refactor: deduplicate recording artifacts and use JPEG thumbnails (#65)

- Remove docs/artifacts/full/ (was a copy of waa_recordings/ PNGs)
- Thumbnails now link to originals in waa_recordings/ for full-res
- Switch thumbnails from PNG to JPEG (1.5 MB vs 3.0 MB for same images)
- Un-gitignore waa_recordings/ (research data, should be tracked)
- Gitignore docs/artifacts/full/ instead (regenerable)
- Untrack benchmark_results/ (mock test output, already gitignored)
- Move os import to module level in generate_demo_review.py

Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>

Author: abrichr

.gitignore

@@ -30,6 +30,8 @@ benchmark_live.json
 # Cost reports (generated during evaluation runs)
 cost_report.json
 
-# WAA recordings (PNGs + meta.json from record-waa sessions)
-# Gitignored to protect from `git stash -u` and `git clean -f`
-waa_recordings/
+# WAA recordings are tracked in git (research data).
+# External backup to ~/oa/recordings/ protects against git stash -u.
+
+# Generated artifacts from generate_demo_review.py (regenerable)
+docs/artifacts/full/