Add vocab_labels.parquet builder (#148 step 1)

[rdhyee]

First step of #148.

Summary

• Adds `scripts/build_vocab_labels.py` — parses the same 10 SKOS TTLs that `scripts/generate_vocab_docs.sh` already pulls (core + Earth Sci + OpenContext + Biology), emits a single `vocab_labels.parquet` with columns: `uri, uri_form, pref_label, lang, scheme, definition, alt_labels, source_ttl`.
• 537 rows / 535 unique URIs / 10 schemes / `en` + `de`.
• Coverage against the 55 distinct `IdentifiedConcept` URIs in the current wide parquet: 51/55 (93%) by URI, >99.99% sample-weighted on every facet (material, context, object_type).

What it doesn't do (yet)

• Publish to `data.isamples.org`.
• Wire any consumer (Explorer, notebooks, pqg facet-summaries) to use it.

Both follow-ups, in their own PRs once we agree on the artifact shape.

Two surprises handled (full write-up on the issue)

1. `/1.0/` URI mismatch. TTLs declare concepts without a version segment, but iSamples export records carry `/1.0/` in the URI. Confirmed: the version segment is in the export itself, propagated faithfully by pqg. The builder emits dual-form rows tagged `uri_form = "vocab" | "data_v1"` so JOINs work on either form. Per-prefix oddities (OpenContext uses `/0.1/`; biology data has inconsistent casing — `Animalia/Fungi/Plantae` but `bacteria/protozoa`) are handled.
2. Cross-vocab redeclarations. 17 material URIs declared in 2-3 TTLs each with slightly different labels (whitespace, casing). Builder dedupes `(uri, lang)` preferring the canonical TTL; losers' labels preserved in `alt_labels`.

Residual unresolved (4 URIs, ~169 / ~6M samples)

Upstream data-quality issues — concept URIs in source data that no TTL declares:

• `opencontext/material/0.1/organicanimalproduct`
• `opencontext/material/0.1/plantmaterial`
• `vocabulary/specimentype/1.0/othersolidobject` (declared as `msot:othersolidobject` in the wrong namespace)
• `vocabulary/specimentype/1.0/physicalspecimen` (likewise)

Not fixable in this artifact — flagged for #148 as design debt.

Run / deps

```
python scripts/build_vocab_labels.py [-o out.parquet] [--also-csv]
```

Deps: `rdflib pandas pyarrow`. Not run in site CI; this is a one-shot artifact builder.

Test plan

• `python scripts/build_vocab_labels.py -o /tmp/v.parquet` succeeds with no warnings beyond the documented dedup.
• Sanity-check resolved labels for representative facet URIs (`Animalia`, `Fungi`, `Whole organism material sample`, `Past human occupation site`, `Gaseous material`).
• Coverage check against current wide parquet: 51/55 URIs, >99.99% sample-weighted.

🤖 Generated with Claude Code

[rdhyee]

Both Codex findings addressed in 30b3c81:

P2 #1 — partial-fetch failures producing a successful artifact
Default behavior is now fail-loud (exit 3) if any TTL source fails to fetch/parse. Since this artifact is intended for publishing, silently emitting a partial parquet is the wrong default. Added --allow-partial for explicit override (useful during local debugging).

Verified by injecting a bogus URL into VOCAB_TTLS:

• without flag → exit 3, error listed on stderr
• with --allow-partial → exit 0, parquet written

P2 #2 — script deps not in repo environment
Added scripts/requirements.txt pinning rdflib>=6.3, pandas>=2.0, pyarrow>=14. Kept separate from the site-build ../requirements.txt (the scripts here are one-shot artifact builders, not run in CI — pulling them into every Quarto render would be noise). Updated docstring to reference it.

Run:

pip install -r scripts/requirements.txt
python scripts/build_vocab_labels.py -o vocab_labels.parquet

Thanks for the review.