Add proactive compaction via experimental.chat.messages.transform hook

Uses Morph CompactClient to compress older messages before they hit the
LLM context window. Intercepts OpenCode's message pipeline, serializes
older messages (text + tool outputs), and replaces them with a compact
summary. Results are cached by message IDs to avoid redundant API calls.
Preempts OpenCode's built-in auto-compact (95% context threshold).

Configurable via MORPH_COMPACT_* env vars (threshold, preserve_recent,
ratio, URL, timeout).