v3.9.4 — remote recall takes the indexed path

3.9.4 — remote recall takes the indexed path

This patch removes the main latency bottlenecks found when SurrealDB and inference services run across the network.

Fixed

• Small neuron-state batches use deterministic record ids with bounded concurrency. Larger consolidation batches retain the set-based query to limit request fan-out.
• Fiber expansion runs independent anchor lookups concurrently while preserving input order and per-anchor limits.
• Existing databases gain idx_fiber_neurons over neuron_ids.* and brain_id; membership queries explicitly select it.
• Exact-content neuron lookups explicitly select idx_neuron_content instead of allowing the optimizer to choose the slower type index.
• instant recall now stops after one hop, and other fallback paths honor the requested depth.
• CLI freshness collection fetches matched fibers concurrently.

Measured result

On the production remote topology, smem recall improved from roughly 23–40 seconds to 7.14 seconds wall time, with the retrieval phase measured at 3.12 seconds. A fresh MCP process completed initialize and smem_recall in 9.60 seconds, including startup.

Changes

• #239 — reduce remote recall latency