fix(compat): map non-stream messages←responses stop_reason

[ax128]

What

Non-streaming Anthropic /v1/messages requests forwarded to a Responses-API upstream returned stop_reason: "end_turn" even when the converted reply contained a tool_use block. Anthropic clients (Claude Code / SDK) key off stop_reason == "tool_use" to execute the tool and continue the turn — reporting end_turn silently breaks the tool loop in this path.

responses_response_to_messages_response (mapper.py) now derives stop_reason from the converted content blocks and the Responses object, mirroring the already-correct streaming path (compat_bridge._messages_stop_reason_from_response):

• tool_use block present → tool_use
• status=incomplete + incomplete_details.reason ∈ {max_output_tokens, max_tokens} → max_tokens
• otherwise → end_turn

Why a local helper (not reuse)

mapper cannot import compat_bridge — compat_bridge imports mapper, so the dependency must stay one-directional. A ~12-line local helper keeps that constraint; logic is identical to the streaming mapper.

Scope

• Only the non-streaming messages←responses stop_reason. No interface/shape change; text-only replies still report end_turn.
• Streaming path was already correct and is untouched.
• Out of scope (separately tracked): stop_sequence (upstream doesn't return the matched sequence), streaming input_tokens (unknown at message_start emit time), document/PDF forwarding (needs upstream support check), dead chat_response_to_messages_response.

How to test

pytest aegisgate/tests/test_openai_request_mapping.py

New regression tests (written test-first, watched fail then pass):

• ..._preserves_function_calls now also asserts stop_reason == "tool_use"
• ..._text_only_is_end_turn
• ..._maps_max_tokens

Validation performed

• pytest test_openai_request_mapping.py → 9 passed
• Related suites (streaming_router, response_sanitization_routes, sanitize_helpers, forwarding_renderers, forwarding_once_execution, payload_compat) → 159 passed, 3 skipped
• ruff check clean, mypy mapper.py clean
• Remote CI (3.10/3.12/3.13): pending — not yet observed

Risks / breaking changes

None expected. Behavior only changes for non-stream messages←responses replies that are tool calls or hit max_output_tokens — those were already incorrect.

🤖 Generated with Claude Code