[BUG] CUDA 12 latest-deps wheel tests fail: pylibcudf/libcudf nightly ABI skew (undefined symbol read_parquet_column_chunk_bounds)

[ramakrishnap-nv]

Describe the bug

CUDA 12 wheel test jobs fail at import time with an undefined symbol error from pylibcudf, before any cuOpt code runs:

ImportError: /pyenv/versions/3.11.16/lib/python3.11/site-packages/pylibcudf/io/parquet_metadata.abi3.so:
  undefined symbol: _ZN4cudf2io32read_parquet_column_chunk_boundsESt4spanIKNS0_7parquet12FileMetaDataE...

demangled: cudf::io::read_parquet_column_chunk_bounds(...)

The import chain that triggers it is from pylibcudf.io.experimental.hybrid_scan import ....

Root cause

The installed libcudf and pylibcudf nightlies are skewed:

libcudf-cu12-26.10.0a291.post260821111824     <- built 11:18
pylibcudf-cu12-26.10.0a292.post260821162058   <- built 16:20, ~5h later

pylibcudf a292 was built against a libcudf that exports read_parquet_column_chunk_bounds; the resolver pinned libcudf a291, which does not. Both satisfy ==26.10.*, so pip is free to pick mismatched nightlies.

Observed in

Run 32495102132:

job	result
wheel-tests-cuopt / 12.2.2, 3.11, arm64, ubuntu22.04, a100, latest-deps	fail
wheel-tests-cuopt-server / 12.9.2, 3.14, amd64, ubuntu24.04, h100, latest-deps	fail
wheel-tests-cuopt / 12.9.2, 3.11, amd64, l4, oldest-deps	pass
wheel-tests-cuopt / 12.9.2, 3.14, amd64, h100, latest-deps	pass

Only latest-deps jobs are affected, which is consistent with nightly drift: oldest-deps resolves to an older, self-consistent pair.

Failure mode is a collection error, so it takes out the whole module -- one of the runs reported 0 passed, 1 failed, 20 errors.

Not cuOpt code

The failure happens while importing pylibcudf, before cuOpt is exercised. It is unrelated to whatever a given PR changes; it surfaced on an unrelated routing fix.

Suggested next steps

• Confirm whether this is transient nightly skew or needs a constraint pinning libcudf/pylibcudf to the same nightly build
• If the former, consider whether latest-deps jobs should pin matched RAPIDS nightly builds rather than ==26.10.* independently