v0.12.2

Billing impact

This release corrects underbilling and invalid symmetry discounts. Affected
calls can cost more than in 0.12.1; ordinary calls outside the corrected paths
retain their existing pricing.

• Unary ufuncs price the actual loop selected by dtype=, signature= or
  sig=, and wider out= destinations contribute independently. With
  production weights, four float32 values passed to
  exp(signature="d->d") cost 128 instead of 64; a complex128 loop costs
  396 instead of 64. modf and frexp follow the same rule; natural output
  destinations remain price-neutral.
• choose with a positional output and module-level random.shuffle now
  invalidate stale symmetry tags, removing discounts on asymmetric data.
• histogram2d and histogramdd apply per-axis bin-search pricing to supported
  scalar counts, count vectors, shared edges and mixed count/edge forms.
  histogram and histogram_bin_edges also charge edge construction, and
  histogram(density=True) charges its normalization arithmetic.
• Each executed length-one FFT stage bills 2 base FLOPs per batch entry for
  complex working-buffer movement, before dtype rates and operation weights.
  Ordinary butterfly formulas are unchanged.
• polyfit(cov=True) and cov="unscaled" bill additional covariance work,
  sharing matrix work across RHS columns. For eight float64 samples and degree
  three, production charges rise from 4,328 to 5,150 (scaled) or 5,148
  (unscaled). full=True still skips covariance.
• polyval, var/std with an enlarging mean=, trapezoid with N-D x,
  and isclose/allclose with array tolerances account for broadcast work.
• Periodic interp charges its preprocessing; matrix_rank charges broadcast
  tolerance comparisons; relocated vecdot/matvec/vecmat axes determine
  the contraction being priced.

Symmetry and correctness

• Caller-declared einsum(symmetry=...) results are canonicalized before
  tagging, matching as_symmetric: values within the validation tolerance are
  made exactly symmetric instead of preserving arbitrary asymmetric data.
  An exact S5 general-matmul exploit regression covers this boundary.
• Positional destinations share keyword out= write tracking, including
  aliases, while preserving deferred custom-ufunc writes on NumPy 2.4.
• Cumulative scans, reordered einsum outputs and vector quantile/percentile
  outputs retain symmetry on the correct axes. Quantile fixes include NaN
  variants and dimensions introduced by q.
• Unary dtype/signature constraints are frozen once for consistent billing and
  execution; invalid loops and incompatible output casts reject without charge.

Compatibility

• All three packages release together at 0.12.2. NumPy requirements remain
  >=2.0.0,<2.5.0; this release does not raise the NumPy minimum.
• FFT, histogram and polyfit changes preserve upfront budget enforcement.