fix(modeler): dedup a multi-active satellite's CDK from its attributes

[mischa76]

Summary

Re-running the health_insurance demo to refresh the walkthrough figures exposed that it failed validation 4/4: every run tripped E_SAT_DUP_ATTR on sat_insured_person_address. The modeler listed the multi-active satellite's child dependent key (address_type) also among its attributes; attributes + child_dependent_key share one column namespace (src_payload + src_cdk), so the generated satellite would emit that column twice and cannot build. The re-model loop couldn't recover within MAX_MODELING_ATTEMPTS.

Fix

Both, since LLM steering alone failed 4/4 even with the error fed back through the retry loop:

• A [GUIDE] line in DV_MODELING_RULES: a multi-active CDK is a key column, not payload — never also list it among the attributes.
• Deterministic rules.attributes_without_cdk(attributes, child_dependent_key), applied by dv2_modeler._validate_model — drops payload attributes normalising to a CDK label (order-preserving; the CDK column still ships via src_cdk). Meaning-preserving — genuine attribute-vs-attribute duplicates are left for E_SAT_DUP_ATTR to flag.

Verification

• Health demo now PASSES 3/3 (0 issues, was 0/4). A representative run: 4 hubs / 3 links / 8 satellites → 15 raw-vault + 8 staging models, matching the walkthrough.
• 333 tests green (+2: attributes_without_cdk unit + modeler-dedup integration), ruff clean, uv run mypy clean (32 files).
• Refreshed docs/demos/health-insurance-walkthrough.md (staging count in the figures).

Note

A separate, sporadic gate (E_SAT_ATTR_OVERLAP on hub_policy) appeared once pre-fix and not in the post-fix runs — a different, lower-frequency modeler-quality issue, not addressed here.

🤖 Generated with Claude Code