[wberri13]

Summary

• Add parallel.ai-judge.decide as a direct Parallel Decisions catalog tool for one or many named noul, choice, and score questions over shared state.
• Preserve Parallel's raw answer metadata, token usage, and shared citations through treg's normal relay. Reuse team-owned Parallel credentials without charging the treg balance.
• Record Parallel's documented $0.001-per-fast-request price, while gating platform-funded calls until shared-key billing and settlement are live verified.
• Add a sanitized synthetic fixture, CLI examples, focused relay tests, and a verification ledger.

Verification

• Parallel API reference, question types, limits, rate limits, and pricing checked on 2026-10-30.
• Mocked Decisions tests cover single and multiple questions, request/response preservation, 401/422/429/503 responses, and own-key precedence with shared-key gating.
• uv run --frozen --with pytest-xdist pytest -n auto -q --tb=short: 6,888 passed, 132 skipped.
• uv run --frozen python scripts/catalog_validate.py parallel: 0 errors.
• uv run --frozen lint-imports and focused Ruff check: passed.

Activation requirement

The fixture and tests use mocked upstream responses. No live Parallel Decisions call or billing observation was made. Before enabling treg-funded access, verify the platform key can call Decisions, observe account charges and error billing, and confirm the settlement rule. Until then the catalog entry is team-key only.

Context fragments

Upstream now keeps provider-specific catalog details in the catalog row and examples, so this PR leaves docs/context/architecture/catalog.md unchanged after syncing with main. The Decisions recipe and verification ledger remain in examples/parallel-decisions.md.

Closes #1285.