feat(site): surface a datasets count metric

[benjibromberg]

Summary

Adds a top-line `datasets` count (129) — the number of individual catalogued datasets, not just the 14 dataset pages — and surfaces it across the homepage hero, the "Datasets by species" card, and the "By the Numbers" dashboard, replacing the old "Species / Dataset pages" headline. The species-page count (14) is retained in a detail position on the dashboard.

How the count works

"Datasets" is not one uniform shape in the corpus, so each `Datasets/` page contributes its natural, deterministic unit (no canonical Markdown was changed — the parser accommodates the existing shapes):

Source	Pages	Unit	Count
`## Complete data inventory` tables	10 species + CrossSpecies	table rows	108
Reference pages	HumanReference, CHOReference	`###` entries	7
Benchmarks	Benchmarks.md	`##` entries	14
Total			129

A new `datasets.ts` parser module (`computeDatasetBreakdown`) is the single source of truth: `counts.ts` reads its total for `counts.json`, `metrics.ts` reads its breakdown for the dashboard, and `generate-data.ts` asserts the parts sum to the total and match `counts.json` — extending the project's existing "a stat can't drift from the page it links to" invariant. `metrics.speciesInventory` now delegates to the shared `inventoryRowCount` helper.

Changes

• Parser: new `datasets.ts` (+ tests + fixtures for all three page shapes); `datasets` field added to `CountsSchema` and a `datasets` breakdown to `MetricsSchema`; `counts.ts` / `metrics.ts` / `generate-data.ts` wired with no-drift assertions.
• Site: `Hero.astro`, `SectionsGrid.astro`, `MetricsDashboard.astro` show the 129 total instead of the 14 page count; dashboard caption shows the breakdown by source-page shape.

Verification

• `pnpm test` — 289 passing (incl. new `datasets.test.ts`; ground-truth total pinned at 129)
• `pnpm parse` — `counts.datasets: 129`; breakdown 108 + 7 + 14; cross-validation asserts green
• `pnpm build` — full build passes
• Browser-verified: hero, card, and dashboard all render `129 Datasets`

🤖 Generated with Claude Code