Industry-funder keyword matching is punctuation-dependent

[hherb]

Split out of the review of #35.

_INDUSTRY_KEYWORDS in bmlib/transparency/analyzer.py is a substring test:

["pharma", "biotech", "therapeutics", "inc.", "corp.", "ltd.", "gmbh", "laboratories"]

The three org-suffix forms carry a trailing dot because a bare "inc" matches "Lincoln", "Vincent" and "province" as a substring. The cost is a false negative whenever the name is written without it — confirmed:

"Genentech Inc."  -> industry_funding_detected True
"Pfizer Inc"      -> industry_funding_detected False

This has always applied to CrossRef funder[].name. #35 widened the same list to PubMed <Grant><Agency>, where NLM-normalised strings drop the punctuation more often, so the false-negative rate is now higher than it was.

The fix is not a one-line edit. Word-boundary matching (\binc\b\.?) handles all the cases above and rejects Lincoln/province correctly, but it changes detection on the existing CrossRef corpus too, and industry_funding_detected feeds a HIGH-risk rule (industry_funding and restricted data). It needs calibrating against real funder names from both sources — measuring the false-positive rate, not assuming it — rather than being slipped into a PR under review.

Scope:

• Replace the substring test with a word-boundary regex covering inc, corp, ltd with or without the dot, and consider llc, plc, ag, bv, nv.
• Measure precision/recall against a sample of CrossRef funder names and PubMed agency strings before merging.
• Note that a false positive here can push a paper to HIGH risk, so precision matters more than recall.