[logins-pathbit]

What happens

sync_combos upserts the standard combos on every daemon cycle with hardcoded model lists that reference every supported provider. On a gateway where only some providers are connected, every probe of those combos logs one line per dead entry (Model X failed, trying next) and All models failed: No active credentials for provider when the chain empties — every five minutes, indefinitely, for providers that were never connected to this gateway.

Steps: connect only one API-key provider (e.g. Groq) to a fresh gateway, let the daemon sync combos, then call any combo such as arsenal-rapido. The combo still lists mistral/* and ag/* models and each call fails through them before reaching Groq.

Why

get_default_combos returns static lists and sync_combos writes them verbatim (src/nine_rtksync/combos.py). The sync never looks at what is actually registered in the gateway database.

What this changes

sync_combos now reads the registered providers (providerConnections for vendors, providerNodes for custom OpenAI-compatible endpoints) and:

• keeps only combo models whose provider prefix is registered;
• removes standard combos that would be left with zero models;
• recovers automatically on the next cycle once a provider is connected.

The ag/ plan prefix used in combos is mapped to the antigravity / gemini-cli / google vendor names recorded in the connections table.

User-created combos are never touched: only the standard ids returned by get_default_combos are filtered or removed.

Regression coverage

tests/test_combos.py gains connection-aware cases against a minimal gateway schema: no connections (standard combos removed), partial connection (only the connected provider's models survive), Antigravity connected (ag/* kept), custom node registered (offline combo kept). Full suite: 448 passed.