feat: provider/model selection for personas and runtime-aware env injection

[wpfleger96]

Summary

Decouples provider/model configuration from goose-specific config files. Personas now carry an explicit provider field, and the desktop injects runtime-appropriate env vars when spawning agents. Pack imports filter derived provider/model env keys so stale values cannot shadow the structured fields after UI edits.

Rebased on top of #868 (Wes's Databricks-defaults-env-only PR) — all goose config compat code removed from sprout-agent.

---

Changes

sprout-agent

• Flatten config/ directory back to flat config.rs — removed goose_compat.rs (345 lines of YAML parsing, superseded by sprout-agent: make Databricks defaults env-only #868's env-only approach)

sprout-persona

• resolve.rs: runtime_env_vars() branches on the runtime field to emit runtime-appropriate env var names (SPROUT_AGENT_MODEL/SPROUT_AGENT_PROVIDER for sprout-agent, GOOSE_MODEL/GOOSE_PROVIDER for goose)
• New tests/e2e_env_flow.rs: 5 integration tests covering env var emission for both runtimes

desktop/src-tauri

• types.rs: Add PersonaRecord.provider: Option<String> field
• runtime.rs: Add runtime_metadata_env_vars() — injects model/provider env vars from discovery table metadata. Placed before build_databricks_defaults() and user env vars.
• discovery.rs: model_env_var: Some("SPROUT_AGENT_MODEL") for sprout-agent runtime; stale doc comments updated to reflect vestigial MCP field and scaffolding supports_acp_model_switching field
• migration.rs: Add reconcile_provider_mcp_commands() and migrate_persona_provider_to_runtime()
• lib.rs: Wire reconcile_provider_mcp_commands() into app startup (between reconcile_persona_pack_paths and migrate_persona_provider_to_runtime)
• persona_card.rs: Emit llmProvider in exported persona JSON
• env_vars.rs: Add DERIVED_PROVIDER_MODEL_ENV_KEYS constant, is_derived_provider_model_key(), and filter_derived_provider_model_env_vars() — identifies env keys that are derived from structured fields and must not be persisted
• env_vars/tests.rs: 6 regression tests for the derived-key filter
• personas.rs: import_persona_pack now filters runtime_env_vars through filter_derived_provider_model_env_vars() before persisting to PersonaRecord.env_vars, preventing stale GOOSE_MODEL/GOOSE_PROVIDER values from shadowing structured fields after UI edits
• commands/agents.rs: build_deploy_payload resolves effective_model and effective_provider from the persona's structured fields (with agent record override) and projects them into the deploy JSON

desktop/src (TypeScript)

• PersonaDialog.tsx: Provider selection UI
• personaDialogState.ts + test: provider field round-trip
• tauriPersonas.ts, types.ts: provider field in API types

---

Env Var Precedence

In spawn_agent_child, env vars are set in this order (last write wins):

parent env
→ runtime_metadata_env_vars()   # SPROUT_AGENT_MODEL / SPROUT_AGENT_PROVIDER
→ build_databricks_defaults()   # DATABRICKS_HOST / DATABRICKS_MODEL (Block builds only)
→ persona env_vars              # user-defined, from persona pack (derived keys filtered at import)
→ agent env_vars                # user-defined, per-agent override

User intent (persona/agent env vars) always wins. Build-time Databricks defaults are a fallback for Block employees — OSS builds bake nothing (intentional).

---

Not in scope

• OSS Databricks gap (no DATABRICKS_HOST for OSS users) — intentional, Block-only convenience feature
• System prompt forwarding — not a bug; ACP delivers prompts at the protocol layer