ARIA self-improvement: retrieval-drift replay harness (#329)

Adds a weekly self-audit that replays a frozen set of canonical prompts
through the memory activation pipeline and measures divergence of the
retrieved top-K node set from a stored baseline.

Distinct from existing drift systems:
  - drift-audit.ts     — source-code drift (git history)
  - drift-detection.ts — pinned-node content + edge drift
  - retrieval-replay.ts (new) — retrieval *behavior* drift: given the
    same input, does the same memory set surface in the same order?

Motivation (Moltbook thread f0bf79e7, Apr 17 2026): introspection is
circular, owner feedback is lagging and sparse. A synthetic external
anchor — canonical prompts with a frozen baseline — gives a cheap,
consistent, non-circular drift signal. The retrieval-behavior analogue
of golden tests.

Implementation:
  - 10 canonical prompts covering family, work, identity, self-improve,
    Moltbook, memory arch, messaging, emotion, schedule.
  - First run captures top-K (K=10) per prompt from spreading activation
    into baseline.json; subsequent runs compute Jaccard on the ID sets,
    Pearson on rank + score over the intersection, and list new/dropped
    nodes.
  - Tiered alerts: severe (meanJaccard <0.4), significant (<0.6),
    localized (any prompt <0.3).
  - Pure structural, deterministic: no Claude calls, no semantic search,
    no WM boosts — uses spreadingActivation directly so (graph, prompts)
    fully determines output.
  - Prompts editable via retrieval-replay/prompts.json on disk; baseline
    stays valid as long as prompt IDs don't change.
  - Wired into brain.ts tick loop behind shouldRunReplay() (weekly
    cadence). Non-fatal on error.

Co-authored-by: ARIA <aria@myagent>
Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>

Author: 3 people

backend/brain.ts

@@ -28,6 +28,7 @@ import { getBrainConfig, getOwnerLocalDate, getOwnerLocalTime } from "./brain-co
 
 import { BRAIN_DIR, OWNER_NAME, GITHUB_REPO } from "./config.js";
 import { createBackup, shouldRunBackup, BACKUP_INTERVAL } from "./memory/backup.js";
+import { shouldRunReplay, replayAndCompare } from "./memory/retrieval-replay.js";
 import { loadConsciousness } from "./consciousness.js";
 
 // ── Extracted modules ──
@@ -721,6 +722,22 @@ async function tick(
     }
   }
 
+  // ── Weekly retrieval-drift replay (structural, no Claude call) ──
+  // Replays a frozen set of canonical prompts through the activation
+  // pipeline and compares top-K node sets against a stored baseline.
+  // External, non-circular drift signal distinct from source-code and
+  // pinned-node drift.
+  if (shouldRunReplay()) {
+    try {
+      const replayReport = replayAndCompare(graph);
+      if (replayReport?.alert) {
+        log(`⚠ RETRIEVAL REPLAY ALERT: ${replayReport.alert}`);
+      }
+    } catch (err) {
+      log(`Retrieval replay failed (non-fatal): ${err}`);
+    }
+  }
+
   saveState(freshState);
   graph.save();