fix(ai-gateway): resolve permission error when allow all models is configured (#3629)

[Adityakk9031]

close: #3629

Description & Root Cause

When configuring an AI Provider (such as Custom, OpenRouter, or Vercel AI Gateway) with Allow List set to "Allow all models" (*), requests from AI clients like Open Code (e.g., testing with models such as unsloth/Qwen3.8-27B-GGUF:UD-Q4_K_XL) failed with a proxy/permissions error (404 Not Found returned as a forbidden/proxy failure).

Investigation revealed two root causes:

1. Model Discovery on GET /v1/models/:model:
   • Because * is a wildcard pattern, expandProviderModels does not expand it into explicit model IDs. For aggregator/custom providers, provider.catalog was empty because they lack a fixed vendor prefix catalog.
   • When clients query GET /v1/models/:model for a model allowed by wildcard, handleV1Models was only searching the pre-enumerated models array (models.find), returning 404 Not Found instead of checking if any attached provider allows the requested model via wildcard patterns (isAllowedByLists).
2. Provider Selection Tie-Breaking for Custom / Unknown Models:
   • In catalogOwnershipScore (server/lib/aiProviderSelection.ts), unknown/custom models not in the static catalog scored 0 for all providers.
   • When tie-breaking in compareProviderCandidates, typed providers (e.g. openai, class rank 2) were preferred over custom providers (class rank 0), mistakenly routing non-catalog / self-hosted models to native vendor endpoints that cannot serve them.

Changes Made

• server/routers/aiGateway/v1Models.ts:
  • In catalogMetadataForType: Populated catalog entries for custom, openRouter, and vercelAiGateway using aiModelCatalog.getAll() so wildcard matching has full catalog metadata.
  • In handleV1Models: For GET /v1/models/:model, added fallback wildcard validation with isAllowedByLists(requestedModel, p.allows, p.blocks) across discovery providers before returning 404.
• server/lib/aiModelDiscovery.ts:
  • Included non-pattern configured model keys in expandProviderModels candidates.
• server/lib/aiProviderSelection.ts & docs/ai-gateway-provider-selection.md:
  • Updated catalogOwnershipScore so aggregator/custom providers score 1 for unknown/custom models while typed providers score 0 (typed catalog miss), properly routing custom models to custom/aggregator providers.
  • Updated documentation table and Example H accordingly.

Verification

• npm run dev:check: All TypeScript type checks and Prettier formatting passed (0 errors).
• Server build (node esbuild.mjs -e server/index.ts -o dist/server.mjs) succeeded.