Restructure the audiobook-curator example as a composed React application - #315

Summary

The flagship framework-mode example previously funneled every MCP tool through one 40-line switch producing a single Agent.Text sentence, and 14 of 15 CLI commands never rendered at all. This PR restructures examples/audiobook-curator into a genuinely composed React application that demonstrates the framework's thesis by construction.

• src/components/ — a shared, typed component library (CuratorDocument, DataList/Field/Callout, AudiobookCard, LibraryShelf, CandidateRanking, EvidenceTrail, MutationReceipt, ChapterOutline, IntegrityReport, CurationShelf, async LibraryAnalysis) reused across MCP tools, the catalog resource, the curate prompt, and rendered CLI commands. src/result.tsx (the switch) is deleted.
• src/providers/library.ts — request-time ffmpeg/ffprobe probing and workflow stages, consumed by the catalog resource via (await agent()).providers.library with honest degraded rendering.
• src/state.ts — workspace-durable curation shelf (editionSelected / mutationApplied / shelfCleared); select/apply routes dispatch and render the shelf; new read-only review_curation_shelf tool and rendered shelf CLI command make persistence visible across tool calls.
• Suspense streaming — audit_library (and CLI library-audit) render an immediate shell and stream the duplicate/multipart analysis behind <Suspense fallback={<Agent.Progress/>}>; tests/route-unit/streaming.test.ts proves staged progress through the real MCP projector (renderRouteEvents + projectTargetCapabilities).
• Rendered CLI — inventory, select, audible-search, convert, audit promoted to .tsx, reusing the same components as their MCP twins; --json output remains the identical canonical receipt (parity proven in tests/route-unit/cli-rendered.test.ts). Plain .ts commands remain to demonstrate both modes.
• README rewritten as the application tour; flagship Workbench e2e pins updated (16 tools, 32 CLI commands, 50 routes, visible provider:library row) with re-captured 1440×900 screenshots.

Route graph: 15 tools + 30 commands + 47 routes → 16 tools + 32 commands + 50 routes. Receipts, schemas, and canonical --json output are unchanged; no framework/package source changed (examples + one workbench test file), so no changeset.

Composition gaps found during the work were filed instead of hacked around: #312 (no shared layout convention) and #313 (providers do not mount for routed CLI/rendered scripts).

Test plan

• AGENT_BUNDLE_TEST_TIME_SCALE=2 pnpm examples:check (all examples) green post-rebase
• Example suite: 33 tests + 30 route-unit tests green (includes streaming projector proof, provider injection, state persistence, CLI receipt parity)
• pnpm typecheck and pnpm lint green
• pnpm test:examples:browser 5/5 in real Chrome at 1440×900, covering populated + stale-diagnostic + repair flows with re-captured screenshots

Refs #107.