Summary

• infer likely vision-capable model IDs in onboard-custom for non-Azure custom providers
• default inferred vision models to input: ["text", "image"] instead of hardcoded text-only
• keep non-vision defaults unchanged
• add regression test for claude-sonnet-4-6 on non-Azure custom providers

Why

openclaw onboard --custom currently writes non-Azure models with input: ["text"] only, which silently disables image attachments for vision-capable models unless users manually edit config.

Fixes openclaw#51869

Validation

• corepack pnpm test src/commands/onboard-custom.test.ts