GeezerVison

Adding a method to Matrix/Vector/Scalar without updating the automethods
name sets left the expression classes silently missing it, and no test
failed. The new introspection test closes that:

- Forward: every public attribute (and value-forwarding dunder) of
  Scalar/Vector/Matrix/TransposedMatrix is either auto-computed on the
  expression classes or listed in an explicit OPT_OUT table with a reason
  (52 entries: mutators, constructors, native metadata, storage flags).
- Reverse: every generated name still exists on the concrete type.
- Hygiene: OPT_OUT entries must be live and not redundantly covered, so
  the table cannot rot.

Coverage is derived at runtime from what the generator emitted (getter
__module__ is graphblas.core.automethods), not from a second copy of the
name sets, so reorganizing the sets does not break the test. Failure
messages name the attribute and point to the sets and
scripts/autogenerate.py.

Teeth: adding a fake public method to Vector makes the test fail with
that message.

core/infixmethods.py copies the same generated automethods surface onto
the infix expression classes (VectorInfixExpr and friends, minus the
private _get_value), so a name missing only from the infix classes was a
gap test_automethods.py could not see: coverage was the UNION across
expression classes, so a name present on the plain expression class
masked its absence elsewhere. Coverage is now the INTERSECTION across
plain, index, and infix classes (identical sets today: Scalar 42/42/41,
Vector 49/49/48, Matrix 58/58/57, differing only by _get_value).

Teeth: hiding to_coo on MatrixInfixExpr alone makes the Matrix forward
test fail, naming the attribute, all three expression classes, and
scripts/autogenerate.py.

test_automethods.py guards the NAMES of the generated expression surface
but not the generated file content: hand-edit a block between the
auto-generated code markers, or edit the name sets in automethods._main
without rerunning the generator, and the on-disk blocks go stale while CI
stays green. scripts/autogenerate.py --check now regenerates every
generated file into a scratch tree and compares, naming the files that
drifted. The generator mains gained internal base-dir and callblack
parameters to support this.

The two guards are easy to confuse, so the scope is worth stating. The
generator emits from the literal name sets in automethods._main rather
than by introspecting the classes, so adding a method to Matrix does not
change its output and cannot surface here; test_automethods.py covers
that side. This covers the other one, generated files that no longer
match the generator claiming to produce them.

The check compares parsed syntax, not bytes. The generators shell out to
black when it is on PATH and skip it when it is not, so a byte comparison
reports drift on automethods.py and infixmethods.py in every environment
lacking black, and black is in no test extra. Guarding that with
skipif(black is None) meant the test never ran in the pytest_normal CI
jobs, which is exactly where it needs teeth. Comparing ASTs makes the
result independent of whether black is installed, so the check runs
everywhere, and layout is already enforced repo-wide by black in
pre-commit and the lint job. The residual gap is a layout-only edit
inside a generated block, which black --check catches and this does not.

The script also puts its own repo root on sys.path before importing
graphblas. For a script sys.path[0] is the script's own directory, so a
bare import resolves to whatever is installed; that coincides with the
checkout in an ordinary dev setup and diverges in a git worktree, where
--check would validate a tree nobody asked about while reporting green.
The check now prints the package it validated and the test asserts on
that line rather than inferring correctness from an exit code.

Scratch files go to a TemporaryDirectory outside the repo. Comparing
parsed syntax removed the reason to keep them inside it (letting black
discover the project pyproject.toml), and an interrupted run no longer
leaves .autogen_check_* directories in the working tree.

scripts/autogenerate.py --check regenerates into a scratch copy with
_main(base_dir=..., callblack=False), but the drift test runs it in a
subprocess, and CI's auto-generated code check runs the generators with
the defaults, so neither argument was ever covered. Copy the generated
files, make every generated block in the copies stale, run both
generators on the copies, and check that the copies are regenerated to
the same code, the package's files are untouched, and black never runs.
If a generator ignores base_dir and rewrites a package file, the test
fails and puts the file back byte for byte, so the checkout is left as
it was.

_fix_hint builds the assertion message of the expression-surface
guards, so it runs only when one of them fails.

"Let numpy's operators defer to graphblas objects" sets __array_priority__
on BaseType, BaseExpression and TransposedMatrix. The guard counts only a
class's own dunders, so TransposedMatrix is the one it flagged. The name is
a class constant numpy reads, like the networkx dispatch attributes already
in the machinery set, not a value of the result that could auto-compute,
and the expression classes define it themselves. In the machinery set it
also holds below that commit, where an OPT_OUT entry would fail the
liveness check for naming an attribute TransposedMatrix does not have.