feat: add Dataflow Gen2 data visuals skill for Claude and ChatGPT

[Miguel Escobar (ptyx507x)]

Adds client-side knowledge for generating Dataflow Gen2 visualization documents in Power Query M. Visual is the canonical artifact name; dashboard, report, and chart are discovery synonyms.

Skills

• claude-skills/datafactory-data-visuals.md (Claude)
• chatgpt-skills/knowledge-data-visuals.md (ChatGPT)
• docs/datafactory/data-visuals.md (in-repo agents, wired into the copilot-instructions.md routing table)

Both client indexes route visual/dashboard/report/chart/KPI requests, and claude-skills/SKILL.md gains four visual symptom-triage rows.

Content follows the published contract: flat five-column table (Name, Parent, PartType, Properties, Data), the closed 11-value PartType set, one-root/parent/cardinality rules, and Preview limitations. Safety guidance covers metadata-only discovery, no dynamic sibling evaluation, execute_query for non-persistent testing, and validateOnly plus connection rebinding around save_dataflow_definition.

Evals

• EVAL-INT-M-021..023: five-column contract, chart-in-card column mappings, unsupported PartTypes
• evals/dataflow-visuals-parity.md: provider-neutral checklist for comparing Claude and ChatGPT output

Eval runner fixes

These are pre-existing bugs found while wiring up the new evals:

1. SKILLS_DIR pointed at a nonexistent path. Path(__file__).parent.parent resolved to evals/claude-skills, so the with-skills arm never loaded any skill and every integration eval compared baseline against baseline.
2. Encoding. read_text() used the OS locale, so on Windows the arrow separator mangled and skill names failed to parse. Now UTF-8, with UTF-8 stdout to avoid UnicodeEncodeError on legacy consoles.
3. datafactory-destinations mapped to a file that does not exist; that content lives in destinations/ subfiles. SKILL_FILES values are lists now.

Also corrected count drift in evals/README.md (pipelines 15 to 18; totals 97 + 23 = 120).

Verification

python evals/run_evals.py --dry-run
Dry run complete. 97 scenarios parsed.

python evals/integration/run_integration_evals.py --dry-run
Dry run complete. 23 scenarios, 99 total rules.

All six skills confirmed loading real content (core 3053, destinations 10407, performance 2533, advanced 4280, pipelines 6576, data-visuals 3096 chars).

Notes

No .csproj or NuGet packaging changes. These are client-side companion files; installing the MCP NuGet package does not register skills with Claude or ChatGPT.

Reviewers may prefer the three eval-runner fixes split into a separate PR, since they touch shared infrastructure affecting all skills rather than visuals alone. Happy to split if preferred.

[Codecov Comments Bot (codecov-commenter)]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 21.88%. Comparing base (5cefbdc) to head (e626a4f).
⚠️ Report is 48 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #117      +/-   ##
==========================================
+ Coverage   14.85%   21.88%   +7.02%     
==========================================
  Files         129      160      +31     
  Lines        3884     4908    +1024     
  Branches      398      461      +63     
==========================================
+ Hits          577     1074     +497     
- Misses       3295     3812     +517     
- Partials       12       22      +10     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.

[Mihir Wagle (mavaali)]

I traced the validation failure to the workflow’s PR-results comment step, not the tests. All restores, builds, integration tests, test reporting, and coverage passed. Because this PR comes from a fork, GitHub downgrades GITHUB_TOKEN write permissions for the pull_request workflow, so actions/github-script fails with Resource not accessible by integration.

The validated fix is to remove the unsupported include-prerelease input, always publish body through core.summary, and skip the PR-comment API call when the PR head repository differs from the base repository. I couldn’t push the patch directly because the authenticated mavaali account has read-only access to both the base repository and this fork branch.