fix: signal Bedrock max_output_tokens truncation on Responses API

[jeremym-tanium]

Summary

When a /v1/responses request to a Bedrock model is truncated by max_output_tokens, Bifrost drops the OpenAI-canonical truncation signals on the wire response:

• response.status is left unset (OpenAI sets "incomplete")
• response.incomplete_details is left unset (OpenAI sets {"reason": "max_output_tokens"})
• The streaming terminal SSE event is response.completed (OpenAI emits response.incomplete)

Consumers using the OpenAI Responses-API contract therefore cannot detect Bedrock-side truncation. This is especially damaging for streaming tool calls, where the truncated bytes parse as valid-but-incomplete JSON (required fields silently missing or {}), so a downstream parser cannot fall back on a parse error.

Changes

Two narrow edits in core/providers/bedrock/responses.go:

• Non-streaming path (BedrockConverseResponse.ToBifrostResponsesResponse): after the existing bifrostResp.StopReason = &stopReason, set Status/IncompleteDetails from the mapped stop reason ("length" → incomplete + max_output_tokens; clean reasons → completed). Mirrors core/schemas/mux.go::responsesStatusFromChatFinishReason.
• Streaming path (FinalizeBedrockStream): the terminal chunk's hard-coded ResponsesStreamResponseTypeCompleted becomes conditional — when the mapped stop reason is "length", emit ResponsesStreamResponseTypeIncomplete and set the same fields on the embedded Response.

Value flow: Bedrock stopReason "max_tokens" → Bifrost finish reason "length" → OpenAI status "incomplete" + incomplete_details.reason "max_output_tokens". No new code paths or API surface; status/incomplete_details are existing optional fields, so consumers that ignore them are unaffected.

Changelog entry added to core/changelog.md.

Type of change

• Bug fix
• Feature
• Refactor
• Documentation
• Chore/CI

Affected areas

• Core (Go)
• Transports (HTTP)
• Providers/Integrations
• Plugins
• UI (React)
• Docs

How to test

cd core
go test ./providers/bedrock/ \
  -run 'TestBedrockStopReasonMappingResponsesPath|TestFinalizeBedrockStream' -count=1 -v

Covered by:

• TestBedrockStopReasonMappingResponsesPath — extended with Status/IncompleteDetails expectations per stop reason (truncation → incomplete, clean → completed, unmapped → unset)
• TestFinalizeBedrockStream_MaxTokensTruncation — terminal event is response.incomplete with the fields set
• TestFinalizeBedrockStream_CleanCompletionUnaffected — non-truncation stop reasons still produce response.completed

Breaking changes

• Yes
• No

status and incomplete_details are existing optional fields. The streaming terminal event type changes from response.completed to response.incomplete only when Bedrock reports max_tokens — which is the bug fix (consumers now see the OpenAI-canonical type).

Related issues

None currently tracks this. Happy to open one first if maintainers prefer.

Security considerations

None — no auth, secret, or PII surface changed.

Checklist

• I read the contributing guidelines and followed them
• I added/updated tests where appropriate
• I updated documentation where needed (no behavior outside the existing OpenAI contract)
• I verified Go builds and tests succeed (UI not touched)
• I verified the full CI pipeline locally (the provider suite needs live credentials; ran gofmt -s, goimports, go vet, go build ./..., and the bedrock tests above)

[jeremym-tanium]

Superseded by the upstream PR against the main repo: maximhq#4680. Closing this fork-internal review PR; the same branch backs the upstream one.