fix(catalog): restore LiteLLM bundled metadata fallback

[roboomp]

Repro

With models.dev mocked as empty and LiteLLM /model_group/info returning glm-5.2, litellmModelManagerOptions({ baseUrl, fetch }).fetchDynamicModels() returned a LiteLLM model with reasoning: false and no thinking block. The recorded repro used the catalog module directly with a mocked fetch implementation matching omp models refresh litellm --json enrichment inputs.

Cause

packages/catalog/src/provider-models/openai-compat.ts loaded models.dev references in litellmModelManagerOptions and used modelsDevReferences.get(id) as the only reference resolver. When models.dev lacked glm-5.2, both mapLiteLLMRichEntry and mapLiteLLMOpenAICompatibleModel received undefined and fell back to default non-reasoning metadata instead of the bundled catalog entry.

Fix

• Changed LiteLLM dynamic discovery to build its resolver with createReferenceResolver(modelsDevReferences), preserving models.dev precedence while adding bundled catalog fallback metadata.
• Added a rich /model_group/info regression for glm-5.2 proving LiteLLM transport fields remain while bundled reasoning/thinking metadata is inherited.
• Added a /v1/models fallback regression for the same bundled metadata path.
• Added the catalog changelog entry for LiteLLM discovery enriches against models.dev only, misses bundled catalog reasoning/thinking metadata #4695.

Verification

bun test packages/catalog/test/litellm-provider.test.ts -t "enriches LiteLLM rich models missing from models.dev with bundled reasoning metadata" passed; bun test packages/catalog/test/litellm-provider.test.ts -t "enriches LiteLLM /v1/models fallback entries missing from models.dev with bundled reasoning metadata" passed; bun test packages/catalog/test/litellm-provider.test.ts passed with 15 tests; bun --cwd packages/catalog check passed. gh_push_branch also completed its pre-publish gate before pushing. Fixes #4695

[chatgpt-codex-connector]

Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits.
Repo admins can enable using credits for code reviews in their settings.

[roboomp]

No code change requested here. The Codex review bot could not run because the repo has reached code-review usage limits; a repo admin needs to add/enable credits in the Codex cloud settings before that review can proceed.