Use native OpenCode compaction lifecycle

[ishaanxgupta]

Summary

Fixes #69, #85 and #88 for OpenCode V1. OpenCode 2 already worked this way.

The V1 plugin ran its own compaction: it estimated context usage, called session.summarize, then sent a synthetic "Continue" message. That caused four bugs:

• Premature compaction (Failed model-limit lookup falls back to 200K and prematurely compacts large-context models at ~160K #85, Preemptive compaction fires at ~16% of context instead of 80% (model-limit lookup falls back to 200K) #88). When the model-limit lookup failed or found nothing, the plugin assumed a 200K context window, so 1M-token models were compacted at about 16% full.
• Wrong agent afterwards (Compaction: post-summarize "Continue" runs on the compaction model; summarize context-window mismatch; races with native compaction (3 bugs + feature gap) #69). The "Continue" message picked its agent from the newest stored message, so sessions ended up on a different agent than the one they were using.
• Wrong model limit (Compaction: post-summarize "Continue" runs on the compaction model; summarize context-window mismatch; races with native compaction (3 bugs + feature gap) #69). The trigger used the main model's context size, but OpenCode summarizes with the compaction model, which can be smaller.
• Double compaction (Compaction: post-summarize "Continue" runs on the compaction model; summarize context-window mismatch; races with native compaction (3 bugs + feature gap) #69). It raced OpenCode's own auto-compaction.

Now OpenCode decides when to compact, which model summarizes, and how the session continues. The plugin hooks experimental.session.compacting to add bounded project memories to OpenCode's own compaction prompt. It then saves each successful summary as a memory, with a stable id per summary and the same 3 s timeout as other automatic saves. Failed or empty compactions are not saved, and a failed save retries on the next idle event.

• compactionEnabled turns this off for both generations. A legacy compactionThreshold of 0 or false now disables it too, instead of silently falling back to 0.8.
• The installer no longer disables Oh My OpenCode's recovery hook, since Supermemory no longer starts compactions of its own. The README is updated to match.

Testing

Real OpenCode V1 1.18.35 and 1.18.23 in a sandboxed home, with a mock model and a mock Supermemory API. The model's context size is unknown, which is the lookup-failure case from #85. One turn reports 170K tokens used, then OpenCode compacts the session.

	Before	After
After a 170K-token turn	plugin compacts and sends "Continue"	nothing, OpenCode decides
Agent after compaction	switched from build to general	stays on build
Project memories in OpenCode's summary request	no	yes
Summary saved as a memory	no	yes, once

• bun run typecheck, bun test (33 pass) and bun run build.

Testing on V1 needed #111. Without it, every published version since 2.0.15 fails to load in OpenCode V1, so merge #111 first.