Add Benchmarking section; rewrite semantic-scholar description as example

[jbragg]

Summary

Adds a Benchmarking section to README documenting how to run Asta skills via agent-baselines' inspect-swe solver against Inspect-compatible eval suites, including paired comparisons for measuring skill changes. The semantic-scholar rewrite below serves as the worked example referenced from that section.

Rewrites skills/semantic-scholar/SKILL.md's description: field. The current description is a long list of trigger-phrase strings ("get paper details", "look up a paper", "find citations", ...) that the agent must match against the user's prompt. It misses single-fact queries about named research artifacts (benchmarks, datasets, models, methods) that don't use one of the literal trigger phrases, and it doesn't mention snippet-search (paper-body text search) at all.

Replace with a two-sentence form — capability first, then use-when:

Look up or search papers, authors, citations, and full-text snippets on Semantic Scholar.
Use for fast, targeted queries about a paper, author, or specific named research artifact
(benchmark, dataset, model, method, etc.) — not comprehensive reports.

What changes:

• Surfaces snippet-search (full-text body search) as a first-class capability so the agent routes single-paper fact lookups here instead of dropping to PDF parsing.
• Adds research artifact (benchmark / dataset / model / method / etc.) as a trigger object so the skill activates on operationally-named-item queries (e.g. "what does SVAMP stand for", "does OLMES recommend N in-context examples") that the literal-phrase trigger list misses.
• Keeps the original "Use this for fast, targeted queries (not comprehensive reports)" qualifier verbatim — preserves the differentiation from the find-literature skill's comprehensive-search lane.

Validation

Paired comparison on the asta_multitool eval, four cases that operationally reference a named research artifact. claude_code 2.1.128 · sonnet-4-6 · ghcr.io/allenai/asta:v0.17.0 · authed (ASTA_TOKEN set) · no working-limit. n=3 per case per arm.

case	prompt	baseline	this PR	ablation: drop "about X"
cbbd1475	SVAMP acronym	0.00	0.33	0.00
fee5ee2a	AI2 reading-comprehension dataset recall	0.00	0.67	0.50
1c13faea	OLMES in-context-examples recommendation	1.00	1.00	1.00
9ad75f7d	LLaMA acronym	1.00	1.00	1.00
mean		0.50	0.75 (+0.25)	0.625

Mechanism: baseline samples on SVAMP (cbbd1475) and AI2-dataset (fee5ee2a) routinely activate the wrong skill (find-literature, or none at all) and drop to web-search/PDF fallback. This PR routes them to semantic-scholar → asta papers search / snippet-search — visible in the per-sample skill activations.

Ablation isolates the trigger-object clause ("about a paper, author, or specific named research artifact (...)"). Dropping it (third column) loses most of the AI2-dataset routing fix (fee5ee2a: 0.67 → 0.50) and all of the SVAMP score lift (cbbd1475: 0.33 → 0.00), confirming that phrase is what's doing the work — not the broader rephrasing or the new snippet-search mention.

(OLMES + LLaMA are at ceiling on baseline — no headroom for this change to demonstrate.)

Companion PRs

• allenai/agent-baselines#26 — the resolver + reproducibility metadata that made this paired comparison measurable.
• allenai/asta-bench-private#225 — adds an inspect_ai entry-points registration for astabench/ai2/evals/ tasks so the multitool can be loaded by registered name (astabench/asta_multitool_challenge) from agent-baselines' solvers/inspect-swe subproject via uv run --no-group astabench --with ../asta-bench-private ....

[rodneykinney]

Neat!