GeezerVison

Behavior fix. SuiteSparse:GraphBLAS 10.5.2 turned GxB_NONBLOCKING_GPU
and GxB_BLOCKING_GPU from GrB_Mode values of their own (7099 and 7098)
into #defines of GrB_NONBLOCKING and GrB_BLOCKING, and
python-suitesparse-graphblas built against it no longer has them.
Before, graphblas.ss (gb.ss.about, gb.ss.config, gb.ss.concat, and so
on, and `from graphblas import *`) failed to import with that build:

    AttributeError: cffi library 'suitesparse_graphblas._graphblas' has
    no function, constant or global variable named 'GxB_NONBLOCKING_GPU'

After, it imports, and gb.ss.about["mode"] reads "blocking" or
"nonblocking" as before. The GPU modes go into About._modes only when
the library has them, with setdefault, so a library that exposes them
as aliases cannot turn "nonblocking" into "nonblocking_gpu".
test_about now checks the mode against the one graphblas was
initialized with.

10.5.2 also checks that imported values are long enough, which older
versions got wrong (they compared the byte size of the values with the
number of entries). test_ss_import_export imported a 14 x 7 bitmap with
values for 7 x 7, which 10.5.2 rejects with InvalidValue and older
versions filled from past the end of the values; it now gives values
for the whole bitmap.

pandas 3.1.0 came out on 2026-10-22, on PyPI so far, and three
test_formatting.py tests fail with it, on main and here:

* test_sparse_vector_repr and test_sparse_matrix_repr: the pandas-based
  repr renderer appends its "..." row with ``df.loc["..."] = ...`` to
  integer columns, and pandas 3.1 warns ("Pandas4Warning: Setting an item
  of incompatible dtype is deprecated"), which ``filterwarnings = error``
  turns into a failure.
* test_autocompute: under pandas' default display.max_columns=0, pandas 3.1
  fits one column fewer on each side of a wide repr's "..." than 3.0 did,
  and the test pins 3.0's output.

conda-forge has pandas 3.1.0 now, so every CI draw of the newest pandas
hits these. #605 renders reprs without pandas, which makes both moot, so rather
than change the renderer it replaces, CI stays below pandas 3.1 until #605
is on main.

The pools keep "" for the newest release. A new LATEST_CEILINGS table caps
it when the pick is written out, so an unpinned pandas becomes
``pdver='<3.1'`` and the log's summary reads ``pd=latest<3.1``. Pins and
rules are unchanged; an unpinned draw now gets the newest 3.0.x. A MAINT note says when to remove the cap, and check_versions.sh
points to it.

Only test_and_build.yml installs pandas to run the tests. imports.yml (pip,
``.[default]``) and debug.yml (conda) also install the newest pandas, but
they only import graphblas, which pandas 3.1 does not affect.

Checked: ``--validate`` passes (160000 combos). Dry-run solves with the cap
land on pandas 3.0.6 for Python 3.15 on linux-64 and for 3.13 with numpy
2.2 on win-64, and on 2.3.3 for 3.11 with numpy 1.24 on osx-arm64. Locally,
the three tests fail with pandas 3.1.0 and pass with 3.0.6.