AI-as-Infrastructure/aiinfra-atlas: Inter-rater upgrade

This release prepares ATLAS for the inter-rater reliability focus group study (15 paid reviewers × 20 ratings each, 100 prompts).

Highlights since v0.2.5:

• Inter-rater session seeding (make seed / seed-dry / seed-reset) with 100 curated 1901 parliamentary prompts across AU/NZ/UK/all corpora.
• Count-aware allocator (sort by inter_rater_count ASC, sha256 tiebreak)
  • 60 s per-user cache: delivers the gold-standard 1:1 design of exactly 3 independent ratings per prompt under full attendance.
• INTER_RATER_DEFAULT_UI mode: reviewers land directly on the inter-rater workflow with the New Session button and privacy toggle hidden.
• Per-scale comment fields on all 6 Likert scales, mandatory for extreme ratings (1, 2, 5).
• INTER_RATER_SESSIONS_PER_USER bumped 10 -> 20 for the focus-group workload.
• Security: python-jose -> PyJWT migration, FastAPI security headers, prod-gated debug endpoint, dependency bumps for python-multipart, python-dotenv, langchain-text-splitters, langchain-openai, pytest.
• Documentation: canonical 'Default Focus-Group Study Configuration' section with a 10-item researcher checklist; seed commands in README.