[VickyXAI]

Why

#43 pinned all four Free tiers to nvidia/llama-3.2-11b-vision, because every other free id showed available: false. That flag is not a delisting. /v1/models sets it to isFreeModelUsable(model.id) (blockrun src/app/api/v1/models/route.ts:59), which is the gateway's per-instance free-health circuit breaker (src/lib/free-health.ts). It flips between instances and recovers by itself.

The gateway already handles fallback

The Go router has no per-request fallback, and adding one would only stack waits on top of the gateway's own.

• Before dispatch: a requested free model whose breaker is open is rerouted to the first usable, capability-compatible rung of FREE_RESCUE_LADDER (chat/completions/route.ts ~1811, pickFreeRescue).
• On failure: the call cascades nemotron-3-super-120b → nemotron-3-nano-omni-30b-a3b-reasoning → llama-3.2-11b-vision within a 90s budget (route.ts:394/422/452/465, 3463+).
• Verified live today: a request for nemotron-3.5-lightning came back x-health-reroute: true and x-served-model: nvidia/nemotron-3-super-120b. A request for nano-omni came back served by nvidia/llama-3.2-11b-vision.

Pinning a primary is therefore safe, and the gateway's ladder supplies the fallbacks, ending at llama-3.2-11b-vision.

Free table

router-core (ae6990f) has no free profile; its only free rungs are the eco SIMPLE head. This table mirrors the sibling SDK's free profile, blockrun-llm router_adapter.py FREE_TIERS (a5af583):

Tier	Primary	Python SDK fallbacks (reference)	Gateway reroute ladder
Simple	nvidia/nemotron-3-nano-omni-30b-a3b-reasoning	nemotron-3.5-lightning, llama-3.2-11b-vision, laguna-xs-2.1	super-120b → nano-omni → llama-3.2-11b-vision
Medium	nvidia/nemotron-3.5-lightning	nano-omni, laguna-xs-2.1, north-mini-code	same
Complex	nvidia/nemotron-3.5-lightning	north-mini-code, nano-omni, llama-3.2-11b-vision	same
Reasoning	nvidia/nemotron-3.5-lightning	nano-omni, north-mini-code, laguna-xs-2.1	same

The fix from #43 for hidden ids stays: nvidia/nemotron-3-super-120b (hidden:true in models.ts; it is a gateway-internal cascade rung) and moonshot/kimi-k2.x remain in the retired-id test, and Eco/Auto Simple are unchanged.

Other changes

• TestRoutingTableAvoidsRetiredIDs no longer lists ids whose only problem is available: false.
• The opt-in live test still fails on unlisted ids, and on free tiers that are not billed as free. A breaker-open routed id is now logged instead of failing the test.
• The README free-model list and notes and the CHANGELOG Unreleased entry are updated.

Checks

• go build ./...: ok
• go vet ./...: ok
• go test ./...: ok
• BLOCKRUN_LIVE_CATALOG=1 go test -run 'TestRoutingTable|TestReadme|TestFreeRouting' .: ok. It logged that all four free primaries are breaker-open right now, which the gateway reroutes.

Known limitation

When the gateway reroutes, SmartChatResponse.Model still names the requested primary rather than the model that served the call. The served model is reported in the x-served-model header and the response model field.