starplast.hf_publish

Publishing the study corpus to HuggingFace, license-gated.

The 97 proximity-labeling and pulldown studies come to ~400 MB, which is past what belongs in git but unremarkable for a HuggingFace dataset repo. The obstacle is not size, it is redistribution rights.

Measured from the JATS license blocks of the local full texts:

52  CC-BY               redistributable with attribution
 1  non-commercial      must be excluded
 6  vague               "open access", "other" -- needs a human decision
38  no local full text  unknown

So this module does two things and refuses to do a third. It builds a derived release — the parsed gene-membership table plus a manifest of source URLs — which sidesteps most of the ambiguity, is what is actually useful downstream, and is roughly a hundredth the size. And it can mirror the raw supplementary files, but only for studies whose license it has positively confirmed as permissive; anything unknown or non-commercial is skipped and listed, never silently included.

Nothing here uploads without an explicit call. plan() shows what would go, and is the intended first step.