Mock-driven supervisor state machine tests

[ImpulseB23]

Follow-up to PRI-13. Current supervisor test coverage is only the pure functions (next_backoff, SidecarStatus serialization, default config constants). The state machine itself — spawn → run → terminate → backoff → respawn, healthy-uptime backoff reset — is only validated by manual E2E.

Scope

Introduce a SidecarLauncher trait inside sidecar_supervisor.rs:

trait SidecarLauncher: Send + Sync + 'static {
    async fn launch(&self) -> Result<LaunchedChild, io::Error>;
}

struct LaunchedChild {
    events: tokio::sync::mpsc::Receiver<CommandEvent>,
    writer: Box<dyn StdinWrite + Send>,
}

Production impl wraps app.shell().sidecar(SIDECAR_BINARY)?.spawn() as today. Test impl is a scripted launcher that:

• Emits a configurable sequence of CommandEvents into the rx channel
• Terminates on cue to drive the respawn branch
• Records bootstrap + twitch_connect writes so tests can assert them

Tests to add:

• Healthy run resets backoff: seed a launcher that stays alive past the healthy threshold, then terminates → next run_once starts at initial_backoff
• Unhealthy run grows backoff: seed a launcher that terminates immediately twice → second backoff is 2× first
• Bootstrap + twitch_connect replay on every respawn
• sidecar_status events are emitted in the expected order per iteration

Out of scope

• Real cross-process respawn (this lives in manual E2E and the release smoke test)
• Mocking the ring buffer — drain loop stays out of these tests; focus is the supervisor lifecycle

Depends on

• PRI-13 landing on main