v2.18.0 — Maturation reachability

2.16.0 made the consolidation report honest about what it counts. It did not make
consolidation_ratio reachable. This release closes the gap between the metric and the
mechanism that is supposed to move it: five separate defects that each, independently,
kept fibers from ever reaching SEMANTIC or kept the report from saying why.

The remedy stopped recommending the wrong command

smem_health's low-consolidation_ratio penalty, smem_stats' low-consolidation hint, and
the NO_CONSOLIDATION warning all told the operator to run smem consolidate to raise the
ratio. Nothing about consolidate moves a fiber toward SEMANTIC — only spaced recall does
(reinforcement spread across 3+ distinct days, or 15+ rehearsals across 5+ time windows).
All three surfaces now name the actual mechanism and say explicitly that consolidate will
not move it on its own.

Recall's rehearsal step was capped at a hardcoded 10, twice over

Every recall reinforces the fibers connected to its top-activated neurons — the mechanism
spaced recall depends on. Two independent, redundant caps limited this to the first 10
neurons regardless of how many a recall actually activated: one in the retrieval pipeline's
neuron ranking, a second, wholly redundant one inside the reinforcement step itself on the
fibers that made it through the first cap. A recall that activated 50 neurons still only
rehearsed fibers connected to 10 of them, silently discarding the rest of that recall's
contribution to the spacing requirement.

Both caps are fixed: the redundant fiber-side cap is removed, and the neuron-side cap is now
a configurable brain.reinforcement_neuron_limit (default 15, up from the old
unconfigurable 10) in config.toml. The default was set from a live measurement against the
production storage backend rather than assumed — raising this value is not free on every
backend, and an installation whose storage can absorb more latency can raise it.

Merging fibers used to erase their progress toward SEMANTIC

Consolidation's merge strategy deletes the source fibers and writes one new fiber in their
place. It never carried the sources' maturation records forward, so a fiber one reinforcement
away from SEMANTIC that got merged came back at the beginning: no maturation row at all,
stage STM. The merged fiber now inherits the higher of the sources' stages, the union of
their reinforcement timestamps (not a plain concatenation — a fiber reinforced on the same
calendar day by both sources counts as one distinct day, not two), and the oldest of their
stage-entry timestamps, so a source that had already dwelt in a stage far longer than its
merge partner does not have that dwell time reset to "now."

The consolidation report and health diagnostics explain themselves now

• The report's flat Stages advanced: N is now followed by a breakdown of exactly which
  hop advanced (stm→working, working→episodic, episodic→semantic); the three hop
  counts sum to the same total, so nothing about the existing field changes for anything
  already parsing it.
• smem_health now includes a stage_distribution (how many fibers sit at each maturation
  stage) and a semantic_gate_blockers breakdown of every EPISODIC fiber by what is actually
  blocking it from SEMANTIC: waiting on dwell time, waiting on reinforcement spacing, or
  already eligible and waiting for the next consolidation pass. Found and fixed along the
  way: smem evolution's own "closest to semantic" classifier duplicated this threshold
  logic locally and checked only the distinct-days path, so a fiber that qualified through
  the rehearsal-count-and-windows alternative was misreported as still blocked. Both
  diagnostics now share one classifier.
• A consolidation run's own backfill of maturation rows for fibers that predate the
  maturation subsystem is now a visible line in the report summary, not a number that only
  existed in a field nothing printed.

Distillation can now explicitly load its model before the first request

reasoning_training.distill_llm_unload_cmd (2.17.0) explicitly releases a distillation run's
chat model when the run ends. Loading was still implicit — whatever happens on that
endpoint's first request — so a run had no way to control how the model got loaded: with
what context size, how many GPU layers, or with a projector it did not need. A new,
symmetric distill_llm_load_cmd runs once, before the first naming request, with the same
argv-only execution (no shell, {model} substitution, empty or failing command degrades
silently to the old implicit behavior). The unload command still fires even if a run that
explicitly loaded ends up naming nothing, closing the leak that would otherwise leave the
model resident.

Fixed

• A configured distill_llm_model is sanitized against a looser character set than the
  distill_llm_load_cmd/distill_llm_unload_cmd argv template it gets substituted into; the
  substituted result is now re-validated against the stricter argv allowlist before
  execution, closing a defense-in-depth gap an independent security review found.
• A distill_llm_load_cmd/distill_llm_unload_cmd longer than the configured part limit is
  now voided outright, consistent with every other invalid-command case, rather than
  silently truncated to a shorter, different command.
• SurrealDBStorage.clear() never deleted maturation rows, leaving them orphaned after a
  brain's fibers and neurons were otherwise fully cleared.