Use native OpenCode compaction lifecycle

[ishaanxgupta]

Summary

• replace Supermemory's independent token-threshold → session.summarize → synthetic Continue flow with OpenCode's experimental.session.compacting hook
• keep project-memory injection and post-compaction summary capture while bounding injected memory context
• add compactionEnabled and preserve compactionThreshold: 0 / false as legacy disable values
• stop the installer from disabling Oh My OpenCode's recovery hook now that Supermemory no longer owns the compaction trigger
• add focused coverage for stale summaries, transient failures, failed compactions, bounded context, and disable semantics

Root cause

The plugin independently calculated context usage from the session model, invoked summarization, and sent a delayed continuation resolved from the newest stored message. OpenCode can run summarization with a separately configured compaction agent/model, so the plugin could trigger against the wrong context limit, select the compaction agent for the continuation, and race OpenCode's own auto-compaction.

User impact

OpenCode now owns the trigger, compaction-model selection, and continuation. This prevents Supermemory from double-triggering compaction or leaving a session on the compaction agent. Supermemory only contributes bounded context to the native summary and captures the completed summary afterward.

Memory lookup or storage failures do not fail native compaction. Failed/overflowed compaction output is not saved as durable memory, and transient summary-storage failures retry on the next compacted/idle event.