Add OpenCode and Mistral Vibe agent backends

[simbo1905]

Summary Of Draft State

Adds two optional agent backends:

• opencode, with its provider/model selection passed as ordinary role arguments.
• vibe, using Mistral Vibe's authenticated programmatic interface in the assigned worktree: --trust --agent accept-edits --output streaming -p <role prompt>.

Vibe role arguments pass through before -p, allowing bounded smoke configurations to provide --max-turns, --max-tokens, and --max-price. The backend never adds --auto-approve or --yolo by default.

Also included:

• focused launcher tests;
• an opt-in, smoke-only role instruction profile for bounded diagnostics;
• SWARMFORGE_CONFIG selection of an alternate project-local configuration; and
• a handoff-daemon detachment fix so the daemon follows its explicit stop-file lifecycle rather than a terminal bridge.

Closes #39.

Validation

• bb test in the mountless aarch64 Lima guest: 27 tests / 104 assertions passed.
• Direct bounded Vibe programmatic request passed with streaming output and an exact marker response.

Known limitation

This is intentionally not presented as a successful Vibe end-to-end handoff smoke. Vibe's programmatic streaming invocation exits after responding. A callback attempt reaches Vibe's approval gate but is immediately denied in this mode, so end-to-end delivery to a Vibe receiver has not been proven. The PR establishes and tests the safe launch contract; it does not claim a benchmark result or a persistent Vibe receiver workflow.

Scope and security

No credentials, user Vibe configuration, project-specific runtime shim, NOOA material, telemetry, or benchmark artifacts are included. Vibe must be installed and authenticated by the runtime owner.

[simbo1905]

Hey @unclebob

This one is a little half baked so far as is making some arbitrary decisions such as it picking run this open model to keep my token costs down while I am trying things out myself

"--model opencode-go/qwen3.7-plus --auto"

I think there is an case to be made that I split out mistral.ai as its a mono provider backend; so it has its own standard model defaults. So that can probably get onto main faster.

On the question of what is the default provider for OpenCode (Go at $10 a month, or Zen with pay-as-you-go), and the default models. There are Deepseek, Kimi and others hosted in US, EU and Singapore on OpenCode. That makes things more complex but a ton of fun https://opencode.ai/docs/zen/#pricing

It could be fun to come up with some swarm-forge benchmark tasks and to then different mix and matches of open and proprietary models perhaps a "Swarm Bench"...

Give me a thumbs up if you want me to pursue this.

Simon