v3.18.1 — neural status: native ruvllm training path no longer misreported as Unavailable (#2549)

Patch release fixing #2549 (reported by @pacphi — exemplary report with dist-verified line numbers and a capability-vs-environment proof).

Fixed

• neural status reported the bundled @ruvector/ruvllm training path as Unavailable — Install @ruvector/ruvllm even though the module was installed and functional. Two defects: a dead _trainingBackend variable (declared, returned, never assigned) and contrastive availability read only from an in-process global a fresh status process never populates.
• Backend now comes from a capability probe (resolveTrainingBackend() — module resolution, pipeline stays lazy); Contrastive Trainer has three honest states (Active with session counts / Available — ready on demand / genuinely Unavailable); the Install hint only shows on true resolution failure.
• Dropped the 'ruvllm checkpoints enabled' claim — upstream saveCheckpoint() (@ruvector/ruvllm 2.5.6) verifiably writes no file; tracked separately.

Not in this release (follow-ups from #2549 triage)

• Routing neural train through the native TrainingPipeline
• Upstream checkpoint persistence fix

PR: #2554 · Regression test pins the capability contract in both resolve/no-resolve environments.