[Saurabh Ravindra Badenkal (saurabhrb)]

Skill guidance and routing fixes for failures that recur in the live evaluations. Each change traces a repeated failure to skill guidance, not to the platform. This PR now also includes the changes from #158 (closed in favor of this one) so they are reviewed and merged together.

Why

Recurring failures in the live evaluations, with the cause in skill guidance:

1. Currency columns created as Decimal (metadata_001_table_and_simple_columns). dv-metadata listed "money" as a supported type and used it in an example, but the SDK maps "money" to DecimalAttributeMetadata. Org verification then fails with "expected column exma_estimatedcost with type Money, got ... Decimal". This was still the most common run failure in the first week of November.
2. Record association through raw HTTP (data_008_associate_contact). dv-data listed association as an SDK gap with a raw POST .../$ref alternative. The agent guessed a client.records.associate method, got AttributeError, then fell back to the SDK's private _odata helper or a hand-built requests POST, which the grader marks non-compliant.
3. Skill not loaded. connect_001 / connect_002 (verify an existing connection or that DATAVERSE_URL matches the live org) answered correctly but never loaded dv-connect: dv-overview sends "connect" and "set up" requests to dv-connect but tells the agent to skip setup when auth already works, and "verify the connection" matched only the skip path. overview_001 / overview_003 ask which skills to use without saying "Dataverse", so dv-overview did not load; on overview_003 the agent also skipped the connect and solution steps.
4. Activity state and partial re-runs (data_005_support_system). The agent created a phone call already closed with a status reason from another state, then re-ran the whole script after the failure and duplicated a company.
5. dv-admin OrgDB reads. A request for named OrgDB settings could leave out a setting that was never written.
6. ERP routing (PR-gate dry-run suites). erp_data_001_update_contract, erp_admin_001_list_running and erp_metadata_001_runtime_schema answered from templates in dv-overview's references/erp-target.md without loading the owning skill.

Changes

dv-data

• No raw $ref; states there is no client.records.associate; new "Link existing records (lookup)" example with client.records.update + parentcustomerid_account@odata.bind (CLI: dataverse data associate).
• New references/activities.md: bind the regarding row with the per-table navigation property, create the activity open, then close it with one update that sets a matching statecode + statuscode pair.
• references/multi-table-fk-import.md: if a create script fails partway, look up what already exists and create only what is missing instead of re-running.

dv-metadata

• "money" documented as a "decimal" alias; the constraint example uses "decimal".
• New references/currency-columns.md: create a real MoneyAttributeMetadata column with dataverse api request, including the solution header and a verify step.

dv-connect / dv-overview

• dv-connect description: verifying which environment or identity is connected (WhoAmI, DATAVERSE_URL match).
• dv-overview description: requests for which skill, tool, or sequence of steps to use; "verify connection" requests load dv-connect.
• dv-overview change lifecycle: states the skill sequence for "new table to source control" (dv-connect -> dv-solution publisher and solution -> dv-metadata -> dv-solution export and unpack). The duplicated publisher/solution snippet is shortened to a pointer (dv-solution owns the full pattern) to stay under the 5,000-token body cap.
• dv-overview/references/erp-target.md is a router again: reads and runtime schema go to dv-query, writes including previews to dv-data, batch jobs to dv-admin. Every removed template still exists in those references. Also corrects "DMF (tier 5 above)" to tier 6.

dv-admin / dv-query

• dv-admin/references/orgdb-settings.md: "read specific OrgDB settings" example that checks requested keys against the documented allowlist, reads only those keys, reports an absent key as not set (platform default), and never writes back.
• dv-admin description names Finance and Operations / F&O for batch jobs; dv-query skill boundaries send ERP writes to dv-data and batch jobs to dv-admin.

Version 1.17.0 -> 1.18.0 across all manifests (minor: new references and trigger keywords).

Testing

• python .github/evals/static_checks.py: passed (9 skill files, 51 Python blocks); all bodies stay under the 5,000-token cap.
• python .github/evals/version_bump_check.py: passed.

Live validation

The combined changes were run on a proof branch (users/saurabhrb/evalarm-skills-155-158) against a live test environment on 10/17:

• 45 of 46 live tests passed and none failed; 604/604 P1 assertions and all 50 org-state checks passed. connect_002 produced no output in the evaluation harness (infrastructure), so it was not graded in that run.
• metadata_001 created a real Money column; data_008, connect_001 and overview_003 passed; data_005 scored effort 5/5 (3/5 before) with exactly 3 companies.
• dv-admin 5/5 tests passed; the OrgDB settings read scored 5/5 on every grader.
• PR-gate dry run: 8/8 suites green, 15/15 tests; each ERP test now loads its owning skill.

The dv-overview skill-sequence line was added after that run; the live dry-run gate on this PR covers it.