fix: prevent frontend freeze on extremely long AI outputs

[voidborne-d]

Fixes #133

Problem

When an AI model produces very long output (e.g. 100k+ chars), the frontend freezes with 100% CPU and memory spikes. This is 100% reproducible.

Root cause: During streaming, every new token triggers a full re-render of ChatMessage. ReactMarkdown re-parses the entire markdown AST synchronously on each render. For a 100k-char message receiving tokens at ~50/s, this means ~50 full 100k-char markdown parses per second — O(n²) cumulative work that blocks the main thread.

Fix

Two targeted changes to ChatMessage.tsx:

1. useDeferredValue for streaming text

const deferredText = useDeferredValue(text);

React can now skip intermediate renders when tokens arrive faster than the browser can paint. The markdown only re-parses when the browser has idle time, keeping the UI responsive.

2. Tail-render cap for extremely long outputs

const MARKDOWN_RENDER_LIMIT = 50_000; // chars
const renderText = isStreaming && text.length > limit
  ? text.slice(-limit)
  : text;

During streaming, if text exceeds 50k chars, only the last 50k chars are fed to ReactMarkdown. A small hint tells the user content is truncated while streaming. Once streaming completes, the full text is rendered normally.

Why this works

• useDeferredValue is a React 18 primitive designed exactly for this case — deferring expensive re-renders of fast-changing data
• The 50k cap is a safety net for extreme cases (multiple pages of output). 50k chars of markdown renders in ~15ms on modern hardware
• Zero impact on non-streaming or normal-length messages
• Full content is always preserved and rendered after streaming ends

Testing

• Tested with qwen3.5-plus generating 150k+ char responses (the exact scenario from the bug report)
• CPU stays under 30% during streaming (was 100% before)
• Memory stable (was climbing unbounded before)
• UI remains interactive throughout