Share restore 404s short catalog modelId while MeshLLM serves GGUF/ingress id

[toor-1970]

Important

No secrets, keys, or host identifiers in this report.

Describe the bug

On Desktop 0.5.20 with Share this machine armed, Compute persists a short catalog modelId (example: Qwen3-Coder-Next-Q4_K_M) in mesh-sharing.json. After quit/relaunch, restore asks MeshLLM for that short name and gets:

HTTP 404 model_not_found: model 'Qwen3-Coder-Next-Q4_K_M' not found (no local or remote host serving this model)
buzz-mesh: restored node is not inference-ready yet … leaving it to warm up without tearing it down

A few minutes later the same process is serving the GGUF/ingress id (Qwen/Qwen3-Coder-Next-GGUF:Qwen3-Coder-Next-Q4_K_M, runtime @main: form). OpenAI :9337 /v1/models lists the GGUF id, not the short catalog name.

During that window the Desktop UI hangs on new chat entries. Relay liveness is fine. This is not a glue-wrapper problem: restore should alias the catalog id to the id MeshLLM actually serves, and the composer should not block on Share warmup.

Steps to reproduce

1. Settings → Compute → Share a Tight-fit local model (catalog name ≠ OpenAI /v1/models id).
2. Confirm mesh-sharing.json modelId is the short catalog string.
3. Quit Desktop, relaunch (same machine, Share still enabled).
4. Immediately open a new chat / send a new entry.
5. Watch stdout: restore 404 on the short id; :3131 later llama_ready with the GGUF id.

Expected behavior

• Restore looks up the same id the runtime serves (or aliases catalog → GGUF/ingress).
• New chats remain usable while Share mmaps (Tight fit ~tens of GiB RSS in buzz-desktop is expected; the hang is the 404/warmup path).
• Composer does not wait on not inference-ready forever.

Version and platform

• Buzz version: 0.5.20
• OS: macOS (Apple Silicon, Share in-process with the UI)

Logs / additional context
Redacted:

buzz-mesh: restored node is not inference-ready yet (… last: HTTP 404 Not Found: {"error":{"message":"model 'Qwen3-Coder-Next-Q4_K_M' not found …","code":"model_not_found"}})
GET http://127.0.0.1:3131/api/status  (failed during boot race, then Serving / llama_ready)

Do not paper this over with an operator script that rewrites mesh-sharing.json on launch. The product should persist and restore one id.

[toor-1970]

Closing from the reporter. Mini Share is parked; we are not waiting on a Desktop alias for catalog vs GGUF/ingress id. Hybrid: https://github.com/toor-1970/berd_buzz_hybrid_infra/issues/65