Complete Agent Models workspace follow-ups

[crstdr]

Summary

• make the Setup Wizard compute and apply Agent Models recommendations against the active global/workspace profile, reject stale targets, and label the affected scopes truthfully
• expose and clear every stored role pin consistently, and validate synchronized Oracle/Compose profiles at the durable settings boundary
• distinguish blocked future/incompatible settings schemas from corruption in telemetry handling, with focused regression coverage and persistence documentation
• emit canonical notifications for every durable scope actually changed, including global MCP Presets combined with workspace Agent Models updates

Current-main integration

This follow-up is based directly on current main after #534. It preserves #534's AgentModelsOperationIdentity, stale-action revision guards, exact-scope mutation boundaries, and service-tier normalization.

Validation

• Oracle final review: APPROVE
• Fable 5 final review: APPROVE
• RecommendationWizardScopedTargetTests: 5/5 passed
• AutoRecommendationEngineScopedSettingsTests: 8/8 passed
• MCPAgentRoleDefaultsServiceTests: 9/9 passed
• SettingsJSONOnlyPersistenceTests: 75/75 passed
• exact test ledger verification: 3,257 IDs
• make dev-swift-build PRODUCT=RepoPrompt: passed
• full make dev-test: one unrelated durable-artifact GC timing test failed; isolated suite rerun passed 9/9
• targeted SwiftFormat/SwiftLint on touched files: passed
• commit and push contribution preflights: passed
• hosted GitHub CI: all required checks passed; shard 3 passed on rerun

Repo-wide lint remains blocked by existing SwiftFormat baseline findings in untouched files.

Closes #508