fix: constrain PoweredExponential.power to (0, 2]

[thomaspinder]

Summary

PoweredExponential.power is now constrained to $(0, 2]$, where the kernel is positive definite in every dimension. Fixes #813.

 PoweredExponential.__init__(power)
-  self.power = power                         # any value; a float is not trained
+  self.power = _wrap_power(power)
+
+_wrap_power(power)
+  if not 0 < val(power) <= 2:  raise ValueError
+  if power is a parameter:     return power                 # user's choice
+  if power == 2.0:             return non_trainable(2.0)    # RBF; cannot sit on a sigmoid bound
+  return SigmoidBounded(power, low=0, high=2)               # trainable, stays in (0, 2)

Evidence

• Before:

  power=3.0 accepted; min eigenvalue -3.299
  power=1.5: type float | trainable array leaves: 2
  

  After:

  power=3.0 -> Expected `power` in (0, 2], where the powered exponential kernel is positive definite. Got 3.0.
  power=1.5: type SigmoidBounded | trainable array leaves: 3
  

• New tests in tests/test_kernels/test_stationary.py:
  • a large unconstrained step keeps the power in $(0, 2]$
  • power=2.0 is fixed
  • −1, 0, 2.5, 3 and PositiveReal(3.0) raise ValueError
  • a valid user parameter passes through unchanged
  • the Gram matrix is positive definite for powers 0.3–2.0
• The existing docstring test with power=2.0 still passes.
• uv run poe test: 3245 passed, 1 skipped. Lint, doctests and the CI docs build (-E -W) pass.

Merge Danger

Door: two-way

Blast Radius: small

• Fitted models change. A float power in $(0, 2)$ is now trained by fit. Before this change, fit did not train it, because it was a plain Python float. This makes it consistent with Periodic.period and RationalQuadratic.alpha, but a model that used to keep its initial power will now learn it. To keep the old behaviour, pass paramax.non_trainable(jnp.array(power)).
• Code can break. Code that read kernel.power as a float must now use val(kernel.power). A power outside $(0, 2]$ now raises an error instead of giving an invalid kernel.
• Nothing in the library or the example notebooks reads kernel.power.
• The changelog entry is under Fixed.

🤖 Generated with Claude Code

[github-actions]

📖 Docs preview: https://pr-816--endearing-crepe-c2d5fe.netlify.app

Smoke render — the expensive notebooks run with reduced budgets, so
figures are not publication fidelity. /render-mode.txt says smoke.