hermes - 💡(How to fix) Fix Verify retrieval fix on fresh unrelated queries

Summary

Live verification after the #103 junk-classification fix and #104 relevance-ranking fix shows the system is behaving correctly on fresh, unrelated queries.

What was tested

I intentionally avoided the same old repeat queries and used different prompts to check whether retrieval was still overfitting to previous topics.

Fresh query set

1. Portfolio site / readme-roulette
2. GCP billing / trial
3. HP Linux Mint SSH
4. SimpleFin Chase cards
5. Nonsense query: quantum blockchain NFT parrot recipes

Observed behavior

1) Topic-relevant queries now return topic-relevant context

• Portfolio / readme-roulette returned the deployed web app context.
• GCP billing / trial returned profile-level GCP context.
• HP Linux Mint SSH returned the wiped media-server / local-network context.
• SimpleFin Chase cards returned the card / transaction context.

2) Irrelevant queries degrade gracefully

• The nonsense query did not return some stale default topic.
• It reported fallback_reason: "no_topic_relevance_candidates".
• lexical: 0
• relevance_signals: []
• current_state_only: true
• current_state_only_penalty: 0.35

3) Ranking is no longer dominated by the most recent write

Before the fix, unrelated queries repeatedly surfaced the same recent VPN / JioHotstar facts. That failure mode did not appear in this test run.

Signals seen in the new retrieval path

These diagnostics were visible in the latest results:

• lexical_hits
• lexical_overlap
• keyword_hits
• relevance_signals
• current_state_bias
• current_state_only_penalty
• fallback_reason

The key behavior change is that lexical relevance is now acting as a primary signal instead of recency/current-state dominating everything.

Verification result

• Junk classification stays fixed: spam-like input remains quarantined / forensic-only.
• Retrieval ranking fix appears stable on fresh, unrelated queries.
• Graceful fallback for no-match queries is working.

Remaining note

The profile still contains some overrepresented VPN-era facts from the earlier test session, but that is a profiling / consolidation issue rather than a ranking failure.

Why this matters

This is the difference between a memory system and a glorified last-write cache. The system now appears to answer the actual query instead of whatever it saw most recently.

Suggested follow-up

If further tightening is needed, the next thing to inspect is profile consolidation / topic clustering so repeated historical topics do not remain overrepresented in the static profile view.