Fix empty funder chips on dataset pages

[neuromechanist]

Closes #204.

Problem

Funding.funder did not exist. data.nemar.org sends funder_name, so
DetailRail.astro rendered an empty <span class="rail__funder"></span> for
every funding entry. Live on production right now: nm000103 has two funding
entries and shows two blank spans. Award numbers were unaffected
(award_number was already the right name), which is why this read as a
partial rendering quirk rather than a broken field.

Change

• neuroschema.ts: Funding.funder -> funder_name, and the two fields the
  API sends but the type omitted (award_uri, funder_identifier_type).
  Typed nullable, matching the rest of the file: a shape violation must render
  an incomplete rail, never throw and drop the page.
• format.ts: new displayableFunding() resolves and trims the name and drops
  entries without one, so a null can never reproduce the blank-chip symptom.
• DetailRail.astro consumes it.

No compatibility branch for the old key: I surveyed 11 funding entries across
10 datasets on production and zero carry a funder field.

Tests

Five cases in src/lib/format.test.ts, including one against
test/fixtures/funding-nm000103.json captured from
data.nemar.org/nm000103/metadata.json — the exact document that renders
blank today. Full suite 1020 passing, typecheck and lint clean.

Verified rendering against real production data on a dev server: both funder
names now appear ("See https://childmind.org/..." and "NIH"), with the
R01MH125934 award number still alongside.

[cloudflare-workers-and-pages]

Deploying nemar-website with    Cloudflare Pages

Latest commit:	eae527f
Status:	✅  Deploy successful!
Preview URL:	https://397f2f6f.nemar-website.pages.dev
Branch Preview URL:	https://fix-issue-204-funder-field.nemar-website.pages.dev

View logs

[neuromechanist]

Review (code-reviewer subagent, sonnet)

Verdict: approve, no blocking issues. One cosmetic nit, fixed in eae527f.

Checks run independently of CI: typecheck 0 errors (213 files), lint clean (209 files), full suite 1020/1020 including the 5 new displayableFunding cases.

The fixture is real. test/fixtures/funding-nm000103.json was diffed against a live curl of data.nemar.org/nm000103/metadata.json and is a byte-for-byte match, satisfying the no-mocks rule.

The survey went much wider than mine, and it settles the one real design risk. I had checked 11 funding entries across 10 datasets. The reviewer fetched all 754 datasets and every metadata.json, covering 1080 funding entries across 408 datasets:

• every entry carries exactly {funder_name, award_number, award_title, award_uri, funder_identifier, funder_identifier_type}
• zero use a funder key, so the decision not to write a compatibility branch is correct
• zero have an award_number/award_title with no usable funder_name

That last one matters: the open question on displayableFunding was whether dropping nameless entries could hide award data a user should see. It cannot, because that combination does not occur anywhere in the catalog.

Other findings: the generic signature is sound (the spread preserves sibling fields, covered by test); folding metadata.funding ?? [] into the helper rather than duplicating the guard is correct; a grep for \bfunder\b excluding the legitimate names found no remaining reader of .funder anywhere in the codebase.

Nit (fixed): the fixture had no trailing newline. Cosmetic only — lint runs biome check src, which does not cover test/fixtures.

Findings addressed: 1 of 1. Nothing skipped.