improve skill: better consolidation coverage, richer scripts, workspace layout

[joe32140]

Summary

Improvements to the /openaireview skill based on analysis of review quality across multiple methods (skill, progressive Opus/Gemini). The key finding: the consolidation step was over-pruning — 61 raw sub-agent comments collapsed to 18, dropping legitimate singleton findings. This PR fixes that while also improving the underlying scripts.

SKILL.md

Step 4a — Singleton protection:
Read singleton findings (appearing in only one sub-agent) in full before dropping — these are most likely to be unique insights, not noise.

Step 4b — Smarter merging:
When merging a root-cause cluster, check whether any comment makes a distinct argument (different evidence, different claim, different consequence). If so, keep it separate. Goal is to eliminate true duplicates, not compress distinct observations that share a design decision.

Step 4c — Calibration floor:

• Major ceiling: 8 → 10 (papers with many independent validity threats may legitimately exceed 7)
• Added a 15–25 total comment floor: if fewer than 15 comments survive dedup, revisit for over-merging

consolidate_comments.py

Richer stdout output — now prints the first sentence of each explanation and a source-count column alongside each title. Makes singleton findings visible at a glance without requiring a full-text fetch.

prepare_workspace.py

Prefer reviewer.parsers (BeautifulSoup + Marker) when the package is available; fall back to the stdlib-only ArXiv extractor. Outputs workspace to review_results/<slug>_review/, writes metadata.json.

save_viz_json.py / criteria.md / subagent_templates.md

Minor fixes and wording refinements.

Impact

Evaluated on paper 2602.18458 (MechEvalAgent). Before: 18 final comments. After: 26 — recovering 8 findings that sub-agents had found but consolidation dropped, including >80% concealing C3/GT3 outliers, C3 as AND-logic artifact, three-trial GT budget unjustified, and impact statement gaps.

Test plan

• Singleton findings visible in consolidation stdout without full-text fetch
• Consolidated output reaches 15–25 comments on a typical paper
• prepare_workspace.py falls back gracefully when reviewer.parsers is absent