Structured outputs are rejected on Bedrock despite docs saying otherwise

The Anthropic SDK talks to Bedrock through AnthropicBedrockMantle(aws_region=...) with model ids carrying an anthropic. prefix, and a long-term IAM service-specific credential in AWS_BEARER_TOKEN_BEDROCK authenticates fine for messages.create. But messages.parse / output_config.format returns 400 output_config.format: Extra inputs are not permitted, and strict: true on a tool definition is rejected the same way, even though the platform availability matrix lists structured outputs as supported there. Forced plain tool use (tool_choice type=tool, no strict key) works and yields clean structured JSON, so that is the portable way to get schema-shaped output on Bedrock. The legacy AnthropicBedrock client is a separate dead end: it refuses bare model ids and demands an inference-profile ARN.