feat(core): opt models into provider compaction

[R44VC0RP]

Issue for this PR

None — draft proposal for v2, building on #46431 and #46962.

Type of change

• Bug fix
• New feature
• Refactor / code improvement
• Documentation

What does this PR do?

Adds opt-in provider compaction to the existing Session compaction lifecycle. The default remains summary compaction.

{
  "$schema": "https://opencode.ai/config.json",
  "providers": {
    "openai": {
      "models": {
        "gpt-5.3-codex": { "compaction": "provider" }
      }
    }
  }
}

A V2 plugin can also set model.compaction = "provider" through its existing catalog transform; no new hook or separate plugin package is needed.

Core calls the existing LLMClient.compact, persists the complete replacement window, and replays it only for the matching catalog model and runtime provider/model/route. Other models use the readable summary. Forks, reverts, and exports retain their existing lifecycle; sanitized exports omit the replacement payload.

Design tradeoff for discussion: native compaction is followed by summary generation from the compacted window. That preserves the current UI and provides a portable fallback, but costs an extra call. The originating model resumes from the native window, not that summary.

This is standalone compaction only. It does not enable in-band context_management, change thresholds, or claim all models behind a capable route support the endpoint.

How did you verify your code works?

• Live synthetic conversation through the changed Core service: compact → persisted checkpoint → full replacement replay → correct recall on a real OpenAI Responses deployment. No real conversation data or credentials are included in the PR.
• 50 focused compaction/replay tests, plus targeted config, fork/revert, export-redaction, and client projection checks passed. Existing Core, Client, and AI compaction regression checks passed, with affected assertions rechecked after fixes.
• Repository pre-push typecheck passed (33 tasks); regenerated clients.
• Read-only review caught a runtime-alias compatibility gap, now guarded and covered. The live check caught undefined provider metadata during serialization, now normalized and covered.

Not verified: a 350k-token workload or long-running production sessions. Kept draft for API and product-policy discussion.

Screenshots / recordings

No visual component changes; existing compaction status and readable summary remain.

Checklist

• I have tested my changes locally
• I have not included unrelated changes in this PR