[Microsoft.Extensions.AI.Evaluation.Reporting] Evaluation report redesign

[grafanaKibana]

Closes #7593.

Summary

Redesigns the generated AI Evaluation HTML report (the React app under Microsoft.Extensions.AI.Evaluation.Reporting/TypeScript) to the FluentUI 2 mockup agreed in #7593. It's a visual/UX refresh — the same Dataset / ScenarioRunResult / EvaluationMetric data is read through ReportContext, with no data or API changes.

The old single-page layout is replaced by four focused, tab-based views driven by a shared app shell:

• Overview — headline KPIs (pass rate, cases failing, scenarios fully passing, good ratings), biggest movers vs. the previous run, a "needs attention" panel, and pass-rate-by-scenario-group.
• Cases — searchable/taggable case list with inline drill-down (transcript + per-metric panel + diagnostics), pagination, "show failed" filter, and a scenario sort control.
• History — per-metric trend chart and dumbbell deltas across runs.
• Comparison — side-by-side A/B comparison of two runs, per scenario/metric.

The same components drive both the standalone report and the Azure DevOps extension tab.

What changed

Restructured components/ into feature folders (previously a flat directory):

Folder	Contents
shell/	app frame — App, AppShell (top bar + pivot), SidebarTree, ViewRouter, theming (theme.ts / theme.css), ADO host-resize hook
core/	data/state — ReportContext, Summary, viewModels, scoring
overview/, cases/, history/	the four views + their pieces (TranscriptBlock, MetricPanel, TrendChart, dumbbellGeometry, …)
styles/	shared Griffel makeStyles slots
index.ts	public barrel; internal files don't import through it

• Removed 14 superseded components (App, MetricCard, PassFailBar, ScenarioTree, ScoreNodeHistory, TagsDisplay, …) and the old flat ReportContext/Summary/Styles.
• Styling moved onto FluentUI 2 design tokens (a theme.css token layer + per-component Griffel slots), with light/dark parity and an acrylic material for elevated surfaces.
• Added html-report/gen-devdata.js — a deterministic dev-data generator so the report can be run and reviewed locally with realistic multi-run data.

New dependencies

• react-markdown + remark-gfm — markdown/GFM rendering in transcripts and reasons.
• Dev/test only: vitest, @testing-library/react, @testing-library/jest-dom, jsdom, @vitest/coverage-v8.

Testing

Adds a test suite (didn't exist before) — 87 tests across 11 files (vitest + Testing Library), covering the load-bearing behavior the redesign must preserve: chronological run ordering, movers/comparison ordering, per-execution re-scoping, the "New" badge, transcript rendering, ADO theme/persistKey, and the Cases sort/filter.

Verification

• npm run build (tsc composite + vite build) — clean; single-file bundle ~860 kB.
• npm test — 87/87 pass.
• eslint — 0 errors (2 inherent react-refresh warnings on the context/entry files).
• Manually verified both light and dark themes across all four tabs.

Risk

As called out in #7593, the risk is visual-regression in the shared components, since the same React app drives both the standalone report and the Azure DevOps extension — both were exercised. No API or data-shape changes.

Microsoft Reviewers: Open in CodeFlow

[grafanaKibana]

@dotnet-policy-service agree