feat(inference): per-request VLM model override

[Liuhaai]

Description

Adds an optional model field to DescribeRequest and CropDescribeRequest so callers can pick a different upstream model per request without restarting the server. The cortex client uses this to route VLM scene-describe and segmentation ground-region requests to different upstream Qwen models from a single trio-core instance (e.g. qwen3-vl-flash for scenes, qwen3.6-plus for segmentation).

Type of Change

• New feature (non-breaking change that adds functionality)

Changes

• DescribeRequest / CropDescribeRequest: add optional model: str | None (default None)
• /api/inference/describe and /api/inference/crop-describe handlers: forward req.model into engine.analyze_frame(...)
• TrioCore.analyze_video / analyze_frame: accept and forward model
• BaseBackend.generate / stream_generate: gain optional model kwarg
• RemoteHTTPBackend: uses model or self._remote_model in the upstream chat.completions.create call; log line includes the effective model
• Local backends (transformers, mlx, tome, compressed): accept the kwarg and call BaseBackend._warn_model_override_once(model) — one warning per backend instance, then ignore (local backends can't swap models per-request)

When the request omits model, behavior is identical to today.

Test Plan

• All existing tests pass (python -m pytest tests/ -v for backends/engine/config/inference_router → 60 passed)
• New tests added for new functionality
  • test_generate_uses_per_request_model_override — asserts the OpenAI client receives the per-request model
  • test_generate_falls_back_to_configured_model — asserts fallback to _remote_model when override is None
• Regression test passes (python examples/run_regression.py) — not run; this PR doesn't change inference math (only routes the model name through), and remote inference depends on a live DashScope endpoint
• Manual sanity: targeted tests cover the changed code path; cortex-side caller wiring in a separate PR has been validated end-to-end against the new request schema

Checklist

• My code follows the project's code style (type hints, no unnecessary abstractions)
• I have not introduced any new AGPL-licensed dependencies
• This PR contains no breaking API changes — model is optional with default None and existing callers are unaffected
• I have not committed secrets, API keys, or credentials

🤖 Generated with Claude Code