hermes - 💡(How to fix) Fix holographic memory: HRR silently degrades to FTS5 when numpy is missing [2 pull requests]

The problem

The holographic memory plugin advertises compositional retrieval features (probe, related, reason, contradict) via HRR (Holographic Reduced Representation) vector algebra. But when numpy is not installed, every one of these operations silently falls back to plain FTS5 keyword search:

# retrieval.py — every smart method starts with this:
if not hrr._HAS_NUMPY:
    return self.search(entity, category=category, limit=limit)

The system prompt still tells the agent "Active. N facts stored with entity resolution and trust scoring. Use fact_store to search, probe entities, reason across entities" — but probe/reason/related/contradict are all just keyword grep without numpy.

This is a soft, silent dependency that makes the plugin appear fully functional when it is actually running in degraded mode.

Impact

• fact_store(action="probe") returns results ranked by FTS5 token overlap instead of HRR structural similarity
• fact_store(action="reason") which should do compositional intersection ("find facts about BOTH Alice AND Bob") degrades to OR-based keyword search
• contradict returns empty list
• The HRR weighting (30% of the hybrid score) gets zeroed out and redistributed to FTS+Jaccard

A user can run for months thinking holographic memory is working when it is not.

Proposed fix

At minimum:

1. Log a WARNING at plugin init when numpy is unavailable
2. Reflect degraded state in the system prompt block (e.g. "Active but HRR disabled — install numpy for compositional queries")

Ideally: 3. Make numpy a declared dependency in the plugin or at least document it prominently in the README

Discovery

Discovered on a Netcup VPS running Ubuntu 26.04 where the Hermes venv had no pip and numpy was never installed. All 30+ facts had flat trust scores and probe queries returned irrelevant results. Installed numpy, HRR enabled, behavior now correct.