Dashboard: data-driven evidence summary + top candidates per pool

[sr320]

Rewrites the What the evidence says so far section of https://sr320.github.io/AREE/ so that every sentence comes from data.json instead of hardcoded prose.

What was out of date on the live page

• It said "All real studies so far are RNA-seq pathogen challenges", which stopped being true once the heat, OA and starvation studies were added.
• It listed single-study phenotypes as pools with "0 genes measured in both", and said "both" for a 3-study pool.
• The next step it gave ("a second stressor class") is already done.

What the section now computes

• Stressor and assay coverage, plus which assays actually have harmonized evidence.
• Cross-study pools (k ≥ 2 only) as a list. Single-study phenotypes are named as awaiting replication.
• Tier and evidence-class counts.
• Caveats that only show when they apply: the top list being disease-only, the multi-omics tier being empty.
• Next curation steps, worked out from unharmonized assays, unreplicated phenotypes and QC failures.

Tested by rendering against the live data.json and against an older local build (10 studies, one pool). Also guarded the CI case with no evidence.
Top candidates per pool (second commit)

• The overall top-15 list is filled entirely by the 3-study disease-susceptibility pool, so thermal tolerance's replicated candidates never appeared.
• aree build-dashboard now also writes top_real_by_phenotype / top_simulated_by_phenotype: the top 5 per phenotype, each with its evidence card and gene annotations.
• The home page adds a Top candidates per pool section with one table per cross-study (k ≥ 2) pool. The existing overall table is unchanged.
• New test assertions cover per-phenotype ranking order and size limits, and check that the field is empty when nothing has run. Full suite: 564 passed; ruff clean.

🤖 Generated with Claude Code