chore(capabilities): refresh model registry from OpenRouter feed

[github-actions]

Automated weekly refresh of packages/lattice/src/capabilities/registry.generated.ts.

Generated by .github/workflows/registry-drift.yml from the live OpenRouter
/api/v1/models feed via scripts/refresh-model-registry.mjs.

Review checklist

• Any new [classifier] WARN -- unknown prefix ... lines in the workflow logs that suggest a scripts/capabilities/classifier.mjs PROVIDER_PREFIX_RULES extension?
• Diff matches expected upstream changes (new models, retired models, context window adjustments)?
• No unintended schema drift (extra fields, missing fields, reordering)?
• Anchor case study openrouter:openai/gpt-oss-120b still classifies as open_weight_instruct with internal_envelope_leak (Phase 33 anchor; do NOT merge if this regresses)?

This PR is opened by peter-evans/create-pull-request@v8.1.1. If you see no diff, the upstream feed is unchanged since the last refresh -- close the PR; the next weekly run will reopen if drift returns.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 677af4ef6d

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[chatgpt-codex-connector]

Add image modalities for Gemini image rows

When consumers use the public ModelCapabilityProfile registry for modality-aware model selection, this new image-generation entry is missing both inputModalities and outputModalities, even though the upstream OpenRouter feed for this id reports image+text inputs and image+text outputs (https://openrouter.ai/api/v1/models). That makes the registry understate image support for the newly added Gemini image models; please populate the nested architecture.*_modalities fields when regenerating.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Classify North Mini Code as an open-weight reasoning model

For cohere/north-mini-code:free, Cohere's release and model card describe North Mini Code as open-source/open-weights under Apache 2.0 with interleaved thinking/tool-use support (https://cohere.com/blog/north-mini-code, https://huggingface.co/CohereLabs/North-Mini-Code-1.0), but this new row marks it as frontier_rlhf with reasoningSurface: "none". Registry-based negotiation will therefore use frontier prompt/failure defaults and report extendedThinking: false for this model; please add a North Mini Code override or correct these fields in the snapshot.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Keep the Nex-N2-Pro free variant profile

When callers select the OpenRouter free slug, this refresh removes the exact nex-agi/nex-n2-pro:free profile and leaves only the paid base row, even though OpenRouter still advertises nex-agi/nex-n2-pro:free with free pricing and tool/structured-output support (https://openrouter.ai/nex-agi/nex-n2-pro%3Afree). getCapabilityProfile("openrouter:nex-agi/nex-n2-pro:free") now misses, while findCapabilityProfile("nex-agi/nex-n2-pro:free") strips to this paid row, so users inspecting or budgeting the free variant can get the wrong capabilities and pricing.

Useful? React with 👍 / 👎.