feat(language_model): add strict_mode option for structured output

[lelouar]

Summary

Adds an opt-out for the hardcoded "strict": true field that LanguageModel currently injects into response_format for structured-output calls.

Problem

Some OpenAI-compatible servers reject unknown fields in response_format via Pydantic's extra_forbidden validator. Notably, NVIDIA's TensorRT-LLM (trtllm-serve) returns:

BadRequestError: [{'type': 'extra_forbidden',
 'loc': ('body', 'response_format', 'strict'),
 'msg': 'Extra inputs are not permitted'}]

litellm.drop_params = True does not strip this field, so the request fails regardless of the strict value (True or False — the key itself is forbidden).

Today, users hitting such a server have no way to use LanguageModel with structured output.

Change

Add a strict_mode constructor argument to LanguageModel:

• Default: True — preserves current behavior, no breaking change.
• False — omits the strict field entirely from the response_format payload.

# Default (unchanged)
lm = synalinks.LanguageModel(model="openai/gpt-4o-mini")

# For TensorRT-LLM and similar strict-field-rejecting servers
lm = synalinks.LanguageModel(
    model="hosted_vllm/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-NVFP4",
    api_base="http://my-trtllm-server/v1",
    strict_mode=False,
)

Also wired through get_config() / from_config() for serialization roundtrip.

Scope

• Scope is limited to the strict field. No rename/alias for other servers that might use a different parameter name — that can be a separate change if needed.
• All 5 provider branches that set strict are covered uniformly.
• No env vars introduced; configuration stays in the constructor like all other LanguageModel options.

Tests

Added 5 unit tests in synalinks/src/language_models/language_model_test.py:

• test_strict_mode_default_true_hosted_vllm — verifies default still emits strict: true
• test_strict_mode_false_omits_strict_hosted_vllm
• test_strict_mode_false_omits_strict_openai (covers the inside-json_schema variant)
• test_strict_mode_false_omits_strict_ollama
• test_strict_mode_roundtrip_via_get_config

All 10 tests in the file pass (5 existing + 5 new).

Test plan

• uv run pytest synalinks/src/language_models/language_model_test.py -v passes (10/10)
• Default behavior preserved (existing tests unchanged)
• Verified against a live TensorRT-LLM server (trtllm-serve with Nemotron-Nano) where strict_mode=False unblocks structured output calls

🤖 Generated with Claude Code

[YoanSallami]

Are you sure that non strict mode is not the json mode ? which doesn't guarantee the format ?

[lelouar]

It's seems to depend on model + LLM server + parameters. I can dig more into this to draw a clearer view. I only tried vLLM and trtllm so far but I can do more test next week