feat: add kimi-k2.6 and harden structured output

[CollierKing]

Summary

• add @cf/moonshotai/kimi-k2.6 to the tested model matrices and Worker example supported-model list
• harden structured output handling by making method="json_schema" a first-class option, fixing prompt-value schema injection, and normalizing dict-shaped model responses before constructing AIMessage content
• remove the older llama vision variants from general test coverage and keep the integration model lists aligned with the current supported set
• harden Vectorize readiness and namespace search tests for eventual consistency
• make Worker test startup reliable on local machines by using non-interactive Wrangler startup and preferring a modern ~/.nvm Node runtime when the system node is too old for Wrangler 4
• bump langchain-cloudflare to 0.3.4 and update the changelog

Test plan

• cd libs/langchain-cloudflare && make test
• cd libs/langchain-cloudflare && set -a && source ../../.env && set +a && uv run pytest tests/integration_tests/test_workersai_models.py -v -s
• cd libs/langchain-cloudflare && set -a && source ../../.env && set +a && uv run pytest tests/integration_tests/test_worker_integration.py -v -s
• cd libs/langchain-cloudflare && make lint
• pre-commit run --all-files

Validation notes

• make test: 101 passed, 2 skipped
• Full Worker integration file: 145 passed, 6 skipped, 4 failed
• The remaining Worker failures were not K2.6-specific:
  • TestWorkerChat::test_chat_batch[@cf/mistralai/mistral-small-3.1-24b-instruct]
  • TestWorkerStructuredOutputBatch::test_structured_output_batch[@cf/openai/gpt-oss-20b]
  • TestWorkerMultiModal::test_multi_modal_image[@cf/mistralai/mistral-small-3.1-24b-instruct]
  • TestWorkerSessionAffinity::test_session_affinity_basic
• During one long REST full-file run, test_structured_output_json_schema_method_invoke[@cf/moonshotai/kimi-k2.6] failed once, but the isolated rerun passed immediately, so it appears flaky rather than a deterministic K2.6 capability gap.

[theobouwman]

I also received the OutputParserException with the Gemma model.

Installed this version of the package. And can confirm it works out of the box and directly fixes the issue 🚀

[theobouwman]

One questions. Does this significantly increase system message token usage compared to other models?

[CollierKing]

Yes, it does add tokens — in my testing ~3x more input tokens (145 vs 466 for a typical nested schema).

Here's why: by default I automatically inject the full JSON schema into a system message so Gemma knows exactly what fields to return. This is what makes it reliable.

If you want to reduce token usage you can pass method="json_mode" to with_structured_output. In this mode I don't inject anything — Gemma just knows it needs to return valid JSON, but has no idea what fields you want. That means you're responsible for communicating the structure yourself in your prompt, for example:

llm.with_structured_output(Data, method="json_mode").invoke(
    "Extract announcements and return JSON with fields: type, context, entities (name, ticker, role)"
)

Just note that this is essentially trading reliability for token efficiency — unreliable field coverage is the original problem this fix was solving.