Preserve visited Q-values with tiny prior probabilities

[sylvesterkaczmarek]

Normalize the visited prior weights before multiplying by Q-values in _compute_mixed_value. Multiplying a floored tiny prior by a small Q-value first can underflow to zero, losing the visited action's contribution.

For raw value 0.5, Q-values [0.25, 1.0], visits [1, 0], and priors [0, 1], the existing float32 result is 0.25 instead of 0.375. This also affects completed Q-values used by the public search transform.

The one-expression fix preserves the prior floor, no-visits guard, mixing formula, and output dtypes. No API or search-state changes.

Validation

macOS arm64 CPU, Python 3.12.11, JAX 0.11.2, Chex 0.1.92:

• New regression module: 13 failures and 3 controls pass on unchanged 88f9205; all 16 pass after the fix.
• Float16/bfloat16/float32, positive/negative values, multiple visited actions, eager/JIT execution, no-visits/ordinary-prior controls, and the public completed-Q transform on a real Tree.
• Full suite: 40 passed from the _testing directory prescribed by test.sh, including existing search-tree fixtures.
• Built the source distribution and wheel; installed outside the checkout and reran the full suite: 40 passed.
• Flake8's repository gate, changed-file Ruff, source Pylint, new-test Pylint/Pyink, compilation, and git diff --check pass.

Pylint over all existing tests reports three line-length findings, identical on unchanged upstream; no existing lint rule or test was weakened. GPU/TPU execution and end-to-end agent training were not run.