Invalid numeric CLI flags silently fall back instead of erroring

[royalpinto007]

Problem

evalgate compare --tolerance abc silently falls back to 0 instead of erroring. A typo in a CI flag can silently disable regression detection.

Where

src/cli/args.ts (numFlag): Number.isNaN(n) ? fallback : n swallows non-numeric input.

Acceptance

Add a failing test first asserting evalgate compare --tolerance abc exits non-zero (or errors) instead of silently using the fallback. Then fix.