feat(core): schedule provider compaction automatically

[nexxeln]

Summary

• Enable automatic provider compaction at the existing safe Session boundaries when provider mode is selected. Add optional positive threshold, clamped to the model's usable input ceiling; omitted thresholds use that ceiling. Like local scheduling, automatic provider scheduling needs a positive catalog context limit.
• Keep compaction.auto: false authoritative for new automatic work while preserving installed native context. Wait for fresh primary usage after a native checkpoint before scheduling again.
• Recover classified automatic context overflow through existing local compaction over the original durable history (SessionHistory.load(..., "local")). Do not fall back for authentication, rate limits, cancellation, arbitrary failures, or manual native errors.
• Update configuration documentation and generated contracts.

Third layer of the Core provider-compaction stack.

Validation

• Core suite via bun run test: 5,235 passed, 35 skipped; Core/Schema/Protocol/Server/Client/SDK/TUI/Session UI/UI typechecks passed.
• Live end-to-end runs through an isolated V2 server (providers.openai.compaction: { mode: "provider", threshold: 10000 }, gpt-5.4-mini) with both an API-key connection and a ChatGPT credential: automatic compaction fired at the step boundary with reason: "auto" and an installed providerContext, the runner waited for a fresh usage anchor before compacting again, manual /compact worked on top of the installed window, and a fact from before the checkpoints was recalled afterwards. Server logs contained no warnings or errors.
• Default HTTP/SSE transport was exercised throughout; the opt-in WebSocket channel is unchanged by this layer.