aibi-dashboards: theme/color guidance, audit fixes, runnable example

[QuentinAmbard]

Summary

A larger polish pass on the dashboard skill, validated end-to-end by deploying the worked example against synthetic data on e2-demo-field-eng and visually inspecting the result.

Live test dashboard: https://e2-demo-field-eng.cloud.databricks.com/dashboardsv3/01f164c9a2fb1700a60fb84c447a262e

What changed

Theme & color (SKILL.md)

• Rename theme section "Theme & Color (always set this — it makes or breaks the dashboard)"
• Palette-design rules: one coherent family per dashboard, semantic colors pinned as literal hex (not palette position), non-categorical widget coloring, pastel-tweak guidance
• Adjacent-stop contrast rule: viewer must distinguish neighboring categories at a glance
• Single-color faded toward white = a quantitative ramp, not a categorical palette
• 8 starter palettes (bare hex, no names) for first-shot color selection
• Fix: positions are 0-indexed (was incorrectly documented as 1-indexed everywhere)
• Default theme matches the live Customer Support warm-sunset palette

Widget docs (audit findings from 4 real dashboards)

• frame.description / showDescription documented
• Value formatting: number-plain, currencyCode, suffix, formatTemplate; MEASURE() + plain "number" format warning
• Pie: label.show
• Symbol-map: bubble-map encouragement, colorRamp modes (scheme enum + custom-sequential as preferred), mark.opacity
• Heatmap: axis.hideTitle
• Forecast-line: 3-CTE (actuals / bridge / forecast) pattern explained, with the explicit column-list UNION ALL that avoids NUM_COLUMNS_MISMATCH

Worked example (4-examples.md)

• Full incident narrative in the header: "Product release 4.1" spike story
• Forecast chart vertical-line annotation at the release date (coral)
• Reopen rate measure outputs 0-1 (was 0-100) so number-percent renders correctly
• queryLines lines use explicit \n (was relying on trailing whitespace)
• Theme matches the live customer-support dashboard
• Symbol-map colorRamp uses family-aligned gradient (coral → mint)
• Semantic colors (SLA breach, critical cells) pinned as literal hex

Runnable verification harness (.tests/)

• generate_test_data.py: databricks-connect script that seeds main.dashboard_skill_test.support_cases with 5000 rows of synthetic data including a 700-row post-release spike around 2026-02-16, biased toward Critical/High priorities + slower resolution + higher reopen — so the dashboard's incident story is grounded in actual data
• dashboard.lvdash.json: portable dashboard JSON (bare FROM support_cases)
• create_dashboard.py: pushes the JSON via Lakeview SDK with sidecar .dashboard_id for idempotent redeploys (avoids slow lakeview.list() on large workspaces)
• README + .gitignore

Test plan

• Generate data: python .tests/generate_test_data.py → 5000 rows written, sanity-check KPIs non-degenerate
• Deploy dashboard: python .tests/create_dashboard.py → dashboard renders all widgets
• Visual inspection on field-eng: all 4 KPIs, sparkline counter, forecast-line with release annotation, symbol map, pivot/heatmap, pie, detail table — theme coherent, story readable
• SQL fixes verified live (the bare-name + UNION column-mismatch errors that surfaced during the first deploy are fixed in this PR)

This pull request and its description were written by Isaac.