fix(ci): treat max_size=0 as an uncapped pool in check_mempool_hygiene

[LeSingh1]

Problem

ci/tools/check_mempool_hygiene.py enforces the rule in cuda_core/tests/AGENTS.md: a test-created pool without max_size reserves an address-space window sized from installed device memory rather than from what the test allocates, and the whole suite shares one process.

It decides "capped" by looking only for the presence of max_size, never its value:

def_is_capped(node: ast.Call) ->bool:
# ``**kwargs`` (arg is None) may carry max_size; do not guess.returnany(kw.argisNoneorkw.arg=="max_size"forkwinnode.keywords)

But max_size=0 is the default, and CUmemPoolProps.maxSize == 0 asks the driver for its system-dependent size. From DeviceMemoryResourceOptions (and identically PinnedMemoryResourceOptions):

max_size : int, optional — Maximum pool size. When set to 0, defaults to a system-dependent value. (Default to 0)

So these create exactly the uncapped pool the check exists to reject, and every one of them passes today:

DeviceMemoryResource(dev, DeviceMemoryResourceOptions(max_size=0))
DeviceMemoryResource(dev, {"max_size": 0})
PinnedMemoryResource(PinnedMemoryResourceOptions(ipc_enabled=True, max_size=0))

Verified against violations_in() on main: all three return [], while the equivalent DeviceMemoryResourceOptions() is correctly reported.

This is the worst shape for a hole in a lint: max_size=0 is the natural thing to reach for when quieting the hook, it looks like a cap in review, and it reserves the full window anyway.

Fix

Inspect the value and report a literal 0.

• Non-literals are still accepted, unchanged. A named value is usually the suite-wide POOL_SIZE; the checker does not guess.
• **kwargs / **other still short-circuit to "capped", unchanged — including {"max_size": 0, **opts}, where the unpacking may override the literal. Not statically decidable, so not reported.
• The reported message names the actual problem (max_size=0, which is the uncapped default) instead of claiming max_size is missing. The omitted-max_size message is unchanged.

Also updated the corresponding rule in cuda_core/tests/AGENTS.md, which is what contributors read.

Effect on the tree

None: cuda_core/tests has no max_size=0 construction today, so the tightened rule leaves the suite clean. The existing test_the_live_test_suite_is_clean (which runs main([]) over the real tree) still passes.

Tests

ci/tools/tests/test_check_mempool_hygiene.py gains three UNCAPPED cases (kwarg, kwarg among other options, dict literal), two CAPPED regression guards (a non-zero literal, and {"max_size": 0, **opts}), and one assertion on the new message.

Verified with the index-safe swap (cp aside, git show upstream/main:<path> >, run, restore):

# with the fix
20 passed
# with ci/tools/check_mempool_hygiene.py restored from upstream/main
FAILED test_uncapped_pool_is_reported[zero-kwarg]
FAILED test_uncapped_pool_is_reported[zero-kwarg-among-others]
FAILED test_uncapped_pool_is_reported[zero-dict]
FAILED test_zero_cap_message_names_the_zero
4 failed, 16 passed

Full pytest ci/tools/tests: 62 passed. ruff check and ruff format --check clean on both changed files; python -m py_compile clean. No GPU is involved — this tool is pure AST analysis, so everything above was actually executed.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.