fix(models): accept non-standard roles in AnthropicMessage so normalization can run

[Navapon]

What

Widens AnthropicMessage.role from Literal["user", "assistant"] to str in kiro/models_anthropic.py, so requests containing inline non-standard roles (e.g. role: "system" sent by newer Claude Code clients for SessionStart hook context) are accepted and handled by the gateway's existing normalization pipeline instead of being rejected with:

422 {"detail":[{"type":"literal_error","loc":["body","messages",N,"role"],"msg":"Input should be 'user' or 'assistant'", ...}]}

Fixes #190
Fixes #226

Why this approach

• The handling logic already exists — normalize_message_roles() in converters_core.py (added for issue BUG: API error calling from codex App #64) coerces any non-user/assistant role to user before ensure_alternating_roles(). The strict Literal on the Anthropic model rejects the request before that logic can run. This PR removes the gate rather than adding a new mechanism, handling the entire class of unknown-role issues ("system", "developer", future roles), not just one value.
• Consistency across both APIs (per CONTRIBUTING.md): the OpenAI path already does exactly this — ChatMessage.role is str (kiro/models_openai.py:78), and tests/unit/test_routes_openai.py::test_validates_invalid_role already asserts unknown roles pass validation. This brings the Anthropic path into parity. Streaming and non-streaming share the same request model, so both modes are covered by one change.
• Field-tested: we've been running this exact patch locally in front of Claude Code (Sonnet 5 / Opus 4.8) since it started sending inline system messages; the 422s disappeared and conversations flow normally.

Behavior change (intentional)

tests/unit/test_routes_anthropic.py::TestMessagesValidation::test_validates_invalid_role previously asserted unknown role → 422. That assertion enforced the bug, so it is replaced by test_accepts_non_standard_role (mirroring the existing OpenAI-side test of the same name and expectation).

Test coverage

• tests/unit/test_models_anthropic.py — new TestAnthropicMessageRoleAcceptance (5 tests): user/assistant regression guards, system role (primary repro), arbitrary unknown role, and a full AnthropicMessagesRequest with an inline system message mid-conversation (exact Claude Code shape).
• tests/unit/test_converters_anthropic.py — new end-to-end test test_inline_system_role_message_is_normalized: converts a request with an inline system message via anthropic_to_kiro() and asserts the Kiro history contains only userInputMessage/assistantResponseMessage entries and the system content survives as a user message.
• tests/unit/test_routes_anthropic.py — route-level tests: test_accepts_non_standard_role and test_accepts_inline_system_role_message (no 422 at /v1/messages).

pytest -q: 1698 passed (full unit + integration suite).

Related PRs

I'm aware several open PRs address the same 422 (#191, #195, #197, #199, #205, #206, #207, #210, #234...). This one aims to be the minimal, test-covered variant that reuses the gateway's existing normalization rather than introducing new hoisting/merging logic — and it updates the route test that previously locked in the strict behavior. The merge-into-system approach discussed in #190 (comment by @W0n9) is a reasonable semantic refinement that could layer on top of this change later; happy to close this in favor of whichever direction the maintainer prefers.

[cla-bot]

Thanks for the PR! 🎉

Before merge, we need a one-time CLA confirmation.
It confirms that you have the right to contribute this code and allow the project to use it.

Full CLA text:
https://github.com/jwadow/kiro-gateway/blob/main/CLA.md

Please reply once with:

I have read the CLA and I accept its terms

You need to write once, all further messages from me can be ignored.

[Navapon]

I have read the CLA and I accept its terms

[ankitcharolia]

@Navapon could you give a try to this gateway: https://github.com/ankitcharolia/kiro-gateway

It supports all necessary endpoints for OpenAI and Anthropic.