chore: unify credential passing pattern across all integrations #616

Summary

Credential handling across NemoClaw integrations uses inconsistent patterns that evolved organically as each integration was added at different times. This creates confusion about the intended architecture and makes security review harder.

Current state

Problems

1. NVIDIA_API_KEY gets extra file-write treatment that Discord/Slack don't. This is because OpenClaw's inference provider requires a specific JSON file, but it's not obvious why the patterns differ without reading the startup script.
2. TELEGRAM_BOT_TOKEN is not passed through onboard at all — only through the deploy path. If a user sets up Telegram after initial onboard, they have to recreate the sandbox or run the bridge manually.
3. No single place documents which credentials go where and why.

Proposed cleanup

• Standardize on env-var-at-creation for all user-provided secrets (NVIDIA, Discord, Slack, Telegram).
• Where OpenClaw requires a specific file format (auth-profiles.json), the startup script handles the translation — but this should be documented as the exception, not discovered by reading code.
• Add a credential inventory table to the architecture doc.
• Consider whether Telegram bridge should move inside the sandbox for consistency, or whether the host-side pattern is intentional and should be documented as such.

Context

This inconsistency became visible when #555/#588 made openclaw.json immutable, which broke Discord/Slack (#599, fixed in #601). Each fix was scoped narrowly to the immediate problem rather than unifying the pattern.