EndpointMantle defaults to /openai/v1, but AWS documents mantle's base URL as /v1

[anandpathak]

bedrock.NewClient with EndpointMantle builds a base URL of
https://bedrock-mantle.{region}.api.aws/openai/v1/. AWS documents mantle's base URL as
/v1 — see the "Programmatic Access" table on the
Kimi K2.5 model card,
whose Chat Completions sample sets
OPENAI_BASE_URL="https://bedrock-mantle.<region>.api.aws/v1".

Sending a non-OpenAI model to the default fails with an error that names the model, which
points you at model access, IAM and region rather than at the path:

client, err := bedrock.NewClient(ctx, bedrock.Config{
	Endpoint:  bedrock.EndpointMantle,
	AWSRegion: "ap-southeast-3",
})
if err != nil {
	panic(err)
}

_, err = client.Chat.Completions.New(ctx, openai.ChatCompletionNewParams{
	Messages: []openai.ChatCompletionMessageParamUnion{openai.UserMessage("ping")},
	Model:    "moonshotai.kimi-k2.5",
})
fmt.Println(err)

POST "https://bedrock-mantle.ap-southeast-3.api.aws/openai/v1/chat/completions":
400 Bad Request {"code":"validation_error",
                 "message":"model `moonshotai.kimi-k2.5` isn't supported on this route",
                 "param":null,"type":"invalid_request_error"}

Setting BaseURL to the documented /v1 on an otherwise identical config fixes it. Same
client, endpoint, region, credentials and model — only the path differs:

Path	Result
/openai/v1/chat/completions	400 validation_error, "isn't supported on this route"
/v1/chat/completions	routes correctly, proceeds to auth
/openai/v1/models	404
/v1/models	401 — route exists, auth evaluated

The 404-vs-401 split on /models suggests /openai/v1 is a narrower route rather than the
general one.

/openai/v1 is correct for bedrock-runtime, so it looks like the suffix carried over
to mantle in bedrock/auth.go:

return fmt.Sprintf("https://bedrock-mantle.%s.api.aws/openai/v1/", region)

Related: openai/codex#21352 is the same
report against the Codex CLI's Bedrock Mantle provider, closed without a recorded
rationale. BerriAI/litellm#29463 asks
for GPT-5.x on mantle via /openai/v1/responses, which suggests /openai/v1 is a live
route for the OpenAI models.

That last point leaves me unsure which fix you'd want. If both routes exist — /openai/v1
for the openai.* family and /v1 for the wider Bedrock catalog — then the default is
right for some callers and wrong for others, and flipping it would break the first group.

• In that case the fix is probably documentation: say in bedrock/README.md which models each mantle route serves, and make the Config.BaseURL hint ("Use BaseURL if a deployment requires /v1") say when that is
• If /openai/v1 was never intended as the mantle default, it's a one-line change in defaultEndpointURL.

Let me know which way would be preferred? I will be happy to send PR.