Fix long-horizon forecast evidence and filter parsing

[mattfaltyn]

Summary

• Keep the EWMA baseline available in background comparison and forecast evidence beyond 25 sessions; only the shorter-horizon challengers report unsupported.
• Reject malformed filter formatting before stripping currency, percent, commas, or whitespace so invalid inputs cannot silently become different thresholds.
• Add model, ledger, parser, and browser-interaction regressions.

Verification

• New regressions failed before the fixes and passed after them.
• scripts/verify-fast: 401 backend and 151 frontend tests passed.
• scripts/verify: Ruff, 401 backend tests, npm audit (0 vulnerabilities), 151 frontend tests, lint, API generation check, production build, and 27 Chromium tests passed.
• Independent R2 review: PASS.