feat(twitch): truncate messages to 500-char limit instead of splitting

[PyRo1121]

Summary

• Replace split_message() with truncate_message() so Twitch responses are a single message capped at 500 characters, instead of flooding chat with multiple chunked messages
• When truncated, cuts at the last word boundary and appends … to indicate content was trimmed

Problem

Long LLM responses were split into multiple sequential Twitch chat messages (each up to 500 chars). This created chat spam — a single response could produce 3-5+ rapid-fire messages, disrupting the channel and making the bot obnoxious.

Solution

Single-message truncation with a clean word-boundary cut:

fn truncate_message(text: &str, max_len: usize) -> String {
    if text.len() <= max_len {
        return text.to_string();
    }
    let budget = max_len.saturating_sub("…".len());
    let candidate = &text[..text.floor_char_boundary(budget)];
    let cut = candidate
        .rfind(|c: char| c.is_whitespace())
        .unwrap_or(candidate.len());
    format!("{}…", candidate[..cut].trim_end())
}

• Uses floor_char_boundary() for safe UTF-8 handling (no panics on multi-byte characters)
• Falls back to hard cut if no word boundary found within budget

Scope

All 8 outbound code paths now use truncation:

Response Type	Before	After
Text	Multi-chunk split	Single truncated message
RichMessage	Multi-chunk split	Single truncated message
ThreadReply	Multi-chunk split	Single truncated message
File	Sent raw (could exceed limit)	Truncated
Ephemeral	Sent raw (could exceed limit)	Truncated
ScheduledMessage	Sent raw (could exceed limit)	Truncated
broadcast(Text)	Multi-chunk split	Single truncated message
broadcast(RichMessage)	Multi-chunk split	Single truncated message

Testing

• LSP diagnostics: zero errors
• The MAX_MESSAGE_LENGTH constant (500) is unchanged and used consistently across all paths