perf: reduce scalar distance validation overhead

[isty2e]

Summary

• use math.isfinite after the existing float normalization in the scalar distance validator
• add direct contract tests for finite values, NumPy scalars, signed zero, non-finite values, and negative distances
• leave the distance model, workspace ownership, public API, and exception behavior unchanged

Performance

I ran three alternating parent/current smoke passes on the same machine, giving six samples per cell. Median zero-cost results:

Workload	Path	Before	After	Change
Continuous	Direct ask/tell	121.7 us/eval	94.5 us/eval	-22.4%
Continuous	Study scalar	149.2 us/eval	126.9 us/eval	-15.0%
Structured	Direct ask/tell	214.2 us/eval	180.0 us/eval	-16.0%
Structured	Study scalar	234.8 us/eval	198.4 us/eval	-15.5%

The objective-only and legacy control paths varied by about 3% or less. At the 200 us objective band, every variopt path in the paired comparison also improved.

A separate 900-run full matrix covered both workloads, five objective-cost bands, three seeds, and five repeats. All 150 variopt semantic records matched the parent baseline exactly, including result and state digests, evaluation counts, proposal counts, generations, and failures.

Verification

• uv run --extra test pytest: 1710 passed
• uv run --extra test pre-commit run --all-files: passed
• Ruff and BasedPyright: passed through pre-commit