sona: revisit MicroLoRA adaptation when ruvector-sona gains a LoRA import API

[pacphi]

Follow-up from Phase 0d spike (already in `docs/spikes/ruvector-phase0.md`).

Background

`ruvector-sona` 0.1.9 has no public API to import LoRA weights back into a `SonaEngine`. MicroLoRA adaptation also only mutates under multi-step trajectories with reward variance + explicit `engine.flush()`, neither of which we instrument today. Phase 1 shipped ReasoningBank-only as the load-bearing retrieval signal.

If/when upstream adds a `load_lora_state(&LoRAState)` (symmetric with the existing `export_lora_state()`) AND we have intermediate reasoning steps worth recording, revisit:

Scope

1. Upgrade `ruvector-sona` dep.
2. Snapshot LoRA state alongside the existing `sona_state` JSONB column.
3. Restore on boot, log the reload size.
4. Wire multi-step trajectories in the confirm/dismiss handlers so REINFORCE advantage is non-zero.
5. Call `engine.flush()` at a reasonable cadence.
6. Re-run the Phase 0d-style probe to confirm MicroLoRA actually shifts.

Acceptance

• LoRA weights survive a process restart
• `apply_micro_lora` output measurably differs from a fresh engine after N confirmations
• ADR-017 "LoRA adaptation deferred" note replaced with actual behavior

References

• `docs/spikes/ruvector-phase0.md` Phase 0d addendum
• https://crates.io/crates/ruvector-sona (watch for updates)
• PR feat: Tier 2 semantic categorization + SONA-enhanced flow detection #30