feat(models): load chat model selector from server

[nutshellfool]

Summary

Replace the hardcoded ModelPresets.list chat model selector with the full model list from the configured OpenCode server's GET /config/providers API.

• AppState.selectedModel: Message.ModelInfo? is now the canonical selection state; the list index is derived for UI only, so server-side list reordering can never silently switch models
• Model catalog (availableModels + providerModelsIndex) is built once per provider load via buildModelCatalog instead of computed getters on every recomposition
• Selection persists by providerId/modelId globally and per session; legacy index-based storage is read-only for migration (resolution order: session saved → message inference → global saved → server default → legacy → first model)
• Host/profile changes funnel through a single configureRepository choke point that clears the catalog only when the effective target actually changes, so same-profile foreground health checks keep the loaded catalog; a request-generation guard drops late provider responses from a previous server
• buildSelectedModel validates the selection against the current catalog before sending, and omits the model (server default) while providers are loading

Notes

• Rebased onto master @ v0.1.20260715: kept session deep links / AI usage quota / hostRuntimeScope intact; selectedAIUsageQuota now reads selectedModel?.providerId; GPT-5.6 Sol Pro/Fast short labels ported to the dynamic modelLabel logic; resetRuntimeForHostSwitch also clears the new catalog fields
• Went through three review rounds (correctness/efficiency/design); plan doc included at docs/dynamic-opencode-model-selector-plan.md

Test plan

• ./gradlew testDebugUnitTest — BUILD SUCCESSFUL after rebase
• New coverage: provider-load selection resolution, cold-start inference after providers arrive late, same-profile reconnect keeping the catalog, stale provider response rejection, legacy preset migration

🤖 Generated with Claude Code