v2.18.2 — Honest consolidation docs

No behavior change. Documentation and contributor-process only.

The README described consolidation as something a command does

Memory consolidation — episodic memories mature into semantic knowledge was the entire
description, which reads as though smem consolidate performs the maturation. It does not,
and 2.18.0 removed that claim from the code's own remedy text without updating the README
that still implied it. The entry now states the actual mechanism: maturation happens through
spaced recall (7 days dwell plus reinforcement across 3+ distinct days, or 15+ rehearsals
across 5+ windows), names the two smem health fields that show where each memory sits and
what it is waiting on, and points at brain.reinforcement_neuron_limit for widening how many
of a recall's memories get rehearsed.

distill_llm_load_cmd (2.18.0) was missing from the reasoning-training example entirely, so
the documented config could only unload a model it never controlled the loading of.