v0.8.10: frontend perf — batched rendering, DOM pruning, event delegation

Fixes 4 critical performance issues:

1. 🔴 innerHTML += → insertAdjacentHTML('beforeend', ...)
   Avoids full element re-parsing on every token. Tokens are buffered
   via requestAnimationFrame and rendered in batches — O(n²) -> O(n).

2. 🔴 No message cap → pruneMessages() at 100 DOM nodes
   Keeps the message container bounded. After 100 messages+tool blocks,
   oldest nodes are removed. Prevents unbounded memory growth.

3. 🔴 scrollBottom() per event → throttled via requestAnimationFrame
   Coalesces multiple scroll calls into one layout pass per frame.
   Eliminates forced-reflow chains from rapid token bursts.

4. 🔴 Per-item event listeners → event delegation
   Tool-block collapse and session-item click use a single delegated
   listener on the parent (#messages, #session-list, #completion).
   Zero listener allocation during streaming or session reloads.

Plus:
- Debounced @ completion (150ms instead of 10ms timer per keystroke)
- contain: content on #messages to isolate layout scope
- textContent for tool results (no HTML parse needed)

Author: jkyberneees