Multi-Modal Vision selector omits GPT 5.6 and future endpoint models

Issue

GPT 5.6 and similarly named future GPT models configured under Model Endpoints can appear in Metadata Extraction but remain absent from the Multi-Modal Vision Analysis model selector.

Manually configured endpoint models may have an empty canonical modelName, a spaced display name such as GPT 5.6 Luna, or the GPT version only in deploymentName. The Vision selector must recognize those aliases consistently.

Steps to Reproduce

1. Configure an enabled Model Endpoint with an enabled GPT 5.6 model.
2. Leave modelName empty or use a display name with spaces while retaining the GPT version in displayName or deploymentName.
3. Open Admin Settings and compare Metadata Extraction with Multi-Modal Vision Analysis.

Expected Behavior

The Vision Model selector includes GPT 5.6 and later vision-capable GPT versions regardless of whether the version is stored in modelName, displayName, or deploymentName. Existing GPT-4 Vision and o-series behavior remains available, and unrelated models remain excluded.

Actual Behavior

The model appears in Metadata Extraction but is omitted from Multi-Modal Vision Analysis because the Vision filter checks only one alias and relies on narrowly formatted GPT name substrings.

Impact

Administrators cannot select valid current or future GPT vision models for image analysis without renaming or restructuring endpoint model metadata.

Notes

The initial Jinja-rendered selector and the browser-side repopulation path must use equivalent capability rules. Add functional and Playwright regression coverage for GPT 5.6, 5.7, 5.8, 6, and 7 aliases plus a non-vision exclusion case.