Epic: sample onboarding, multi-source metadata, and publication (MVP)

Plan: docs/plans/onboarding-and-publication.md (PR #273).

MVP. About 470 samples from 4 studies go through the whole path:

registered
run under cmgd_mpa4.2 2.3.0
staged to the lake as Parquet
published with cMD and Olm metadata to cmgd-public, Zenodo and a Hugging Face Dataset repo
navigable by a reviewer agent from llms.txt in 4 fetches or fewer

Cohort:

Estimated compute: about 11.3k core-h at the 2.3.x rehearsal rate of 24 core-h per sample.

Decisions to confirm before Phase 1 (plan section "Decisions to confirm"):

D1: which catalog holds the metadata (cmgd_lake or the shared lake's cmgd_meta schema)
D2: add-study write-through to v2
D3: Olm already contains cMD (external.cmd)
D4: per-table licence (Olm data is CC-BY-4.0)

Phases

Later, outside the MVP: #272

Decisions (2026-10-16, autonomous run while user is offline):

D1: metadata goes in the shared lake under schema cmgd_meta (lake.cmgd_meta.*), so a single snapshot covers both results and metadata. This keeps the user's choice of per-source tables in the cmgd lake, just without the retired cmgd_lake catalog.
D2: fix the add-study trigger and add write-through to v2 as the first step of Phase 3.
D3: ingest Olm's long determinations table and exclude external.cmd rows when treating Olm as an independent source.
D4: license Olm-derived tables CC-BY-4.0 per table and don't redistribute evidence_quote.
First release: built from the 125 cmgd_mpa4.2 samples already completed, plus metadata. The 470-sample cohort runs on Alpine and Anvil and lands in the next release. Bridges-2 stays off (max_concurrent_runs: 0).
Zenodo and Hugging Face: no public DOI is minted without the user's sign-off. The mirror code is tested against Zenodo sandbox and a private HF repo if credentials exist.

Co-Authored-By: Claude Opus 5.5 noreply@anthropic.com

Discussion issues opened (no commitment yet): #290 AWS Open Data, #291 R / Bioconductor access, #292 Zenodo deposit strategy at corpus scale.