sgl-node v1.4.3

sgl-node v1.4.3 — liveness-gated model advertising (anti-ghost)

The node now probes its local inference engine (llama-server /health, short timeout) on every heartbeat and stops advertising its model after two consecutive failed probes, so the grid routes elsewhere instead of dispatching jobs the node can't actually run.

• Fixes the "ghost" failure mode: a node whose engine crashed or ran out of memory mid-run (e.g. a 14B model on a too-small box) kept heartbeating and kept being offered work it silently dropped.
• One transient blip is tolerated; advertising resumes automatically the moment /health recovers (operator restart or engine recovery).
• No config change for operators. The desktop app (v1.2.12+) prompts you to this build and adds a one-tap Restart if the engine stalls.

Full Changelog: v1.4.2...v1.4.3