Bug: vault_graph / vault_related / vault_summary don't record note usage

[raphasouthall]

Summary

Several retrieval paths fetch notes but never log usage. This creates a feedback loop: notes reached only via graph/related/summary navigation stay `never_used` forever, biasing the ranker toward already-used notes.

Estimated impact: 30–50 of the 218 `never_used` notes would move to `active` if these paths recorded usage.

Root cause

`_record_note_usage` is called on hybrid / semantic / keyword / triples retrieval (`search.py:838, 1051-1110`) and inside `graph.py:263` for graph-neighbour co-activation. But the top-level MCP tool paths bypass it:

• `vault_graph` — graph retrieval for the root note doesn't record.
• `vault_related` (`related.py:19-125`) — pure cosine similarity, no usage recording on results.
• `vault_summary` (`tools/search_tools.py:157`) — reads and returns summary, no usage recording.

Proposed fix

Add `_record_note_usage(conn, [note])` at exit of each of:

• `vault_graph` — for the root note and the returned neighbours.
• `vault_related` — for the returned related notes.
• `vault_summary` — for the note whose summary was fetched.

Expected effect

• `never_used` count shrinks by an estimated 30–50.
• Hotness scores more accurately reflect follow-the-link navigation, not just direct search.
• Hebbian reinforcement fires on all retrieval paths, consistent with the design docs' engram-network framing.

Key files

• `src/neurostack/graph.py:263`
• `src/neurostack/related.py:19-125`
• `src/neurostack/tools/search_tools.py:157`

[raphasouthall]

Acceptance criteria

• `vault_graph` calls `_record_note_usage` for the root note AND all returned neighbours
• `vault_related` calls `_record_note_usage` for all returned notes
• `vault_summary` calls `_record_note_usage` for the fetched note
• Integration test per endpoint: call the endpoint, assert a new row in `note_usage`
• After a session that only uses graph/related/summary (no vault_search), `never_used` count decreases proportionally

Dependencies

• Blocks: improves signal quality for Bug: excitability decay never auto-runs; hidden side-effect in vault_stats #31 (more usage data → meaningful decay) and Design: reconcile neuroscience framing (CREB/LTP/LTD) with actual implementation #35 (Hebbian propagation needs complete recording surface).
• Depends on: none. Three small call-site additions.

Sequencing

Phase 1 (quick wins). Among the cheapest fixes — ship with #30.