test(core): consolidate multiprocessing child-process helpers

[rwgk]

Description

The VMM IPC and descriptor-leak tests independently implemented child execution and multiprocessing warmup. The first spawned process opens a persistent resource-tracker pipe; warming up after a file-descriptor baseline can therefore produce an order-dependent false leak failure. Child cleanup must also release IPC handles before fixture teardown.

Follow-up to the consolidation discussion on #3146.

Consolidate the common behavior in helpers/child_processes.py:

• run_child(target, *args) applies the sanitizer-aware timeout, kills surviving children when a join times out or is interrupted, checks the exit status, and closes the Process handles before returning.
• noop(*args) supplies one picklable target for warmup and argument-serialization tests.
• warmup_multiprocessing() starts the resource tracker before descriptor counting and caches only a successful warmup.

Reuse these helpers in the VMM IPC and descriptor-leak tests. The serialization-failure case still serializes the resource-owning object before encountering Irreducible; its child target never executes. Tests that communicate with a running child or inspect the Process object's lifetime keep explicit process management.

Add six spawn-based regression tests covering argument delivery, failed exits, hung-child termination, interrupted-join cleanup, warmup retry/caching, and descriptor stability. Add agent guidance explaining when to reuse the helpers and how to guard specialized process-management tests.

Validation

Fresh local build and validation on Linux with Python 3.12.3, CUDA 13.4, and an NVIDIA A10G:

QA suite	Passed	Skipped	Expected failures	Passing subtests
cuda-pathfinder	1,615	5	0	0
cuda-bindings, default stream mode	669	27	0	21
cuda-bindings, per-thread default stream mode	669	27	0	21
cuda-bindings Cython tests	9	0	0	0
cuda-core	3,998	281	3	223
cuda-core Cython tests	1	0	0	0
Total	6,961	340	3	265

• The full build and QA runbook both exited successfully.
• A separate fresh pytest process ran the VMM descriptor-lifetime regression first, followed by all six new helper tests: 7 passed. This checks the cold-start baseline directly.
• pre-commit run --all-files: all applicable checks passed.
• Skips cover unavailable hardware capabilities, multiple-GPU requirements, and optional dependencies.

Review And CI

Ready for human review. This PR remains a draft solely to retain manual control over triggering CI.

Checklist

• New or existing tests cover these changes.
• The documentation is up to date with these changes.

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

[rwgk]

@brandon-b-miller, this is the follow-on to PR #3146. Could you please make a review pass before I run the full CI?

At #3146 (review) @brandon-b-miller wrote:

Since we have a growing number of tests that hit this safety issue with multiprocessing it might be time to centralize the logic and put a note in the agent skill to reuse it.

We don’t yet have a repository-level test-authoring skill, so I added the helper-reuse guidance to AGENTS.md for this PR. I suggest we establish the skills foundation as another follow-up and keep this PR focused on the multiprocessing test helpers consolidation.

[brandon-b-miller]

LGTM