fix(models): correct cache, usage, streaming, and image outputs

[git-jxj]

Summary

This PR fixes six independent model-layer correctness issues:

• Model memoization could return an instance created for a different eval_type.
• ModelScope generation with multiple choices only reported usage from one choice.
• Packed streaming choices could lose their individual finish_reason values.
• Text-to-image and image-editing outputs omitted the data: URI prefix required by downstream consumers.
• LiteLLM requests ignored reasoning_history and the configured base URL.
• Image model outputs stored an epoch timestamp instead of elapsed generation time.

Each behavior is covered by deterministic tests with mocked model or SDK dependencies.

Root Cause

The model cache key did not include eval_type, so otherwise identical task and judge configurations could collide across backends.

ModelScope usage and OpenAI-compatible stream metadata were assigned from a single choice instead of being aggregated or indexed per choice.

Image model adapters serialized raw base64 content and assigned time.time() directly to ModelOutput.time, while downstream APIs expect a data URI and elapsed seconds.

The LiteLLM request builder did not forward the same reasoning-history and base-URL settings used by the OpenAI-compatible path.

Changes

• Include eval_type in the model memoization key and document the cache contract.
• Aggregate ModelScope token usage across every generated choice.
• Track streaming finish_reason by choice index.
• Emit complete image data URIs from text-to-image and image-editing models.
• Report elapsed generation time for image outputs.
• Forward LiteLLM reasoning_history and base_url configuration.
• Add regression tests for all six behaviors.

Reproduction

On the unmodified base:

same model/config with different eval_type -> the same cached backend instance
ModelScope two-choice usage              -> 10 instead of 20
second packed choice finish_reason       -> stop instead of length
image output consumed as media           -> FileNotFoundError
image ModelOutput.time                   -> approximately 1.7e9 seconds
LiteLLM reasoning/base URL settings      -> silently omitted

The regression suite reports 11 failures against the old source. After this change, each configuration produces the intended backend, metadata, image payload, and timing value.

Validation

Targeted regression tests:

python -m pytest \
  tests/models/test_model_cache.py \
  tests/models/test_modelscope_usage.py \
  tests/models/test_openai_stream_finish_reason.py \
  tests/models/test_image_model_output.py \
  tests/models/test_litellm_reasoning_history.py -q

Result:

13 passed

Model-layer regression suite:

python -m pytest tests/models -q

Result:

62 passed

CI smoke test:

python -m pytest tests/cli/test_all.py::TestRun::test_ci_lite -q -p no:warnings

Result:

1 passed

All current Ruff and repository pre-commit checks passed.

Scope

These changes do not alter prompts, generation parameters, cache lifetime, retry behavior, image encoding content, or non-LiteLLM request routing. The branch is rebased onto the current main and contains one commit.

[Yunnglin]

LGTM!