[Enhancement]: Lock model/endpoint selection per conversation (prevent context leaking to a different model after the first message)

[hboie]

Problem / Motivation

LibreChat currently allows switching the endpoint/model mid-conversation at any time, which silently serializes and forwards the entire existing conversation history to the newly selected model/provider. While this is a well-documented and intentional feature for flexibility, it creates a problem for deployments with strict data governance requirements:

• Organizations that need to guarantee a conversation's context never leaves a specific model/provider boundary (e.g., for compliance, data residency, or contractual reasons between different LLM vendors) currently have no way to enforce this per-conversation.
• The existing interface.modelSelect: false and modelSpecs.enforce: true options solve this only globally (disabling model choice entirely, even for new conversations), which is too restrictive for deployments that want to offer multiple models to users — just not allow switching within an already-started conversation.

Proposed Solution

Add a new interface.lockModelPerConversation boolean option (default: false, fully backward-compatible) to librechat.yaml:

interface:
  lockModelPerConversation: true

Behavior when enabled:

• Users can still freely choose any available model/endpoint when starting a new conversation.
• Once a conversation has been persisted (i.e., conversationId !== 'new'), the model/endpoint selector becomes locked for that specific conversation — both in the UI and enforced server-side.
• Attempts to switch endpoint/model for an existing conversation (via UI, URL query parameters, or direct API calls) are rejected, preventing the conversation history from ever being forwarded to a different model/provider.

Implementation Details (prototype already built and tested)

We've implemented and are currently running this in a staging/test environment. Rough outline of the changes, happy to adapt based on maintainer feedback on architecture/placement:

• packages/data-provider/src/config.ts – New field in interfaceSchema + default value (false)
• packages/data-schemas/src/app/interface.ts – Passthrough in loadDefaultInterface so the config value reaches appConfig.interfaceConfig
• api/server/middleware/buildEndpointOption.js – Server-side enforcement: compares the requested endpoint/model against the conversation's originally stored values via getConvo(), rejects mismatches with a 409 when the flag is enabled and the conversation already exists (bypassed for the Agents endpoint and for brand-new conversations)
• client/src/components/Chat/Menus/Endpoints/ModelSelectorContext.tsx – Computes isModelLocked based on the route's conversation ID (not the chat context state, to avoid a race condition during "New Chat" navigation) and guards the selection handlers
• client/src/components/Chat/Menus/Endpoints/ModelSelector.tsx – Visually disables the model selector trigger and shows a localized tooltip explaining why, while preserving hover/focus accessibility (aria-disabled + onClick guard instead of native disabled, to keep the tooltip functional)
• New i18n keys (com_ui_model_locked, com_ui_model_locked_tooltip) added to client/src/locales/en/translation.json
• Unit tests added for the backend enforcement logic and the loadDefaultInterface passthrough

Areas of the Application Affected

• librechat.yaml schema / config validation
• Backend request validation middleware (buildEndpointOption.js)
• Frontend model selector UI and its underlying context
• Localization files