Precomputed biomodel summaries and publication metadata

[androemeda]

Adds an AI-generated, biologist-facing summary for each biomodel, precomputed offline and stored in Supabase, plus the publication metadata it draws on.

Publications are synced from the VCell v1 API into Supabase along with PubMed abstracts (250 publications, 287 links over 258 biomodels, 232 with abstracts), and are shown in a new Publications section on the biomodel Overview tab. Summaries are generated by a batch script following the structure defined in SKILL.md, which is read at runtime and hashed into each row so edits to it invalidate and regenerate the affected summaries. Input is the element extracted from the VCML plus the BNGL export for rule-based models, which keeps a typical model to roughly 9k tokens instead of the 30k+ a raw VCML would cost. The summary renders as a card at the top of the AI Analysis tab and is seeded into the chat as context.

Generation is idempotent — a re-run skips any model whose inputs and SKILL.md are unchanged — and failures are recorded rather than aborting the run. The publications endpoint reads Supabase first and falls back to the live VCell feed, so nothing breaks before a sync. Ten models have been generated and verified against the SKILL.md checklist; the remaining ~1251 are a single command.