Phase 2: LarqlRunner — process supervision

[ttupper92618]

Parent: #173
Phase: 2 — LARQL runner + MoE sweet spot
Depends on: #153, #155
Blocks: #159, #161

Goal

Add LarqlRunner process supervision in isolation, without MLX-side delegation or slice placement.

Scope

• Add a worker runner path that starts larql-server <vindex-path> --port <p> --ffn-only.
• Support MoE expert range arguments when a slice assignment provides them.
• Stream child stdout/stderr into Skulk logging.
• Restart on ordinary crash and fail the worker/runner after a bounded crash loop.
• Report lifecycle through existing event-sourcing and runner-status paths.

Acceptance criteria

• larql-server command construction is deterministic and covered by tests.
• Crash-loop behavior has a clear terminal failure state.
• Child process is terminated on runner shutdown.
• Existing MLX runner behavior is unchanged.

Validation

• Focused unit tests for process spawn, shutdown, restart, and crash-loop failure.
• Existing worker runner lifecycle tests still pass.

[ttupper92618]

Implemented in PR #174: #174

This adds the internal LarqlRunnerSupervisor path: deterministic larql serve command construction, worker-managed lifecycle, readiness polling, stdout/stderr forwarding, bounded restart behavior, clean child termination, and focused supervisor tests. Slice placement remains out of scope.