feat(arxiv): full abstract/authors + surface pdf/categories/comment + new recent <category>

[jackwener]

What

Three changes to the arxiv adapter, all agent-facing data quality / coverage gaps I hit while exploring:

1. paper <id> — stop silently truncating the data

The Atom feed for 1706.03762 has 8 authors and a ~200-word abstract. Old behavior:

• Abstract was cut at 200 chars + … (silent — caller had no signal)
• Authors were .slice(0, 3).join(', ') (also silent)

New behavior: full abstract, full author list. Plus surface fields the API was already returning that we threw away:

• pdf (from <link rel="related" type="application/pdf">, with id-based fallback)
• primary_category (e.g. cs.CL)
• categories (all <category term> entries comma-joined)
• comment (e.g. "15 pages, 5 figures", conference info, etc.)
• updated

2. search <query> — minor cleanup

• Add primary_category to columns (cheap signal of paper area)
• Switch the no-results path from CliError('NOT_FOUND', …) to EmptyResultError('arxiv', …) to match the convention other public-API adapters use (matches the lesson from PR fix(douban): drop unparseable fields from movie-hot, add id/votes #1285).

3. recent <category> — new

You couldn't ask "what's new in cs.CL today?" without inventing a search term. Now:

opencli arxiv recent cs.CL --limit 10
opencli arxiv recent math.PR --limit 5

Uses search_query=cat:<category>&sortBy=submittedDate&sortOrder=descending. Validates the category string against the arxiv <archive>.<sub> shape (also accepts archive-only like math or q-bio). Bad input → ArgumentError.

Why

I'm an agent that calls these adapters. With the old paper, an agent fetching "Attention Is All You Need" got 4 of 8 authors and a clipped first paragraph — silently. That's a worse failure mode than an outright error: caller has no idea data is missing.

recent is just an obvious gap — arxiv search "anything" was the only way to read the Atom feed.

Tested

• clis/arxiv/arxiv.test.js: 4 tests covering command registration, full-abstract / all-authors parsing, XML entity decoding in titles, pdf/categories/comment extraction, and category validation.
• Live runs:
  • opencli arxiv paper 1706.03762 → all 8 authors, full abstract, primary_category=cs.CL, pdf URL present.
  • opencli arxiv recent cs.CL --limit 3 → 3 newest cs.CL papers, sorted by submitted date desc.
  • opencli arxiv search "attention is all you need" --limit 3 → 3 hits with primary_category surfaced.
  • opencli arxiv recent "BAD..STR" → ARGUMENT error (exit 2).