feat(ai): support server-side response compaction

[rekram1-node]

Summary

• expose OpenAI Responses contextManagement and lower it to context_management
• retain completed encrypted compaction items in durable provider state
• replay compaction items for same-model stateless continuation without response-only fields
• preserve compaction-only assistant outputs through Session history projection

Scope

• supports in-stream server-side compaction on normal Responses calls
• does not add the standalone /responses/compact operation

Testing

• bun typecheck in packages/ai
• bun run build in packages/ai
• bun test test/provider/openai-responses.test.ts test/provider/openai-compatible-responses.test.ts in packages/ai (101 pass)
• bun typecheck in packages/core
• bun test test/session-runner-message.test.ts in packages/core (22 pass)

[Enough1122]

AI code review — automated review for reference; please use your judgment.

• packages/ai/src/protocols/open-responses.ts:434 — compactionItems() silently filters out any persisted entry that fails Schema.is, so corrupted or future-schema compaction state disappears without a trace and later requests can fail upstream with confusing "missing item" errors — log (or collect into providerMetadata as a diagnostics field) whenever stored items are discarded during lowering.
• packages/ai/src/protocols/utils/openai-options.ts:38 — resolve() swallows providerOptions decode failures via Option.getOrElse(() => ({})), meaning a typo like compactThreshold: "100k" disables server-side compaction with zero feedback on a feature whose whole point is invisible context management — either propagate the decode error or emit a one-time warning naming the offending option.
• packages/core/src/session/runner/to-llm-message.ts:209 — An assistant message that carries ONLY compaction state (meaningful.length === 0, native defined) now survives as an empty-content message; the replay test covers non-empty content, so add a case asserting how lowering handles the truly-empty-content-plus-compaction shape end-to-end — if lowerMessages emits a message item with an empty content array the Responses API may reject the whole request.
• packages/ai/test/provider/openai-responses.test.ts:1553 — Good happy-path coverage for retaining and replaying compaction items, but the malformed-item guard in onOutputItemDone (missing id or non-string encrypted_content must surface eventError) has no test — one negative case would lock in the validation branch against accidental loosening.
• packages/ai/src/protocols/openai-responses.ts:139 — context_management entries always materialize a compact_threshold key even when undefined, relying on the serializer to strip undefined values before transport; using conditional spread (...(item.compactThreshold !== undefined ? { compact_threshold: item.compactThreshold } : {})) makes the intent explicit regardless of serialization behavior.

— AI code review (automated)