feat: build and refresh the featured models list daily

[ashokgelal]

Stacked on #8. Merge #8 first; GitHub then moves this PR's base to main.

Problem

The featured models list is rebuilt only by hand, with a Python script in the private msty-nexus repo. Sizes and files go stale between rebuilds. Ollama now publishes some sizes as ranges ("6.6GB - 7.6GB"), and the old builder dropped size_bytes for those tags without any error.

Solution

• The builder now lives here, in Node with no dependencies, like the Ollama job. Reasons: msty-nexus is private, so a workflow here would need a cross-repo token to run the Python script. Also, one repo now owns both the data and its builder.
• scripts/featured-curation.mjs holds the curation: which models, names, summaries, tiers, coding tags, collections. The job never edits it. Adding, dropping, or replacing a model is still a reviewed PR.
• refresh-featured.yml runs daily. It refreshes files, sizes, parameters, context, and popularity. It commits only when something other than generated_at or popularity changed. Nexus reads popularity live, so those numbers alone do not need a commit.
• An Ollama size range uses the upper end. An Ollama size that cannot be read is now a build problem, so the old bug fails loudly. Any unconfirmed reference fails the run without writing.
• featured-verify.mjs runs in CI. It checks the file's shape and sizes, and that the file matches the curation. A curation edit without a rebuild fails the PR.
• featured-report.mjs writes leads to the job summary: mlx-community repos that may be newer versions of featured models (each model has a watch term). It changes nothing. Today it finds, for example, MiniMax M3, GLM-5.3 Flash, DeepSeek V4, Ornith 1.5, and MiniCPM-V 4.7. Expect some noise, such as other sizes of a featured family.

Test

• The Node port produces the same output as the Python builder, except the 7 sizes it now reads from ranges.
• Dry run on the published file: 6 rows change, all from upstream. Ollama re-published qwen3.5:9b, qwen3.5:0.8b, qwen3.5:2b, gemma4:12b, and glm-ocr with new sizes and parameter counts, and one Qwen3.8 GGUF grew by 96 bytes. Collections are unchanged. The second commit is that refresh.
• A second run reports "unchanged".
• node --test scripts/*.test.mjs (50 pass) and every verifier pass.

Notes

• After this merges, cloudstack-llc/msty-nexus#585 removes the Python builder from msty-nexus.
• The job uses the existing REFRESH_TOKEN secret, like the other refresh jobs.