[tcyjohn]

Part 3 of 3; depends on #19, which depends on #18. Merge in order: #18 → #19 → #20. This PR targets main, so its current diff includes the two prerequisite parts until they are merged.

Review only this part: EHP resource lifecycle changes.

When a rollout GPU is loaned to Hybrid training, its stopped serving process must leave the weight-publication receiver set. This change publishes to the actual serving receivers while preserving Hybrid ownership across ordinary publication. GPU return remains an explicit planner or sampling-supply operation; restored services install current weights before admission reopens.

Lending treats the requested GPU count as a target total. It selects complete same-host replicas using the existing TP × EP × PP × CP width, counts the rollout reserve in physical serving GPUs, and preserves existing Hybrid replicas, slot IDs and loan generation during expansion. Each newly selected owner must confirm exit before its GPUs enter training. Failed expansion restores and republishes only the new allocations. Runtime candidates respect the existing fixed-Core rank set and the actual rollout GPU slots; overlapping process replacements stop the previous owner before starting the next one.

All Core ranks participate in state capture and release at a shared training boundary. Workers install matching model, master parameters and Adam state through CPU/TCP live-state transfer before becoming eligible for a fixed batch. TP1 workers use the same distributed launch environment. Cold startup keeps its 180-second budget, and warm installation ACKs keep their 60-second budget. After the initial ACK, a ready Hybrid waits for a Core boundary or cancellation without a separate queue deadline. Boundary errors still propagate to the join operation.

Normal resource return and training completion wait for assigned Hybrid tasks to finish. Cancellation and failure stop the owned process, then retire unfinished inputs as cancelled after exit is confirmed. Unused snapshots are released collectively once pending loaders and activations have ended; manifests are removed after successful release. Core incorporates the sample statistics of the Hybrid gradients it actually consumed.

Validation covers complete-replica selection, incremental lending and partial rollback, process ownership, failed snapshot cleanup, task replay and activation timing. The source tree passed FYP job 1856870 on Qwen3-4B with four GPUs and preflight skipped: 50 updates, 200 unique complete GRPO groups, 800 trajectories, 12 successful Hybrid training tasks, and the same 20 held-out tasks at revisions 0, 20, 40 and 50. Final publication matched the actual TP2 + TP1 receiver set, all three physical-worker ACKs and live attestations, and the sender's measured 398 tensors. Audit, cleanup and Slurm exit checks passed. The consolidated commit has exactly the same file tree as the GPU-tested source commit 896a5ba. Its follow-up CI correction changes only test exception references: the asynchronous and Future timeout tests now use asyncio.TimeoutError and concurrent.futures.TimeoutError on Python 3.10. Production code is unchanged; all 69 tests in the four affected files pass on both Python 3.10.22 and Python 3.12.

The GPU run covers this single-node four-GPU workload. Incremental expansion and partial rollback have CPU coverage. Ordinary publication was observed with the Hybrid GPU still borrowed; subsequent training on both sides of that publication for the same active Hybrid was not exercised.

This PR builds on the fixed-batch/optimizer PR and preserves the existing fixed-Core execution model.

This three-part series supersedes #16 and #17.