cdeust/Cortex

Addresses issue #13 (darval) in full

The telemetry patch in v3.10.1 (commit `69d81fb`) gave per-stage observability. This release is the substantive follow-up: the five pathologies identified by the agent+genius audit of darval's 35-min run on a 66K-memory store.

Phase A — Set-based SQL batching (100–500× speedup on decay)

Per-row UPDATE/DELETE loops with per-row commits were the dominant cost. New batch methods on both `PgMemoryStore` and `SqliteMemoryStore`:

• `update_memories_heat_batch` — single `UPDATE … FROM UNNEST` replaces 62K per-row `UPDATE + COMMIT`
• `update_entities_heat_batch`, `archive_entities_batch`
• `update_relationships_weight_batch`, `delete_relationships_batch`
• `insert_stage_transitions_batch`

Converted callers: `decay`, `plasticity`, `pruning`, `cascade`. Erlang analysis predicts the decay stage drops from ~1100 s → 3–10 s on a 66K store.

Phase B — Consolidation-scoped memory cache

`store.get_all_memories_for_decay()` was called 6× per run (decay, compression, memify, homeostatic, sleep, emergence). Now loaded once at handler entry and threaded through every stage that needs it. Stages retain a fallback load when called standalone.

Phase C — Plasticity co-access starvation fixed

Previous `get_hot_memories(limit=50)` sampled 0.5% of a 10k-entity store → 99.95% LTD was distribution collapse, not biology. Sample cap raised 50 → 2000; reuses Phase B cache; precomputes lowercase entity-name index once; exception handler surfaces errors instead of swallowing them.

Phase D — CLS causal-edge signal gate

`_discover_causal_edges` previously hard-capped at 500 episodic on a 25K store (2% sample) and produced 0 edges by construction. Episodic cap raised 500 → 2000; adds an early-exit gate that skips the O(E²) PC pass when too few entities clear the minimum-observations threshold; restricts the PC vocabulary to qualifying entities so the matrix is `E_qualifying²` not `E_all²`.

Phase E — Cascade heartbeat skip + payload trim

Cascade was writing a heartbeat `UPDATE` on every scanned memory (~2000) even when nothing advanced, each with its own commit; and per-transition `INSERT + commit` for stage_transitions (503 fsyncs); and returning the full 503-transition list in the response. Now:

• Heartbeat skipped when `|Δhours| < 1` (the previous value was noise)
• `stage_transitions` INSERTs batched into one statement
• Response payload: `transitions_preview` (first 50) + `transitions_count`
• Exception handler now surfaces errors
• Response adds `scanned / heartbeats_written / heartbeats_skipped` counters

SQLite parity

Cowork mode previously lacked `stage_transitions` / `stage_entered_at` — cascade would error silently. Schema now has the table + column migration so cowork installs get the same cascade behavior as CLI.

Reviewed by

• `code-reviewer` — APPROVE
• `test-engineer` — 2188 tests pass
• `genius:erlang` — queuing theory on the 66K store
• `genius:feinstein` — differential diagnosis of all 6 findings
• `genius:feynman` — independent rederivation from first principles

Breaking changes

None. Stage functions gained an optional `memories=` parameter; all existing callers continue to work. MCP tool schemas unchanged. Response payload adds new fields (non-breaking).

Install

```
/plugin marketplace update
```