fix(storage): stop swallowing the neuron_state write, and reclaim orphans explicitly

[acidkill]

Closes #174.

What

Two defects in opposite directions, both silent.

1. A swallowed write reported success (#174)

add_neuron wrote the neuron, then its neuron_state row inside a bare
except: pass. A failed state write still returned the neuron id, so afterwards
"the state was written" and "the state write failed" were indistinguishable.

This matters more than a missing row. neuron_state holds access_frequency and
activation_level, so a neuron without one looks permanently un-accessed: it never
receives the activation boost, and it is a standing candidate for dead-neuron pruning —
for a reason nothing logged.

Both paths now log and re-raise. The single path additionally rolls the neuron back, so a
caller's retry does not collide with a half-written pair. The batch path does not unwind —
its neurons are already written in bulk, and a partial unwind is a worse outcome than a
loud failure the caller can retry.

2. Nothing reclaimed orphaned states

Both delete paths clean up their own state today, so surviving orphans are historical. They
are not inert: apply_decay iterates get_all_neuron_states(), so each orphan is decay
computed for a neuron that no longer exists, and it inflates any per-pass count taken from
that loop — which would directly corrupt the decay telemetry proposed in #183.

Reclaiming is explicit, not automatic

smem prune-orphan-states            # report only (default)
smem prune-orphan-states --apply    # delete

Deliberately not a consolidation stage. The detection compares neuron ids against
neuron_state.neuron_id, and the two sides spell the same neuron differently:

neuron.id              →  neuron:0001182a_b6f6_...   (_to_surreal_id folds - to _)
neuron_state.neuron_id →  0001182a-b6f6-...          (original dashed uuid)

Comparing them unnormalised marks every live state as an orphan and empties the table.
Both sides now go through the same fold, and a regression test asserts it. Given that
failure mode, the operation defaults to reporting and runs only when someone asks for it —
it is not something that should execute unattended on a schedule.

count_orphaned_neuron_states exists so the destructive path can be inspected before it is
taken; both it and the delete share one detection function, so they cannot drift apart.

Verification

• full suite: 7260 passed, 0 failed (an earlier run showed one unrelated failure in
  test_unified_config; it passes in isolation on this branch and on main, and did not
  reproduce on re-run — cross-test state, not a regression here)
• lint / format clean
• dry-run exercised against a live database: reported orphans, deleted nothing, and neuron
  state coverage was verified unchanged afterwards