[FEATURE] Gate server-side compaction through model metadata capabilities #5466

Human Overview

Whether a model can compact server-side, and whether its thinking is bound to history, should come from model metadata rather than a hardcoded list in AnthropicModel.

SDK Language

Both / Not language-specific

Problem Statement

#5401 gates native compaction on a private list of model IDs and a temporary AnthropicModelConfig.serverCompaction override. That has to be kept by hand, is specific to Anthropic, and doesn't let other providers declare the capability.

Proposed Solution

Once #5305 lands:

• Gate on model.metadata.capabilities.serverCompaction (server_compaction in Python) and capabilities.reasoning.historyBound.
• Move the model ID list into AnthropicModel.declareMetadata(), using the defaults from [FEATURE] Test native compaction on other providers that serve preserved-thinking Claude models #5465.
• Replace the temporary config key with the metadata config key, for example metadata: { capabilities: { serverCompaction: true } }. Keep the old key as a deprecated alias for one release.
• BedrockModel declares serverCompaction: false.

Done when:

• No compaction code reads the private ID list or the temporary key.
• Unit tests cover declared, overridden, and unknown capability values.

Use Case

A gateway user sets serverCompaction: false through model metadata, the same way they override any other capability.

Additional Context

• Depends on: docs(designs): model metadata and capabilities #5305, [FEATURE] Summarize through Anthropic native compaction #5401, [FEATURE] Test native compaction on other providers that serve preserved-thinking Claude models #5465
• Design: team/designs/0022-server-side-compaction-support.md (docs(designs): server-side compaction support #5242), open question "Ship Part 1 with the temporary override, or wait for docs(designs): model metadata and capabilities #5305?"