Native capture, hydration, and startup calibration

The optional native hydrator writes an immutable snapshot blob directly into engine-owned final CUDA allocations. It is below the vLLM adapter boundary and therefore has no dependency on vLLM types. The engine-neutral library default remains python; the qualified vLLM process adapter selects direct.

The staged buffered and direct backends use an aligned pinned-host ring, bounded concurrent reads, cudaMemcpyAsync, and per-stage CUDA events. They can verify CRC32 or CRC32+SHA256 inline before publishing a restore. The gds backend uses a runtime-loaded cuFile API and writes directly to device memory; because those bytes do not pass through host memory, it requires an immutable artifact that has already passed full verification.

Build both native providers:

make native

Calibrate a bounded 512 MiB transfer on the local GPU:

.venv/bin/python benchmarks/harnesses/hydration_backends.py \
  --library build/native/libcoldsnap_hydration.so \
  --bytes 536870912 --chunk-bytes 67108864 --queue-depth 2

The current vLLM adapter selects the qualified direct backend and queue depth for capture and restore. An experimental engine/plugin launch outside that adapter can select the same runtime settings with COLDSNAP_HYDRATION_BACKEND and COLDSNAP_HYDRATION_QUEUE_DEPTH. The manager-driven request-JSON CLI is the operating surface.

COLDSNAP_HYDRATION_BACKEND=gds additionally requires COLDSNAP_HIBERNATE_VERIFY_MODE=preverified and its existing trust-marker admission. GDS availability is filesystem/device-specific; a present libcufile is only a dependency probe, not proof that a particular snapshot path supports GDS.

The same library exposes a separately versioned capture ABI. Its buffered and direct backends copy final CUDA allocations into a bounded aligned pinned-host ring, overlap device-to-host copies with positional file writes, and calculate requested per-extent CRC32/SHA-256 digests from the staged bytes. Every successful capture calls fdatasync; preverified capture additionally rereads the durable file and compares its digests before publication. Set COLDSNAP_CAPTURE_BACKEND=auto to prefer direct I/O for a fully aligned layout and otherwise retain the native pipeline with buffered I/O, including when the target filesystem rejects a direct open. Explicit direct remains fail-closed. python disables native capture without changing hydration selection.

Managed engine launches use native capture auto. vLLM applies it to the full live weight blob, while SGLang uses it through the engine-neutral semantic artifact writer. Both paths retain their existing Python implementation as an explicit compatibility fallback.

Fresh n580 restores resume at the context-free pre-exec boundary and select the ColdSnap loader before vLLM starts. The loader reconstructs the stable model layout from checkpoint metadata and lets the chosen native or recovery provider hydrate the final allocations; the retired pre-worker synthetic-loader bootstrap is not part of current artifacts.

The startup matrix keeps vLLM intact and compares full versus text-only model initialization in both in-process and ordinary multiprocess LLM() topologies:

.venv/bin/python benchmarks/harnesses/vllm_startup_matrix.py \
  --cases inproc-full inproc-text multiproc-full multiproc-text

For cache-hit measurements, bind writable persistent directories and set VLLM_CACHE_ROOT, TRITON_CACHE_DIR, and TORCHINDUCTOR_CACHE_DIR before the fresh process starts. The matrix records those paths in its output so a cache claim remains auditable.

Add serve-full and serve-text to measure HTTP readiness and streaming TTFT through the standard vLLM server.