docs(deploy): connect the evaluation guide to the Policies & Decisions docs

[ghchinoy]

Follow-up to #40: the evaluation guide only linked the policy docs in one sentence.

docs/deploy/evaluate.md

• §5.1 Design your policy (was "Write a policy"): a "what to read" table:

  • applications + decision matrix (does the use case fit?)
  • template catalog + creator's guide (start from an existing policy)
  • first policy (step by step)
  • authoring §2 (fast and reliable)
  • datasets step 2 (co-design with MCP decide_custom_questions)

  It also lists the policy rules that matter for an evaluation: fixed schema with per-row state, samples: 1, the 26-option limit, depends_on/ask_if costing a second pass, and evaluating the policy you will ship.

• §5.4: maps each error pattern to its fix:

  • errors cluster on one question → reword the options (authoring, first policy §3)
  • inputs fit no option → taxonomy discovery
  • too many hesitant answers → Stage 2 cascade. Noted that it calls Gemini on Vertex AI, so a self-hosted evaluation needs a GCP project (cmd/cascade_gemini.go).
  • need an error bar per decision → samples: 4

  Run the test set once with the final policy and any cascade enabled.

• Step 5 label in the overview table updated; Next links the full Policies & Decisions set.

Link-backs

• policies/first-policy.md "Where to go next" → evaluation guide.
• policies/datasets.md intro: callout for self-hosters to verify the install first, and to use a local gateway with the runners.

Checks

make docs-sync-check (44/44), docs_link_check.py (0 broken) and make check-public pass. The docs-site build passes (45 pages), and all 11 in-page anchors on the three changed pages resolve in the built HTML.