Bug Description

When messages.tts.auto is set to "always" and the agent produces a long reply, the TTS audio is sent as a Telegram voice note with the full text as caption. If the caption exceeds Telegram's limit (~1024 chars for voice captions), the entire message is silently dropped — no error is logged, and the user receives nothing.

Steps to Reproduce

1. Configure messages.tts.auto: "always" with a TTS provider (e.g., OpenAI)
2. Send the bot a message that triggers a long reply (>1024 chars)
3. The bot generates audio + attaches the full reply text as voice caption
4. Telegram rejects the message (caption too long for voice notes)
5. No error is logged, no fallback delivery occurs

Expected Behavior

• OpenClaw should catch the Telegram API error (caption too long) and fall back to sending text + audio separately
• At minimum, the error should be logged so operators can diagnose delivery failures
• Ideally: auto-truncate caption or split into voice note + separate text message

Actual Behavior

• Message is silently dropped
• No entry in gateway logs
• User receives no reply at all
• Only discoverable by manually checking session history vs. delivered messages

Environment

• OpenClaw version: 2026.1.30
• Channel: Telegram (polling mode)
• TTS provider: OpenAI (gpt-4o-mini-tts)
• streamMode: "partial"
• replyToMode: "first"

Workaround

Set messages.tts.auto: "inbound" instead of "always" to only attach audio when replying to inbound voice notes. This reduces the frequency but doesn't fully prevent the issue for long voice-reply captions.

Suggested Fix

1. Wrap Telegram sendVoice in error handling that catches caption-length errors
2. On failure: retry as separate sendVoice (no caption) + sendMessage (full text)
3. Log all delivery failures with context (message length, caption length, API error)