feat(core): metaserver flush barrier + embeddable P2P transfer service

[xiaguan]

What

Three additions that let an in-process Rust embedder (e.g. openinfer) serve cross-node RDMA fetches without running the full pegaflow-server binary:

1. MetaServerClient::flush() — a barrier over the fire-and-forget registration queue. Resolves once every insert/remove enqueued before the call has been delivered to the MetaServer or dropped after a failed attempt. Exposed as PegaEngine::flush_saves_and_registrations() (write-pipeline drain first, so block-seal registrations are enqueued before the barrier runs). This is the P/D handoff barrier: a prefill node flushes before signalling KV-ready, so a decode node's MetaServer query observes every registration that made it through; dropped registrations degrade to decode-side recompute.

2. P2pTransferService — the minimal Engine gRPC surface a serving peer must expose (RdmaHandshake / QueryBlocksForTransfer / ReleaseTransferLock + Health), with every CUDA-IPC registration RPC answered unimplemented. pegaflow-server keeps its full service; this is for embedders that drive PegaEngine in-process. Includes serve_with_incoming so embedders can bind first and fail loud on a taken port.

3. logging::init uses try_apply — RcBackend::new calls init_logging on first use; when pegaflow-core is embedded as a library the host process has already installed its own logger, and logforth's apply() panicked, taking down the host's engine loader thread. The host's logger receives our records; its existence is not an error.

Also bumps the cudarc floor to 0.19.7 and adapts to its regenerated sys bindings (CUmemAllocationHandleType is now a newtype over c_uint; the multicast handleTypes bitmask reads .0 instead of an enum cast, and cuda_sys.rs re-binds the handle-type constants as consts since associated constants cannot be pub used).

Tests

• 3 new fake-metaserver tests for the flush barrier: delivered-before-ack, ack-on-drop (permanent insert failure), no-hang after shutdown. cargo test -p pegaflow-core --features rdma --lib internode: 7/7 pass.
• End-to-end validated by the openinfer P/D disaggregation M2 run on a jiuzhang H200 node: 1 prefill + 1 decode openinfer instance, each embedding PegaEngine with this branch; decode discovered the prefill node's KV via the MetaServer and pulled 33/33 blocks (74.2 MiB) over RDMA READ; greedy output token-identical to the single-instance baseline across block-boundary prompt lengths; killing the MetaServer or the prefill node degrades to local prefill with no crash or hang.

🤖 Generated with Claude Code