Problem

Live Telegram direct replies can leak useless fallback prose like "Nothing further on my end" when the agent emits an exact NO_REPLY sentinel. This is worse than silence because it looks like an answer while hiding the actual tool/session failure.

Evidence

• Live transcript showed Telegram direct turns where SAB/Prowlarr auth/tool failures led to stale Plex fallback content or exact NO_REPLY.
• Default direct silent-reply rewrite was enabled, and direct-chat prompt guidance told the model to use NO_REPLY so OpenClaw could send a short fallback reply.

Acceptance

• Direct Telegram/WebChat prompts forbid NO_REPLY unless direct silence is explicitly allowed.
• Direct delivery never rewrites exact NO_REPLY into visible fallback chatter.
• Outbound/CLI delivery resolves silent-reply policy from the active session context.
• Regression tests cover direct prompt text, dispatcher behavior, outbound payload planning, outbound delivery, and reset transcript rotation.