake] Save exact Mamba boundary states #51358 handles exact Mamba boundary states emitted during the CoW path; it does not finalize a producer that finishes before CoW.
• [Bugfix] Stop mamba align prefill at the replay boundary #50409 changes local replay-boundary scheduling and does not implement connector finish-time handoff.
• test(prefix cache): pin two mamba "align" + EAGLE prefix-cache defects #52371 adds tests for separate EAGLE/Mamba prefix-cache defects and has no production fix.

No open PR was found for the head branch or for searches covering producer partial-tail finish-time offload.

Validation

Focused prefix-cache, Mooncake Store scheduler, and MultiConnector suites:

.venv/bin/python -m pytest \
  tests/v1/core/prefix_cache/test_partial_prefix_cache_hits.py \
  tests/v1/kv_connector/unit/test_mooncake_store_scheduler.py \
  tests/v1/kv_connector/unit/test_multi_connector.py \
  -q -k "not multi_example_connector_consistency"

Result: 103 passed, 1 skipped, 1 deselected. On macOS, only the repository global cleanup hook was disabled because the installed PyTorch build raises an unrelated MPS allocator assertion during teardown; all selected test bodies passed.

The four finish-time handoff tests also passed independently. Pre-commit passed on every changed file, including ruff, formatting, mypy, SPDX, and repository policy hooks.

The deselected MultiConnector test requires a runnable accelerator and model engine, which were unavailable in this environment.

Model evaluation

Not run. This conflict resolution was validated at the focused unit-test level; serving/model evaluation still requires a Mooncake deployment.

AI assistance

AI assistance was used to resolve conflicts, adapt the change to the current per-store-job pinning design, run validation, and prepare this PR. The human submitter must review every changed line and be able to understand and defend the change end-to-end.

[claude]

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.

[ywang96]

/ci run

[github-actions]

✅ Triggered Buildkite CI #84497 for commit c5eb44e4a3da.

[ZJY0516]

/ci run

[github-actions]

✅ Triggered Buildkite CI #84953 for commit 1ac894ea9d2b.

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @Dao007forever.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[zhewenl]

/ci run

[github-actions]

✅ Triggered Buildkite CI #86522 for commit ee994c2671c2.

[khluu]

/ci run

[github-actions]

✅ Triggered Buildkite CI #86674 for commit ca92162b1e9f.