feat(retrieval): genuinely multi-cortical recall — link channels via content-cued spreading activation

## Summary

Recall today fuses three independent channels (vector + keyword + graph) with RRF and then a cross-encoder rerank, but it does **not** genuinely *combine cortical regions* the way the README's brain-inspired framing implies. The "graph" channel matches query tokens against tag nodes, expands one hop, and surfaces the result as a UI-side `related[]` list — it never participates in ranking, and there is no spreading activation, pattern completion, or region-aware combination.

This issue proposes making recall **genuinely multi-cortical**: let activation flow *across* channels so that evidence found in one region (e.g. a vector hit) can cue and strengthen related memories in another region (graph neighbors, same-partition siblings), more closely mirroring how cortical areas co-activate during retrieval.

## Current state (as shipped)

- `MemorySearcher.search` runs three parallel paths — **vector** (embedding similarity), **keyword** (FTS5 / tsvector), **graph** (tag match → neighbor expansion) — fused with RRF (k=60), then optionally cross-encoder reranked.
- The graph path is **tag-keyed and content-blind**: `search_tags` does token-wise matching of the query against tag ids/labels, traversal is fixed depth-1, and the expanded neighbors feed a `related[]` list rather than the ranked results. When it does contribute, hits carry a static frequency-derived score, not a content-relevance score.
- Channels are **siloed**: a strong vector hit cannot cue its graph neighbors or same-partition siblings into the candidate pool. There is no cross-channel activation.
- The knowledge graph is only populated by **consolidation**, so raw-write deployments have an empty/sparse graph — the "multi-cortical" path is effectively inert there.

Net: the "multi-cortical recall / CA3 cell-assembly" language is aspirational relative to what the pipeline actually does. The Hebbian/CA3 mechanics that *are* real live in the consolidation + forgetting loop, not in recall.

## Proposed feature — link the cortices

Add a genuine cross-region activation stage, all behind **default-OFF flags** and **A/B-gated** (see constraints):

1. **Content-cued spreading activation** as a real 4th RRF channel.
Seed activation from the *tags of the top vector + keyword hits* (not a query substring), then propagate over weighted graph edges with distance decay (e.g. `activation *= decay ** hops`). Collect the memories on activated nodes as a ranked channel that joins the RRF fusion — so a vector hit can pull in its graph-connected relatives.
2. **CA3-style pattern completion** — when a subset of an assembly (co-occurring tags / co-consolidated cluster) is activated, boost the remaining members so a partial cue completes the pattern.
3. **CLS / region-aware recall** — use the existing partition plumbing to recall per partition-group and combine, instead of pooling all partitions undifferentiated (the search router currently never sets `partition_ids`).
4. **Populate the graph on raw writes** (lightweight tag edges), so the spreading-activation channel is non-empty without requiring consolidation first.

## Constraints & acceptance criteria

- Every new mechanism ships behind a **default-OFF flag**; shipped default behaviour stays bit-for-bit until proven.
- **Eval per house rules**: full dataset (no train/test split), isolated server on ports 8401–8409 with its own `hebb.db`, `.venv/bin/python`, run with rerank **ON and OFF**, and include **multi-hop** MemBench slices (single-hop recall is near-saturated, so the lever should show up on multi-hop).
- **Ship-gate**: ≥ +1pp headline or ≥ +2pp multi-hop vs. the current default; record the in-tree `run-N` report before any number lands in `repo_pages/benchmarks/`.
- Honest-README correction can land independently and immediately (zero code risk): scope the CA3/Hebbian language to consolidation, describe recall as hybrid RRF + rerank until the multi-cortical channel proves out.

## Risks / open questions

- Tag graph may be **too sparse** (< ~1.5 edges/node) for spreading activation to fire — measure edge density first; consider embedding-kNN edges as a denser substrate.
- **Wash risk**: rerank is the dominant retrieval lever and already recovers in-pool answers; LongMemEval R@10 is saturated (~0.994). The new channel must earn its keep on multi-hop, not on saturated single-hop.
- RRF normalization when a channel is ablated needs a look (the per-hit cap can under-normalize a missing channel).
- Latency / cost budget for the extra traversal + re-fusion.

## References

- Internal audit (2026-06-07): "multi-cortical recall is a pseudo-claim as shipped."
- `src/hebb/retrieval/searcher.py` — three-path recall + RRF + rerank.
- `src/hebb/graph/knowledge_graph.py` — tag graph, `search_tags`, neighbor expansion.