fix(view): out= handling, scalar division, and broadcasting in structured views

[henryiii]

🤖 AI text below 🤖

Part of #1143.

Fixes three verified findings in the structured-view (view.py) ufunc machinery:

B2 — out= silently corrupted reductions

The "reduce" branch of __array_ufunc__ forwarded the caller's out= into _reduce, where np.add.reduce(raw[field], **kwargs) ran once per field with the same out buffer, so each field's reduction overwrote the previous one (e.g. np.sum(v, axis=0, out=np.empty(2)) on a 2-D Weight view returned [(4., 4.), (9., 9.)] instead of [(2., 4.), (3., 9.)]).

Now out= is popped before reducing: if the out dtype matches the view dtype, the result is computed fresh and copied into the out buffer (which is returned); otherwise a TypeError is raised. This applies uniformly to WeightedSumView and the mean views. _MeanArithmetic._reduce now also raises TypeError on reduction arguments it cannot honor (dtype, where, initial) instead of silently ignoring them; axis and keepdims keep working.

B11 — scalar / view produced meaningless statistics

2.0 / weighted_sum_view went through the scalar-first _scale branch and produced variance = s**2 / variance, which is statistically meaningless. The mean views already raise for exactly this case; WeightedSumView now raises a matching TypeError for divide/floor_divide with the view as the divisor. scalar * view and view / scalar are unchanged. (The existing test asserting the old 1 / v behavior was updated to expect the error.)

B12 — legal broadcasts failed

Binary ops preallocated np.empty(self.shape, self.dtype), so a (3,) view plus a (2, 3) array raised ValueError: non-broadcastable output operand. The result is now allocated with np.broadcast_shapes over all inputs; tested in both operand orders for + and *.

Cleanup (from D2)

np.divide is np.true_divide, so the redundant np.true_divide entries in the match alternation and the mean _scale tuple were dropped.

Testing

• Regression tests added to tests/test_views.py for each finding, including the exact issue reproductions.
• uv run pytest -n auto --benchmark-disable: 1006 passed, 1 xfailed.
• mypy (strict) clean; prek -a clean.

🤖 Generated with Claude Code