Phase 2: LarqlRunner — health and readiness events

[ttupper92618]

Parent: #173
Phase: 2 — LARQL runner + MoE sweet spot
Depends on: #157, #158
Blocks: #163, #164, #165

Goal

Expose LarqlRunner readiness and serving metadata through Skulk state.

Scope

• Health-check the supervised LARQL child via its readiness endpoint.
• Emit readiness metadata for served preset, layer range, expert range, vindex URI, and measured RAM footprint.
• Wire the readiness event through apply() so master-side placement can target available cold-tier runners.

Acceptance criteria

• Readiness metadata is strict, serializable, and replay-safe.
• A not-ready or failed LARQL child is not eligible for slice placement.
• Dashboard/API consumers can distinguish ordinary MLX runner status from LARQL slice readiness without breaking existing status handling.

Validation

• Unit tests for readiness parsing and event application.
• Worker planning tests cover ready vs not-ready LarqlRunner eligibility.

[ttupper92618]

Implemented in PR #174: #174

This adds replay-safe LarqlRunnerReadinessUpdated state carrying runner id, vindex URI, preset, served layer/expert metadata, address, status, RAM footprint when available, and cleanup on runner shutdown or instance deletion. Existing RunnerStatus handling remains separate.