refactor(scoring): reuse the shared measurement mean

[drewstone]

RAG scoring and KB evaluation maintained copies of measurement averaging already defined in src/statistics.ts.
Use the existing mean in six callers and delete both copies and one redundant filter.
Empty inputs remain zero, and non-finite measurements remain excluded.
The published API does not change.

Validation:

• Lint, source and contract typechecks, build, and packed-package checks pass.
• The API record remains current: 995 exports across six entry points.
• Hosted Linux CI: 855 tests pass and 7 optional tests skip; network tests are enabled.
• Official GEPA and SkillOpt integrations pass both tests against the packed package.
• The local Linux activation suite passes all 9 tests with one worker and unchanged deadlines.
• All 55 claim-evidence tests pass in Docker with --init, which reaps orphaned processes.