Support GPT 5.6 and later models in Multi-Modal Vision selector

[paullizer]

Issue

The Multi-Modal Vision Analysis model selector omits configured GPT 5.6 deployments even though the same enabled deployments appear under Global Endpoints. This affects GPT 5.6 variants such as Luna, Sol, and Terra and may affect later GPT model versions that use aliases or different model-name fields.

Steps to Reproduce

1. Enable Multi-Model Endpoints and configure an enabled endpoint with GPT 5.6 deployments (for example Luna, Sol, or Terra).
2. Save the endpoint with those models enabled.
3. Open Admin Settings > Search and Extract > Multi-Modal Vision Analysis.
4. Open the Vision Model selector.

Expected Behavior

Every enabled vision-capable GPT 5.6 deployment appears in the selector, and future GPT 5.x or later naming variants are recognized when their model metadata or deployment name identifies them as GPT models.

Actual Behavior

The selector shows GPT 5.4 deployments but omits the enabled GPT 5.6 deployments shown under Global Endpoints.

Impact

Admins cannot choose newer multimodal deployments for image analysis, forcing them to remain on GPT 5.4 even when newer models are configured and governed.

Notes

The client-side vision filter currently checks only a subset of model naming fields and uses narrow literal patterns. Validate Azure OpenAI, New Foundry, and classic Foundry endpoint model objects, including aliases with spaces or hyphens.

Acceptance Criteria

• Enabled GPT 5.6 Luna, Sol, and Terra deployments can appear in the Vision Model selector when their metadata identifies them as vision-capable GPT models.
• Capability matching considers model, display, and deployment naming fields instead of depending on one field.
• Separator and case differences do not prevent supported GPT model families from being recognized.
• Existing GPT 5.4, GPT-4o, GPT-4.1, GPT-4.5, vision, and supported o-series behavior remains intact.
• A focused UI regression test covers the newer multi-endpoint model variants.

[paullizer]

Implemented in PR #1087.

Summary:

• Added GPT 5.6 and later capability matching across modelName, displayName, deploymentName, deployment, and name fields.
• Applied the same matching contract to initial Jinja rendering and client-side endpoint updates.
• Covered Azure OpenAI and New Foundry aliases including Luna, Sol, Terra, provider prefixes, normalized separators, and a future GPT 5.7-style deployment.
• Preserved disabled endpoint/model and unsupported-family filtering.
• Updated version 0.250.066, fix documentation, feature guidance, and release notes.

Validation:

• 17 affected UI, vision, and route-policy tests passed.
• JavaScript syntax, Python compilation, Pylance diagnostics for new files, editor diagnostics, and whitespace checks passed.

Commit: dafcdc1

[paullizer]

Follow-up fix pushed in commit ea7e590 after reproducing a local admin-page 500:

• Root cause: a long-lived Flask worker had the pre-change route module while Jinja reloaded the new template, so is_vision_capable_model was absent from the render context.
• Added a guarded Jinja fallback for all vision selector branches. During worker/template skew, the selector renders its placeholder and the existing client-side population runs instead of returning UndefinedError.
• Extended the Playwright regression to render the actual selector both with and without the route helper.
• Re-ran the affected suite: 17 passed. Syntax, diagnostics, and whitespace checks passed.

PR: #1087