fix open-weights metadata: 29 missing weights links, 3 wrong flags

[arimxyer]

What

The repository invariant open_weights = true ⇒ weights links (packages/core/test/generate.test.ts, "repository open-weight model metadata includes weights links") currently fails on 32 registry records. Every one was verified against Hugging Face individually — each URL below was fetched and confirmed to be that model's official repo.

29 records were genuinely open but missing links — each gets its canonical Hugging Face repo as a [[weights]] entry (the label = "Hugging Face" + url form used by all 93 existing entries). Non-obvious repo spellings, all verified live:

• NVIDIA underscore variants: Llama-3_1-Nemotron-Ultra-253B-v1, Llama-3_3-Nemotron-Super-49B-v1_5, Llama-3_3-Nemotron-Super-49B-v1
• The 70B Instruct link points at the Transformers conversion (-HF) rather than the NeMo-checkpoint repo, per the model card's own recommendation
• Nemotron 3 BF16-suffixed repos: Nano-30B-A3B, Super-120B-A12B, Ultra-550B-A55B, Nano-Omni-Reasoning
• Cohere publishes under CohereLabs: command-a-plus-05-2026-bf16, North-Mini-Code-1.0
• poolside Laguna weights are public: Laguna-M.1 (Apache-2.0), Laguna-S-2.1 / Laguna-XS-2.1 (OpenMDW-1.1), Laguna-XS.2
• Plus sarvam (105b, 30b), whisper (large-v3, large-v3-turbo), Kimi-K3, and the remaining NVIDIA Nemotron/safety/embed/rerank repos

3 records had the flag itself wrong and are corrected to open_weights = false, each with a top-of-file citation comment (per AGENTS.md comment placement rules):

• deepreinforce/ornith-1.0-31b — the 31B Dense variant is named in the announcement but was never published: the org ships only 9B/35B/397B, and no 31B repo or community quantization exists anywhere on HF
• nvidia/mistral-nemotron — no public weights; served API-only via NVIDIA NIM, nvidia/Mistral-Nemotron does not exist on HF
• nvidia/nemotron-voicechat — weights are behind an application-only Early Access Program, not publicly downloadable; no HF repo exists. Worth revisiting if NVIDIA opens general access.

Verification

bun validate passes, and the core test suite goes fully green with this change (111 pass / 0 fail — the weights-link invariant was the only repository-data failure). A follow-up PR proposes running that suite in CI so this class of drift is caught at PR time.

🤖 Generated with Claude Code

[github-actions]

Action items

• [high] [violation] models/nvidia/mistral-nemotron.toml:13 - Check: Provider models that set open_weights must stay consistent with the corrected base metadata (primitives replace on merge). Why: This PR sets open_weights = false on models/nvidia/mistral-nemotron and models/nvidia/nemotron-voicechat, but the NVIDIA provider entries still hardcode open_weights = true (providers/nvidia/models/mistralai/mistral-nemotron.toml:9, providers/nvidia/models/nvidia/nemotron-voicechat.toml:9). Those overrides win the merge, so the published provider catalog keeps advertising non-downloadable models as open-weight. Action: Set both provider files to open_weights = false, or drop the field so they inherit the base value.