[Core] Integrate HWR resilience changes for lifecycle validation

[baonudesifeizhai]

Purpose

Integrate the HWR follow-ups to #7107 into one draft for lifecycle fault-injection validation. Original commit authorship is preserved.

Source PR	Change
#7126	Escalate worker shutdown to SIGKILL and retain survivors for retry
#7128	Bound domain-initialization lock waits
#7131	Skip store-size scans when no total limit is configured
#7133	Add explicit quarantine cleanup
#7140	Add opt-in serialized build admission
#7144	Expose full-checksum validation in startup configuration
#7145	Add the retention diagnostic, including the tensor-size follow-up
#7147	Add registration bypass and progress logs

Resolved overlapping configuration/documentation changes and added coverage for combined validation and registration policies.

Defaults remain metadata validation and concurrent admission. Quarantine cleanup is explicit; serialized admission requires a compatible store policy and does not provide strict storage reservations. Registration bypass avoids the risky path; it does not establish or fix a driver root cause. The retention PR is diagnostic only.

Test Plan

vLLM Version: 0.28.0; Python 3.13.12, Torch 2.13.0+cu130, safetensors 0.8.0.

vLLM-Omni Commit: integration HEAD 688446d23af1dda5e1c5d40940b0ccca162252a7, based on ff6e906a25de1ca3122c0d0d5250fde5a0ddbc7b.

With project/test dependencies installed (GPU hidden for this CPU pass):

CUDA_VISIBLE_DEVICES='' PYTHONPATH="$PWD" python -m pytest \
  -o 'addopts=--strict-markers --strict-config' -q \
  tests/host_weight_runtime \
  tests/diffusion/test_multiproc_engine_concurrency.py \
  tests/diffusion/test_result_pump.py \
  tests/diffusion/model_loader/test_diffusers_loader.py \
  tests/config/test_omni_config.py \
  tests/entrypoints/test_async_omni_diffusion_config.py \
  tests/diffusion/offloader/test_distributed_layerwise_backend.py \
  tests/diffusion/offloader/test_cuda_host_registration.py

The addopts override removes the unavailable xdist option from this local environment; strict checks remain enabled.

Test Result

• 615 passed, 2 skipped. Two additional Helios tests failed with No CUDA GPUs are available; both fail identically on the base commit with GPUs hidden.
• 144/144 local lifecycle fault cases passed: 12 phase boundaries × 3 repeats × ext4/tmpfs × concurrent/serialized admission. Real non-root processes, BF16 payloads, files/flock/mmap, SIGSTOP/SIGKILL; four concurrent readers and four explicit recovery attempts per case. Verified tensor contents, producer count, active-reader protection and released FD/mmap/locks. These are CPU store tests, not multi-GPU inference or automatic service recovery.
• Retention diagnostic still shows ~30.67 MiB private-memory growth after one million get_tensor() calls at 64 elements; cached reuse grows ~0.02 MiB in the same environment.
• Applicable pre-commit hooks pass except mypy: 45 errors versus 51 on the base, with no new diagnostics after normalizing line numbers.

Real multi-GPU rank-failure/resource-release validation on this combined branch remains pending. Keep this PR in draft while completing that lifecycle campaign.