fix(agent): recreate large shared allocations on restore

[galletas1712]

Summary

• preserve large shared POSIX-exportable device allocations in anonymous pinned host memory owned by the workload process before CUDA checkpoint
• release the original exportable device backing before native CUDA checkpoint
• let CRIU capture the pinned host carrier as ordinary process memory; no carrier sidecar files are created
• after restore, create fresh exportable device backing, restore bytes with H2D, then unregister and unmap the host carrier
• transfer creator ranks in parallel and use bounded control timeouts
• force the path in test_cucheckpoint.py without changing the production threshold, and assert that no carrier files appear

This is a transparent r615 workaround for large shared allocations whose bytes restore but whose original post-restore POSIX export state is stale. It does not change the driver, disable FLA/FABRIC, require GMS, or retain a second device allocation. Small, non-shared, non-POSIX-exportable, and non-device allocations remain on the native CUDA checkpoint path.

Validation

• exact 578 MiB two-GPU symmetric-memory A/B:
  • PR feat(agent): checkpoint CUDA multicast state #133 without this workaround fails during coordinator restore when creator re-export returns CUDA_ERROR_INVALID_VALUE
  • this patch restores successfully and verifies CUDA-graph output and allocation bytes
• native CUDA checkpoint test builds the shim with HOST_CARRIER_THRESHOLD=1, exercises the carrier path, and rejects any cuinterposer-host-*.bin artifact
• GLM 5.2 / DeepSeek-V4-Flash-NVFP4 SGLang TP8/EP8 on eight B200 GPUs, context length 32768, with warm compile/JIT/FlashInfer caches:
  • source startup: 87.02 s
  • checkpoint: 290.891 s total; CRIU dump 262.580 s; CUDA checkpoint 27.632 s
  • active restore: 37.410 s total; CRIU 12.914 s; CUDA 24.291 s
  • standby restore: 40.165 s total; CRIU 12.921 s; CUDA 27.033 s
  • aggregate CRIU process artifact: about 100 GiB; no separate carrier artifact
  • coherent post-restore and post-promotion inference both returned 17 * 23 = 391
• clean CUDA-image build with -Wall -Wextra -Werror
• git diff --check
• python3 -m py_compile agent/cmd/cuinterpose/tests/test_cucheckpoint.py

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on base/target branches other than the default branch.

🗂️ Base branches to auto review (1)

• main

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: ASSERTIVE

Plan: Enterprise

Run ID: 8594f3c2-bc6d-4d22-abf7-539ca6a02bc0

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

_{Comment @coderabbitai help to get the list of available commands.}

[galletas1712]

Superseded by #218 (lifecycle: host carrier for every tracked creator allocation, restored through one pinned arena per process). The re-cut stack (#212 → #220) supersedes this PR. Its body is quoted verbatim in the replacement's Origin section, and each review thread here has a row in the replacement's "Review threads carried" table with what was done about it.