Design: reconcile neuroscience framing (CREB/LTP/LTD) with actual implementation

[raphasouthall]

Summary

`docs/neuroscience-appendix.md:5-13` invokes CREB excitability windows (~6 hours), engram recruitment, and Hebbian dynamics. The current implementation is effectively a read-counter with frozen flags: decay is mathematically unreachable on realistic usage spans, the "active" flag is seeded once from YAML and never refreshed, LTD is absent, and engram competition isn't modelled.

Evidence

• Decay cliff unreachable. Hotness formula: `sigmoid(log1p(count)) * exp(-ln2/30d * age_days)` with 30-day half-life (`search.py:323-354`). Threshold for dormant is 0.05. A note used once needs `age_days > 130` to fall below — but earliest `note_usage.used_at` is 21 days old. Dormant-by-score is mechanically impossible until ~late July 2026.
• Static CREB flag. `note_metadata.status` is seeded from YAML frontmatter at migration (`schema.py:706`) and only changed by `run_excitability_demotion`. It doesn't refresh on read. The "6-hour CREB window" from the appendix is not implemented.
• LTD absent. Decay exists but doesn't run (see separate issue on auto-decay). Co-occurrence weights have no decay (see separate issue on co-occurrence noise).
• Hebbian only on graph path. `graph.py:263` records usage for neighbours. Semantic / FTS hits don't propagate activation.
• No engram competition. The boost is uniform 1.15× for all active notes.

Proposed fix

1. Calibrate half-life to observable history. Drop default `half_life_days` from 30 to ~7 (or expose in `config.toml`). With 7-day half-life and 0.05 threshold, a note used once goes dormant after ~30 days — matches actual usage envelope. Location: `search.py:323, 357, 397, 455` (hard-coded).
2. Implement a real CREB-like window. Replace static `note_metadata.status` boost with a computed float. On every `note_usage` insert, set `note_metadata.excitability_expires_at = now + 6h`; apply the 1.15× boost only while unexpired (`search.py:710-740`). Effect: recently-read notes genuinely attract linking — matches the appendix claim.
3. Hebbian on all retrieval paths. `search.py:838, 1051-1110` records only the returned notes. Lift `graph.py:263`'s pattern: on semantic/FTS hits, also record a smaller-weight co-activation for 1-hop graph neighbours (via a `weight` column on `note_usage`).
4. Engram competition. Instead of uniform 1.15×, scale boost by `1 / log(active_neighbour_count + e)` so dense hubs compete for retrieval mass rather than dominating.
5. Purge/bucket `note_usage`. Currently `ORDER BY used_at DESC` on every hotness call (`search.py:335`). Bucket to one row per (note, day) with a counter. Effect: O(1) hotness computation, bounded storage.

Expected effect

The neuroscience framing becomes falsifiable: decay demonstrably fires, recent reads demonstrably boost, Hebbian propagation demonstrably changes retrieval. The project's core claim ships as code, not appendix.

Key files

• `src/neurostack/search.py:323-506, 710-740, 838, 1051-1110`
• `src/neurostack/schema.py:175-182, 258-268, 488-501, 690-725`
• `src/neurostack/graph.py:263`
• `docs/neuroscience-appendix.md:5-13`

Relationship to other issues

Depends on the auto-decay timer issue and the co-occurrence LTD issue — those fix the mechanics; this issue fixes the semantics.

[raphasouthall]

Acceptance criteria

• `half_life_days` exposed in `config.toml` (no longer hard-coded at `search.py:323, 357, 397, 455`); default calibrated to observable usage history (~7 days)
• `note_metadata.excitability_expires_at` column exists and is refreshed to `now + 6h` on every `note_usage` insert
• 1.15× active boost at `search.py:710-740` applies only while `now < excitability_expires_at`
• Semantic/FTS hits record smaller-weight co-activation for 1-hop graph neighbours (not just graph-path retrievals as today)
• `note_usage` bucketed to one row per (note, day) with a counter column; `hotness_score` complexity drops from O(N) per call
• Boost scales inversely with neighbour density (engram competition) — dense hubs don't monopolise retrieval mass
• `docs/neuroscience-appendix.md` updated to reflect implementation (or removed if claims remain aspirational)

Dependencies

• Depends on: Bug: excitability decay never auto-runs; hidden side-effect in vault_stats #31 (decay must run for CREB window + half-life calibration to be falsifiable); Bug: vault_graph / vault_related / vault_summary don't record note usage #32 (Hebbian propagation needs complete usage-recording surface).
• Co-design with: Improvement: co-occurrence table is 62% of DB and 82% noise #34 (co-occurrence LTD should share the decay half-life/scheduling pattern).

Sequencing

Phase 4 (rework). Largest-scope issue — depends on most others. Don't start until #31, #32, #34 have merged.