fix(ai): honor Anthropic cache format on Bedrock Converse - #55942

Issue for this PR

Related to #53756. This fixes the caching part only. The Claude reasoning-variant problem with inference profile ARNs is still open, so this PR does not close the issue.

Type of change

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

What does this PR do?

Bedrock Converse decides whether to emit cachePoint blocks only from the model ID (anthropic.claude-). An application inference profile ARN (arn:aws:bedrock:…:application-inference-profile/…) has no model name in it, so ARN-backed Claude models never cache. This was an intended trade-off in #47660, but it leaves no way to opt in.

Models already have a compatibility.cacheFormat override. The cache policy reads it and places the hints, but BedrockCache.block dropped them all because the ID check failed. Now BedrockCache.breakpoints takes the model and treats cacheFormat: "anthropic" as supported. This follows the same pattern as the supportsThinkingBlockBinding override in claude-model.ts. Models detected by ID are unaffected, and "alibaba" is ignored on Bedrock.

Users can opt in with:

"team-sonnet": {
  "modelID": "arn:aws:bedrock:us-east-1:123456789012:application-inference-profile/abc123",
  "compatibility": { "cacheFormat": "anthropic" }
}

Known limit: an opaque ARN can't be recognized as one of the 5-minute-only legacy Claude models, so 1h TTL stays allowed. The default policy sends no TTL, so this only matters if a caller explicitly asks for a TTL of 3600s or more on an ARN backed by an old model.

How did you verify your code works?

• New tests in bedrock-converse-cache.test.ts:
  • An ARN with the override gets cache points on tools, system and messages, with both the default and a 1h TTL.
  • "alibaba" emits none.
  • Both tests fail without the source change.
• The existing tests still cover an ARN without the override and non-Claude IDs getting no cache points.
• bun test on the Bedrock Converse, Mantle and cache-policy suites: 247 pass. bun typecheck passes in packages/ai.
• Not tested against a real application inference profile.

Screenshots / recordings

N/A

Checklist

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

Added by the opencode bot (AI-generated).

Fixes #55938

Related: #53756