[AHMETHAKANBEZIR1]

A traced scalar on the left of a kernel sum, such as offset + kernels.ExpSquared(...), raises TracerBoolConversionError under JIT or differentiation. The reflected-addition method tests offset == 0 in Python to support sum, before it builds the constant kernel.

Apply that shortcut only to builtin numeric zeros. Traced scalars follow the existing Constant/Sum path, while 0 + kernel, 0.0 + kernel, and sum([kernel]) keep their identity behavior.

Validation on Windows / Python 3.12.14 / JAX 0.12.0 / Equinox 0.13.8, CPU:

• On unchanged main: 3 failing regressions / 1 passing zero-sum control.
• New and existing base-kernel tests: 22 passed.
• Full test suite with default float32: 174 passed, 2 skipped; with JAX_ENABLE_X64=1: 174 passed, 2 skipped. Skips are the optional george/celerite comparisons.
• Source doctests, with the declared doctest extra: 3 passed.
• Configured pre-commit hooks and git diff --check pass; a news fragment is included.

Regressions check JIT, the analytic all-ones kernel derivative, vmap, and GP marginal likelihood/gradient against an independently formed dense covariance reference. No change to quasiseparable kernels' constant-addition restrictions is proposed.

Prepared autonomously with OpenAI Codex, including code and tests; no independent human review is claimed.