fix: streaming message length overflow + Telegram error handling

[teknium1]

Summary

When a streamed response grows past Telegram's 4096-char limit (or Discord's 2000, etc.), editMessageText fails with MESSAGE_TOO_LONG and the stream consumer treats it as an edit failure — killing streaming entirely. Normal sends chunk long responses, but streaming didn't.

Fix: Stream consumer overflow splitting

The run() loop now checks accumulated text against adapter.MAX_MESSAGE_LENGTH before each edit. When the text exceeds the safe limit:

1. Finalizes the current message at a clean line break (or at the limit if no good break exists)
2. Resets _message_id so the next edit sends a fresh message
3. Continues streaming into the new message

Works for all platforms — reads the adapter's MAX_MESSAGE_LENGTH at runtime (Telegram 4096, Discord 2000, Slack 39000, etc.).

Safety net: Telegram adapter

If editMessageText still hits MESSAGE_TOO_LONG (e.g. markdown formatting expanding the text), the adapter truncates and returns success=True instead of killing the stream.

Files changed

• gateway/stream_consumer.py — Overflow split loop in run()
• gateway/platforms/telegram.py — MESSAGE_TOO_LONG handler in edit_message()

Test plan

• 1035 gateway + streaming tests pass