[codex] add GLM 5.2 and OpenAI-compatible endpoint mode

[CollierKing]

Summary

Adds two related Workers AI chat-model updates for langchain-cloudflare:

• Adds explicit support for @cf/zai-org/glm-5.2 across model behavior, examples, REST integration coverage, and Worker integration coverage.
• Adds endpoint_format="openai_compatible" to ChatCloudflareWorkersAI so REST callers can route chat requests through Cloudflare's OpenAI-compatible chat completions endpoint.

The default remains endpoint_format="workers_ai", which preserves existing /ai/run/{model} behavior. The OpenAI-compatible mode is REST-only; Worker bindings still use env.AI.run() and now raise clearly if that mode is requested with a binding.

Current Status

CI follow-up commits fixed the first runner failures and the separate Worker example dependency issue:

• 2d6a114 fixes mypy under CI's newer floating dependency resolution and newer langchain-tests model-override conformance.
• 54fef21 fixes the Worker example dependency layout that triggered the separate Dependabot security-update error on main.
• The Worker example no longer declares direct langchain in [project.dependencies]; pywrangler sync keeps only Pyodide-compatible dependencies, and scripts/setup_pyodide_deps.sh remains responsible for the known-working create_agent vendored stack.
• Worker create_agent integration tests now fail on 501 instead of skipping, so unavailable create_agent is no longer hidden.

AI Gateway update: the account now has the test-ai-gateway gateway configured. Targeted Worker binding and REST AI Gateway tests pass against it.

Why

Cloudflare's native Workers AI run endpoint and OpenAI-compatible chat completions endpoint can behave differently for multimodal and structured-output workflows. This exposes the endpoint choice directly without changing existing defaults.

GLM-5.2 is also now registered with model-specific parameter handling so it keeps supported OpenAI-compatible fields like max_tokens and tool_choice while stripping unsupported fields like top_k and repetition_penalty.

User Impact

Users can now instantiate:

ChatCloudflareWorkersAI(
    model="@cf/moonshotai/kimi-k2.6",
    endpoint_format="openai_compatible",
)

and continue using the existing default behavior without changes:

ChatCloudflareWorkersAI(
    model="@cf/moonshotai/kimi-k2.6",
    endpoint_format="workers_ai",
)

Validation

• make lint - passed (ruff check, ruff format --diff, mypy)
• make test - 110 passed, 2 skipped
• pre-commit run --all-files - passed
• uv run pytest --disable-socket --allow-unix-socket tests/unit_tests/test_chat_models.py -q - 57 passed, 2 skipped before the CI follow-up test was added
• uv run pytest tests/integration_tests/test_workersai_models.py::TestOpenAICompatibleEndpoint -v -s - 3 passed
• uv run pytest tests/integration_tests/test_worker_integration.py::TestWorkerAIGateway -v -s - 3 passed
• AI_GATEWAY=test-ai-gateway uv run pytest tests/integration_tests/test_workersai_models.py::TestAIGatewayHeaders -v -s - 3 passed
• cd examples/workers && uv run pywrangler sync --force - passed
• cd examples/workers && ./scripts/setup_pyodide_deps.sh - passed; installs the known-working LangChain 1.0 Worker stack
• Representative Worker create_agent coverage on Qwen - 3 passed:
  • TestWorkerAgentTools::test_agent_tools[@cf/qwen/qwen3-30b-a3b-fp8]
  • TestWorkerAgentStructuredOutput::test_agent_structured_output[@cf/qwen/qwen3-30b-a3b-fp8]
  • TestWorkerAgentStructuredJsonSchema::test_agent_structured_json_schema[@cf/qwen/qwen3-30b-a3b-fp8]
• Full REST integration: uv run pytest tests/integration_tests/test_workersai_models.py -v -s - 128 passed, 64 skipped, 1 failed
• Full Worker integration before test-ai-gateway was configured: uv run pytest tests/integration_tests/test_worker_integration.py -v -s - 158 passed, 6 skipped, 5 failed

Known Unrelated Failures

Full REST integration has the known Mistral structured-output batch failure where one result returns None.

Full Worker integration has known/environment-sensitive model-output failures:

• Mistral structured-output batch returned None
• gpt-oss-20b structured-output batch returned None

A targeted Worker rerun reproduced the Mistral and gpt-oss-20b structured-output batch failures while gpt-oss-120b passed, which points to model-output quality rather than this endpoint-routing change.