Role defaults: default orchestrator and worker agent+model

[liam-russell]

Problem

The orchestration flow (#134) has two distinct roles with opposite economics: the planner wants the strongest model available (reads the whole repo, decomposes the task, reviews merges), while workers want something fast and cheap per subtask. Today every launch would make the user re-pick agent and model each time — and the fan-out dialog (#137) would start blank for its most common case.

Proposal

Two role-based defaults in Settings:

• Default orchestrator: agent (from the roster, Agent roster + ad-hoc ACP agents: configure any ACP-speaking CLI, not just known presets #143) + model (from the searchable picker, AI provider registry + universal searchable model picker (OpenRouter, OpenAI-compatible, Ollama) #144) — e.g. Claude Code + Fable/Opus.
• Default worker: agent + model — e.g. Claude Code + Sonnet.

Consumed wherever a role is implied:

• "Start a big task" (Orchestration worktrees: give the planner agent a checkout and a branch to integrate into #134) launches the planner with the orchestrator default.
• Fan-out (Task fan-out: create worker worktrees with pre-seeded agent sessions in one action #137) prefills every subtask with the worker default; per-subtask overrides still apply.
• The MCP fan-out tool (Wire SproutGit's MCP server into every integrated agent session #145) uses the worker default when the planner doesn't specify one.

Scope notes:

• Per-workspace override on top of the global default (a monorepo at work and a hobby project may want different economics).
• For ACP agents the model is applied via session/set_config_option after session start; if the agent doesn't expose a model option, the model half of the default is ignored gracefully (agent-side config wins).
• Roles stay implicit — a plain single-worktree agent launch keeps using the plain default agent from Agent roster + ad-hoc ACP agents: configure any ACP-speaking CLI, not just known presets #143; only orchestration surfaces consult these two.