[nexxeln]

Summary

• add a built-in vLLM provider that checks /health and discovers extended ModelCards from /v1/models
• project served model IDs and max_model_len with process-wide polling deduplication
• support custom endpoints, path-prefixed proxies, bearer authentication, and live config reloads
• replace and restore same-ID Models.dev entries based on successful live discovery
• conservatively avoid advertising tools or vision because vLLM does not expose its server launch capabilities
• document zero-config and configured vLLM discovery

Testing

• bun test in packages/core (1840 passed, 9 skipped)
• bun typecheck in packages/core
• repository-wide pre-push typecheck
• targeted Oxlint and Prettier checks
• V2 docs validation and production build
• branch server smoke test against a congruent vLLM mock: readiness, discovery, and streamed /api/generate response