fix: the audit list is bounded; metrics count in SQL

CHANGELOG.md

@@ -17,6 +17,10 @@ a decision with visible consequences.
   call refused), evidence processing stored an empty `local_ai_proposal` analysis, so the
   evidence counted as analysed and was never analysed again. Now nothing is stored.
 
+- **The audit list is bounded.** `GET /audit` returned every row ever written; it now returns the
+  newest `limit` rows (default 500, at most 2000). `/audit/metrics` counts with GROUP BY instead of
+  loading every row of the window.
+
 - **The conflict check reads whole words and keeps the owner's words in recall.** It counted
   "not" inside "knot" or "cannot " and matched long words as substrings over the first eight
   only. Now it compares whole words ("not", "never", "no longer", "stopped", "cannot", "-n't"

services/api/app/routes/audit.py

@@ -3,19 +3,25 @@
 from app.core.db import SessionLocal
 from app.models.audit import MemoryAudit
 from fastapi import APIRouter
-from sqlalchemy import select
+from sqlalchemy import func, select
 
 router = APIRouter(prefix="/audit", tags=["audit"])
 
 READ_ACTIONS = {"runtime_context", "runtime_ask", "read"}
 WRITE_ACTIONS = {"write", "upgrade", "edit", "delete", "transcript_write", "transcript_delete"}
 
+DEFAULT_LIMIT = 500
+MAX_LIMIT = 2000
+
+
 @router.get("")
-def list_audit():
+def list_audit(limit: int = DEFAULT_LIMIT):
+    """The newest audit rows, at most `limit` (1..2000): the table grows with every read."""
     db = SessionLocal()
     try:
         rows = db.execute(
-            select(MemoryAudit).order_by(MemoryAudit.created_at.desc())
+            select(MemoryAudit).order_by(MemoryAudit.created_at.desc(), MemoryAudit.id.desc())
+            .limit(max(1, min(limit, MAX_LIMIT)))
         ).scalars().all()
 
         return [
@@ -38,19 +44,16 @@ def audit_metrics(hours: int = 24):
     cutoff = datetime.now(UTC) - timedelta(hours=bounded_hours)
     db = SessionLocal()
     try:
-        rows = db.execute(
-            select(MemoryAudit.action, MemoryAudit.created_at).where(MemoryAudit.created_at >= cutoff)
+        counts = db.execute(
+            select(MemoryAudit.action, func.count()).where(MemoryAudit.created_at >= cutoff)
+            .group_by(MemoryAudit.action)
         ).all()
         by_action: dict[str, int] = {}
-        reads = 0
-        writes = 0
-        for action, _created_at in rows:
+        for action, count in counts:
             key = str(action or "unknown")
-            by_action[key] = by_action.get(key, 0) + 1
-            if key in READ_ACTIONS:
-                reads += 1
-            if key in WRITE_ACTIONS:
-                writes += 1
+            by_action[key] = by_action.get(key, 0) + count
+        reads = sum(count for key, count in by_action.items() if key in READ_ACTIONS)
+        writes = sum(count for key, count in by_action.items() if key in WRITE_ACTIONS)
         return {
             "window_hours": bounded_hours,
             "reads": reads,

services/api/tests/test_audit_list_bounded.py

@@ -0,0 +1,16 @@
+"""The audit list is bounded: it returns the newest rows, at most `limit`."""
+
+from app.models.audit import MemoryAudit
+from memory_stack import stack as stack
+
+
+def test_the_audit_list_returns_at_most_the_limit_newest_first(stack):
+    with stack.sessions() as db:
+        db.add_all([MemoryAudit(id=f"a{n:03d}", action="read", payload_json="{}") for n in range(30)])
+        db.commit()
+
+    assert len(stack.get("/audit?limit=10").json()) == 10
+    assert len(stack.get("/audit?limit=0").json()) == 1
+    assert len(stack.get("/audit").json()) == 30
+    metrics = stack.get("/audit/metrics").json()
+    assert metrics["by_action"]["read"] == 30 and metrics["reads"] == 30