fix(free): the free tier had been dead since June — rebuild it

[VickyXAI]

What was wrong

This repo's model catalog was frozen on 2026-06-08. Five of the six advertised free models died in the eleven weeks since, and nothing noticed:

id	what happened
free/gpt-oss-120b	dead upstream 2026-08-16; the gateway 400s the free/ id
free/gpt-oss-20b	same
free/deepseek-v4-flash	410 Gone at NVIDIA 2026-08-12 (blockrun #367)
free/qwen3-coder-480b	NVIDIA EOL 2026-06-14
free/llama-4-maverick	dropped from the live catalog 2026-07-17
free/nemotron-3-nano-omni…	the only survivor

None of them failed loudly, which is why this sat for months. blockrun hides a retired free id and server-redirects the call, so the caller keeps getting answers — from a different model — and /exclude on the id they can see does nothing.

/model free, the nvidia alias, the eco SIMPLE primary, and two agentic backstops all pointed into that dead set.

What this does

• Ships the current seven free models in blockrun's own auto-pick order, and repoints every default at nemotron-3.5-lightning.
• Retired entries stay in BLOCKRUN_MODELS so explicit pins still resolve. They come off the picker and the cascade only — same convention ClawRouter uses.
• Ports ClawRouter's FREE_UPSTREAM_OVERRIDES. Two of the seven are not NVIDIA-hosted, so the blanket free/ → nvidia/ rewrite in toUpstreamModelId would have put ids the gateway does not serve on the wire. Verified: nvidia/north-mini-code and nvidia/laguna-xs-2.1 are both absent from /api/v1/models; the real ids are cohere/ and poolside/.
• Sweeps the user-facing surfaces that named dead models: README (pitch, routing table, budget-model table, /exclude example, licence note) and skills/xclawrouter/SKILL.md (which also claimed "11 free NVIDIA models" — never right, and the tier is no longer NVIDIA-only).

The guard

top-models.test.ts gains what would have caught this:

• no retired free id may reappear in the picker (12 ids listed, append on every retirement)
• the free tail must equal the live tier, in cascade order, contiguously at the end
• every advertised id must be defined in BLOCKRUN_MODELS — so a picker entry can never be a phantom

Verification

All seven picker ids resolve to an id that exists in the live catalog:

free/nemotron-3.5-lightning   -> nvidia/nemotron-3.5-lightning     OK
free/nemotron-3-nano-30b      -> nvidia/nemotron-3-nano-30b        OK
free/laguna-xs-2.1            -> poolside/laguna-xs-2.1            OK
free/north-mini-code          -> cohere/north-mini-code            OK
free/nemotron-3-nano-omni…    -> nvidia/nemotron-3-nano-omni…      OK
free/nemotron-3-ultra-550b    -> nvidia/nemotron-3-ultra-550b      OK
free/llama-3.2-11b-vision     -> nvidia/llama-3.2-11b-vision       OK

514 tests pass, typecheck clean, brand-numbers --check clean.

Scope

Free tier only. The paid catalog is also a generation behind — 71 model definitions here vs ClawRouter's 113, missing Fable 5 / Opus 5 / Sonnet 5, all three GPT-5.6 tiers, Grok 4.5, Kimi K3, the whole GLM-5 line — but porting 42 model definitions is a re-fork, not a fix, and it is a separate decision. This PR fixes what is actively lying to users.

Originally this branch also carried a sync-brand-numbers --refresh commit. While I was working, the weekly brand-sync cron pushed the identical refresh to main (213068c), so the branch was rebuilt on top of that and now contains only the free-tier change.