Support pi subagents in Loom + Orbit

## Background

pi 0.75 ships a subagent example extension (`packages/coding-agent/examples/extensions/subagent/`) that registers a `subagent` tool. Each invocation spawns a separate `pi` subprocess with its own context window, model, and tool allowlist. Agents are declared as markdown + YAML frontmatter under `~/.pi/agent/agents/*.md` (user) or `.pi/agents/*.md` (project, gated by confirm). Three modes: `single`, `parallel` (max 8, concurrency 4), `chain` (with `{previous}` substitution between steps).

PR earendil-works/pi#4710 (landed in 0.75.4) returns real per-task output to the parent (was truncated to 100-char previews).

## What Loom needs

1. **Vendor or load the tool.** Either include pi's example extension or copy it into `extensions/loom/` so we control discovery paths — e.g. `~/.loom/agents/*.md` and project-local `.loom/agents/*.md`. Lets us ship Galaxy-flavored agents (scout, planner, QC reviewer, …) with sensible tool allowlists.

2. **Orbit rendering.** Pi's `renderCall`/`renderResult` are TUI-only (return `@earendil-works/pi-tui` components). In Orbit, a subagent run arrives on the existing event stream as `tool_call` with `toolName === \"subagent\"` and `details: SubagentDetails` carrying per-child `{agent, task, exitCode, messages, usage}`. No dedicated `subagent_start`/`subagent_end` events — `tool_execution_update` streams in as each child finishes. The renderer switch in `app/src/renderer/app.ts` needs a new case that renders one card per child, probably as an artifact tab so long scout runs don't bury chat.

3. **Galaxy agent presets.** Candidate first-cut agents:
- **scout** — wide search across IWC + Toolshed, locked to `galaxy_search_*`/`galaxy_search_iwc`/`Read`
- **planner** — chains scout results into a draft plan section
- **reviewer** — audits a completed invocation against the plan

## Sharp edges

- **50 KB cap per task return** and subagent stdout doesn't stream tokens to the parent LLM — one shot at end. Long Galaxy exploration will hit the cap. UX needs an escape hatch: have the subagent write a longer report to a file in the analysis dir, return a pointer.
- **Cost / load.** Each subagent is a fresh `pi` subprocess with the configured model. Parallel mode = N× tokens and meaningful local CPU/RAM. Orbit's usage badge should reflect aggregated subagent usage.
- **Tool allowlist per agent** — good fit for Galaxy; lock scouts to discovery tools only.
- **Discovery paths.** Decide: support pi's `~/.pi/agent/agents/` directly, or fork to `~/.loom/agents/` to avoid stepping on other pi installs the user may run.

## Scope suggestion

Two PRs:
1. Loom (brain): vendor the subagent tool with Loom-flavored discovery paths + one or two example Galaxy agents under `extensions/loom/agents/`.
2. Orbit (shell): add the `toolName === \"subagent\"` renderer case + artifact-tab affordance + usage aggregation.

## References

- pi-mono: `packages/coding-agent/examples/extensions/subagent/` (index.ts, agents.ts, README.md)
- earendil-works/pi#4710 (parallel output fix, in 0.75.4)
- pi extension lifecycle: `packages/coding-agent/docs/extensions.md`

Start with packages/coding-agent/examples/extensions/subagent/ (index.ts, agents.ts, README.md) and the pi extension lifecycle docs, then inspect app/src/renderer/app.ts. Define the Loom tool and agent discovery paths, Galaxy presets, and Orbit rendering and usage behavior; done means the proposed Loom and Orbit changes are covered by two reviewable PRs.

Written by the indexing model from the issue text.