apps/api/prisma/migrations/20260731170000_read_events_partitioned/migration.sql

@@ -0,0 +1,76 @@
+-- #224 (ADR 0023): convert read_events to monthly RANGE partitions on
+-- occurred_at. Volume grows unbounded with use; retention then DROPs whole
+-- expired partitions instead of scanning deletes. The primary key gains the
+-- partition column (PostgreSQL requirement); the dedup unique pair
+-- (dedup_key, window_bucket) moves to PER-PARTITION unique indexes — a
+-- partitioned parent cannot carry it without the partition key. A bucket
+-- spanning a month boundary can therefore record one duplicate; documented
+-- in ADR 0023, over-recording is acceptable, gaps are not.
+--
+-- A DEFAULT partition catches rows outside every maintained range, so a
+-- lagging maintenance job can never make classified reads fail (the trail's
+-- hard-failure semantics would otherwise turn an ops miss into an outage).
+
+ALTER TABLE "read_events" RENAME TO "read_events_old";
+ALTER INDEX "read_events_pkey" RENAME TO "read_events_old_pkey";
+ALTER INDEX "read_events_dedup_key_window_bucket_key" RENAME TO "read_events_old_dedup_key";
+ALTER INDEX "read_events_page_id_occurred_at_idx" RENAME TO "read_events_old_page_idx";
+ALTER INDEX "read_events_actor_id_occurred_at_idx" RENAME TO "read_events_old_actor_idx";
+ALTER INDEX "read_events_occurred_at_idx" RENAME TO "read_events_old_at_idx";
+
+CREATE TABLE "read_events" (
+    "id" TEXT NOT NULL,
+    "occurred_at" TIMESTAMP(3) NOT NULL DEFAULT CURRENT_TIMESTAMP,
+    "actor_id" TEXT,
+    "session_key" TEXT NOT NULL,
+    "page_id" TEXT,
+    "pond_id" TEXT NOT NULL,
+    "channel" TEXT NOT NULL,
+    "classification" TEXT NOT NULL,
+    "details" JSONB,
+    "dedup_key" TEXT NOT NULL,
+    "window_bucket" BIGINT NOT NULL,
+    "window_seconds" INTEGER NOT NULL,
+
+    CONSTRAINT "read_events_pkey" PRIMARY KEY ("id", "occurred_at")
+) PARTITION BY RANGE ("occurred_at");
+
+-- Non-unique parent indexes propagate to every partition automatically.
+CREATE INDEX "read_events_page_id_occurred_at_idx" ON "read_events"("page_id", "occurred_at");
+CREATE INDEX "read_events_actor_id_occurred_at_idx" ON "read_events"("actor_id", "occurred_at");
+CREATE INDEX "read_events_occurred_at_idx" ON "read_events"("occurred_at");
+
+-- The safety-net partition, plus the current and the next month — the daily
+-- maintenance job (read-trail-maintenance) keeps creating months ahead and
+-- adds the same per-partition dedup index to each new one.
+CREATE TABLE "read_events_default" PARTITION OF "read_events" DEFAULT;
+CREATE UNIQUE INDEX "read_events_default_dedup_key"
+  ON "read_events_default"("dedup_key", "window_bucket");
+
+DO $$
+DECLARE
+  m DATE;
+  part TEXT;
+BEGIN
+  FOR i IN 0..1 LOOP
+    m := date_trunc('month', now())::date + (i || ' month')::interval;
+    part := 'read_events_y' || to_char(m, 'YYYY') || 'm' || to_char(m, 'MM');
+    EXECUTE format(
+      'CREATE TABLE %I PARTITION OF "read_events" FOR VALUES FROM (%L) TO (%L)',
+      part, m, m + interval '1 month');
+    EXECUTE format(
+      'CREATE UNIQUE INDEX %I ON %I ("dedup_key", "window_bucket")',
+      part || '_dedup_key', part);
+  END LOOP;
+END $$;
+
+INSERT INTO "read_events"
+  ("id", "occurred_at", "actor_id", "session_key", "page_id", "pond_id",
+   "channel", "classification", "details", "dedup_key", "window_bucket",
+   "window_seconds")
+SELECT "id", "occurred_at", "actor_id", "session_key", "page_id", "pond_id",
+       "channel", "classification", "details", "dedup_key", "window_bucket",
+       "window_seconds"
+  FROM "read_events_old";
+
+DROP TABLE "read_events_old";

apps/api/prisma/schema.prisma

@@ -96,8 +96,13 @@ model AuditEntry {
 /// volume, purpose and legal basis all differ. Deliberately WITHOUT foreign
 /// keys: evidence must survive a page purge and a hard user deletion — the
 /// ids stay as recorded (pseudonymous uuids), history is never rewritten.
+///
+/// In migrated databases the table is RANGE-partitioned by `occurred_at`
+/// (monthly, issue #224) — hence the composite id. The dedup unique pair
+/// lives per partition there (a partitioned parent cannot carry it without
+/// the partition key); `db push` test databases get it on the plain table.
 model ReadEvent {
-  id         String   @id @default(uuid())
+  id         String   @default(uuid())
   occurredAt DateTime @default(now()) @map("occurred_at")
   /// Null = anonymous reader (public grant); `sessionKey` still names the
   /// browsing session, so the anonymous marker is explicit, not an accident.
@@ -124,6 +129,7 @@ model ReadEvent {
   /// represents up to this many seconds, so the evidence is not overread.
   windowSeconds Int    @map("window_seconds")
 
+  @@id([id, occurredAt])
   @@unique([dedupKey, windowBucket])
   @@index([pageId, occurredAt])
   @@index([actorId, occurredAt])

apps/api/src/admin/system-admin.controller.ts

@@ -1,9 +1,12 @@
 import { Controller, Get, Param, Post, Query, Req, UseGuards } from '@nestjs/common';
 import {
   auditListQuerySchema,
+  readEventListQuerySchema,
   type AuditListQuery,
   type AuditListView,
   type JobTriggerResult,
+  type ReadEventListQuery,
+  type ReadEventListView,
   type StorageOverviewView,
   type SystemBackupView,
   type SystemJobView,
@@ -45,6 +48,15 @@ export class SystemAdminController {
     return this.system.auditLog(query);
   }
 
+  /** Read-access trail queries (issue #224): "who read page X", "what did
+   * user Y read" — Site-Admin only, like the audit viewer above. */
+  @Get('read-events')
+  async readEvents(
+    @Query(new ZodValidationPipe(readEventListQuerySchema)) query: ReadEventListQuery,
+  ): Promise<ReadEventListView> {
+    return this.system.readEvents(query);
+  }
+
   @Get('storage')
   async storage(): Promise<StorageOverviewView> {
     return this.system.storage();

apps/api/src/admin/system-admin.service.ts

@@ -7,10 +7,13 @@ import {
   AUDIT_PAGE_SIZE,
   BACKUP_FRESH_MAX_AGE_HOURS,
   BACKUP_STATUS_FILE,
+  READ_EVENT_PAGE_SIZE,
   type AuditListQuery,
   type AuditListView,
   type BackupStatus,
   type JobTriggerResult,
+  type ReadEventListQuery,
+  type ReadEventListView,
   type StorageOverviewView,
   type SystemBackupView,
   type SystemJobView,
@@ -160,6 +163,66 @@ export class SystemAdminService {
     };
   }
 
+  /**
+   * The Site-Admin query path over the read-access trail (issue #224,
+   * ADR 0023) — evidence nobody can read is not evidence. Answers "who read
+   * page X" and "what did user Y read" within a period. API-only by design
+   * (no panel yet): the trail is an examiner's tool, not a daily screen —
+   * documented in data-model.md §read_events.
+   */
+  async readEvents(query: ReadEventListQuery): Promise<ReadEventListView> {
+    const where: Prisma.ReadEventWhereInput = {};
+    if (query.pageId) where.pageId = query.pageId;
+    if (query.actor) {
+      const actor = await this.prisma.user.findUnique({ where: { username: query.actor } });
+      // An unknown username matches nothing rather than everything.
+      where.actorId = actor?.id ?? '00000000-0000-0000-0000-000000000000';
+    }
+    if (query.channel) where.channel = query.channel;
+    if (query.from || query.to) {
+      where.occurredAt = {
+        ...(query.from ? { gte: query.from } : {}),
+        ...(query.to ? { lte: query.to } : {}),
+      };
+    }
+
+    const total = await this.prisma.readEvent.count({ where });
+    const pageCount = Math.max(1, Math.ceil(total / READ_EVENT_PAGE_SIZE));
+    const page = Math.min(query.page, pageCount);
+    const events = await this.prisma.readEvent.findMany({
+      where,
+      orderBy: { occurredAt: 'desc' },
+      skip: (page - 1) * READ_EVENT_PAGE_SIZE,
+      take: READ_EVENT_PAGE_SIZE,
+    });
+    // No FK on actor_id (evidence outlives accounts) — resolve what still
+    // exists in one query, show the bare id otherwise.
+    const actorIds = [...new Set(events.map((e) => e.actorId).filter((id): id is string => !!id))];
+    const actors = actorIds.length
+      ? await this.prisma.user.findMany({
+          where: { id: { in: actorIds } },
+          select: { id: true, username: true, displayName: true },
+        })
+      : [];
+    const actorById = new Map(actors.map((a) => [a.id, a]));
+    return {
+      entries: events.map((event) => ({
+        id: event.id,
+        occurredAt: event.occurredAt.toISOString(),
+        actor: event.actorId ? (actorById.get(event.actorId) ?? null) : null,
+        pageId: event.pageId,
+        pondId: event.pondId,
+        channel: event.channel,
+        classification: event.classification,
+        windowSeconds: event.windowSeconds,
+        details: (event.details as Record<string, unknown> | null) ?? null,
+      })),
+      page,
+      pageCount,
+      total,
+    };
+  }
+
   async storage(): Promise<StorageOverviewView> {
     const usages = await this.prisma.pondUsage.findMany({
       where: { pond: { deletedAt: null } },

apps/api/src/audit/audit-actions.ts

@@ -40,6 +40,7 @@ export const AUDIT_EVENTS = {
   'pond.purged': { severity: 'notice' },
   'quota.override_cleared': { severity: 'notice' },
   'quota.override_set': { severity: 'notice' },
+  'read_trail.pruned': { severity: 'info' },
   'settings.changed': { severity: 'notice' },
   'setup.admin_created': { severity: 'notice' },
   'setup.completed': { severity: 'info' },

apps/api/src/audit/audit-catalogue.test.ts

@@ -18,11 +18,12 @@ import { AUDIT_EVENTS } from './audit-actions';
 const doc = readFileSync(join(__dirname, '../../../../docs/architecture/audit-events.md'), 'utf8');
 
 /** Event rows are `| \`ns.event\` | trigger | severity | …` — the dot in the
- * id keeps field-set rows (`msg`, `severity`, …) out of the match. */
+ * id keeps field-set rows (`msg`, `severity`, …) out of the match. The
+ * namespace may carry an underscore since `read_trail.*` (issue #224). */
 function documentedEvents(): Map<string, string> {
   const events = new Map<string, string>();
   for (const line of doc.split('\n')) {
-    const id = /^\| `([a-z]+\.[a-z_]+)` +\|/.exec(line)?.[1];
+    const id = /^\| `([a-z_]+\.[a-z_]+)` +\|/.exec(line)?.[1];
     if (!id) continue;
     const cells = line.split('|').map((cell) => cell.trim());
     // cells[0] is the empty string before the leading pipe.

apps/api/src/read-trail/read-trail-maintenance.service.ts

@@ -0,0 +1,139 @@
+import { Injectable } from '@nestjs/common';
+import { PinoLogger } from 'nestjs-pino';
+
+import { AuditService } from '../audit/audit.service';
+import { ClockService } from '../common/clock.service';
+import { PrismaService } from '../prisma/prisma.service';
+import { InstanceSettingsService } from '../settings/instance-settings.service';
+
+const MS_PER_DAY = 24 * 60 * 60 * 1000;
+
+/** How many months ahead of "now" a partition must exist. Two keeps a
+ * multi-week job outage from ever reaching an uncovered month (the DEFAULT
+ * partition would still catch it — reads never fail on a missing month). */
+const MONTHS_AHEAD = 2;
+
+/**
+ * Read-trail storage maintenance (issue #224, ADR 0023), one daily job with
+ * two duties:
+ *
+ * 1. **Partition upkeep** — `read_events` is RANGE-partitioned by month
+ *    (migration `20260731170000`); this creates the next {@link MONTHS_AHEAD}
+ *    monthly partitions, each with its per-partition dedup unique index
+ *    (#223 — the partitioned parent cannot carry it). A `db push` database
+ *    (tests) has a plain table; partition work skips itself there.
+ * 2. **Retention** — events older than `readTrail.retentionDays` are
+ *    removed: whole months by dropping their partition (no scan), the
+ *    remainder (default partition, plain tables) by a ranged delete. The
+ *    deletion is audited (`read_trail.pruned`) so a gap in the evidence is
+ *    always explainable — same principle as `audit.pruned` (#196), but a
+ *    deliberately separate period.
+ */
+@Injectable()
+export class ReadTrailMaintenanceService {
+  constructor(
+    private readonly prisma: PrismaService,
+    private readonly settings: InstanceSettingsService,
+    private readonly audit: AuditService,
+    private readonly clock: ClockService,
+    private readonly logger: PinoLogger,
+  ) {
+    this.logger.setContext(ReadTrailMaintenanceService.name);
+  }
+
+  async run(): Promise<void> {
+    await this.ensurePartitions();
+    await this.pruneExpired();
+  }
+
+  /** True when read_events is a partitioned parent (relkind `p`). */
+  private async isPartitioned(): Promise<boolean> {
+    const rows = await this.prisma.$queryRaw<{ relkind: string }[]>`
+      SELECT relkind::text FROM pg_class
+       WHERE relname = 'read_events' AND relnamespace = 'public'::regnamespace`;
+    return rows[0]?.relkind === 'p';
+  }
+
+  /** `read_events_y2026m08` for 2026-08. */
+  private partitionName(month: Date): string {
+    const y = month.getUTCFullYear();
+    const m = String(month.getUTCMonth() + 1).padStart(2, '0');
+    return `read_events_y${y}m${m}`;
+  }
+
+  private monthStart(base: Date, offsetMonths: number): Date {
+    return new Date(Date.UTC(base.getUTCFullYear(), base.getUTCMonth() + offsetMonths, 1));
+  }
+
+  async ensurePartitions(): Promise<void> {
+    if (!(await this.isPartitioned())) {
+      this.logger.debug('read_events is not partitioned here; skipping partition upkeep');
+      return;
+    }
+    const now = this.clock.now();
+    for (let offset = 0; offset <= MONTHS_AHEAD; offset += 1) {
+      const from = this.monthStart(now, offset);
+      const to = this.monthStart(now, offset + 1);
+      const name = this.partitionName(from);
+      try {
+        await this.prisma.$executeRawUnsafe(
+          `CREATE TABLE IF NOT EXISTS "${name}" PARTITION OF "read_events"
+             FOR VALUES FROM ('${from.toISOString()}') TO ('${to.toISOString()}')`,
+        );
+        await this.prisma.$executeRawUnsafe(
+          `CREATE UNIQUE INDEX IF NOT EXISTS "${name}_dedup_key"
+             ON "${name}" ("dedup_key", "window_bucket")`,
+        );
+      } catch (error) {
+        // Most likely: the DEFAULT partition already holds rows of this month
+        // (the job lagged past a month boundary). Nothing is lost — those
+        // rows live in the default partition and age out through the ranged
+        // delete below; the month just cannot get its own partition anymore.
+        this.logger.warn({ partition: name, err: error }, 'read-trail partition not created');
+      }
+    }
+  }
+
+  async pruneExpired(): Promise<number> {
+    const retentionDays = await this.settings.get('readTrail.retentionDays');
+    const cutoff = new Date(this.clock.now().getTime() - retentionDays * MS_PER_DAY);
+    let dropped = 0;
+
+    if (await this.isPartitioned()) {
+      // Whole months strictly before the cutoff month go by DROP — no scan,
+      // and the dropped range is exact (every row in them is < cutoff).
+      const partitions = await this.prisma.$queryRaw<{ relname: string }[]>`
+        SELECT c.relname::text
+          FROM pg_inherits i
+          JOIN pg_class c ON c.oid = i.inhrelid
+         WHERE i.inhparent = 'read_events'::regclass
+           AND c.relname ~ '^read_events_y[0-9]{4}m[0-9]{2}$'`;
+      const cutoffMonth = this.monthStart(cutoff, 0);
+      for (const { relname } of partitions) {
+        const match = /^read_events_y(\d{4})m(\d{2})$/.exec(relname);
+        if (!match) continue;
+        const monthEnd = new Date(Date.UTC(Number(match[1]), Number(match[2]), 1));
+        if (monthEnd.getTime() > cutoffMonth.getTime()) continue;
+        const counted = await this.prisma.$queryRawUnsafe<{ count: bigint }[]>(
+          `SELECT count(*)::bigint AS count FROM "${relname}"`,
+        );
+        dropped += Number(counted[0]?.count ?? 0n);
+        await this.prisma.$executeRawUnsafe(`DROP TABLE "${relname}"`);
+      }
+    }
+
+    // The remainder: rows before the cutoff inside surviving partitions,
+    // the default partition, or a plain (test) table.
+    const deleted = await this.prisma.readEvent.deleteMany({
+      where: { occurredAt: { lt: cutoff } },
+    });
+    const count = dropped + deleted.count;
+    if (count > 0) {
+      await this.audit.record({
+        action: 'read_trail.pruned',
+        details: { count, cutoff: cutoff.toISOString(), retentionDays },
+      });
+    }
+    return count;
+  }
+}

apps/api/src/read-trail/read-trail-storage.e2e.db.test.ts

@@ -0,0 +1,308 @@
+import { execFileSync } from 'node:child_process';
+import { join } from 'node:path';
+
+import { INestApplication } from '@nestjs/common';
+import { PrismaClient } from '@prisma/client';
+import { PinoLogger } from 'nestjs-pino';
+import request from 'supertest';
+import { afterAll, beforeAll, describe, expect, it, vi } from 'vitest';
+
+import { ClockService } from '../common/clock.service';
+import { InstanceSettingsService } from '../settings/instance-settings.service';
+import { createTestApp, sessionCookieOf } from '../testing/test-app';
+import { createTestPrisma, hasTestDb, uniqueSuffix } from '../testing/test-db';
+import { UsersService } from '../users/users.service';
+
+import { ReadTrailMaintenanceService } from './read-trail-maintenance.service';
+
+const API_ROOT = join(__dirname, '..', '..');
+const MS_PER_DAY = 24 * 60 * 60 * 1000;
+
+/**
+ * Read-trail storage (issue #224, ADR 0023): its own retention period with
+ * an audited deletion, the Site-Admin query path, and — against a freshly
+ * migrated database, where the real DDL applies — the monthly RANGE
+ * partitioning with per-partition dedup indexes and DROP-based pruning.
+ */
+describe.skipIf(!hasTestDb)('read-trail storage (e2e, issue #224)', () => {
+  const suffix = uniqueSuffix();
+
+  describe('retention and admin query path (shared database)', () => {
+    let app: INestApplication;
+    let prisma: PrismaClient;
+    const password = 'lesetrail speicher 123';
+
+    let adminId: string;
+    let adminCookie: string;
+    let readerCookie: string;
+    // Plain text ids — read_events has no FKs by design (#222).
+    const pondId = `pond-${suffix}`;
+
+    const api = () => request(app.getHttpServer());
+
+    async function makeUser(handle: string, siteAdmin: boolean) {
+      const users = app.get(UsersService);
+      const user = await users.createUser({
+        username: `${handle}-${suffix}`,
+        email: `${handle}-${suffix}@example.test`,
+        displayName: handle,
+        password,
+        locale: 'en',
+      });
+      await users.markEmailVerified(user.id);
+      if (siteAdmin) {
+        await prisma.user.update({ where: { id: user.id }, data: { isSiteAdmin: true } });
+      }
+      const cookie = sessionCookieOf(
+        await api()
+          .post('/api/v1/auth/login')
+          .send({ usernameOrEmail: `${handle}-${suffix}`, password })
+          .expect(200),
+      );
+      return { id: user.id, cookie };
+    }
+
+    function eventRow(overrides: Record<string, unknown>) {
+      return {
+        actorId: null,
+        sessionKey: 'anon',
+        pageId: null,
+        pondId,
+        channel: 'page_view',
+        classification: 'vs_nfd',
+        dedupKey: `k-${suffix}-${Math.random().toString(36).slice(2)}`,
+        windowBucket: BigInt(Math.floor(Math.random() * 1_000_000_000)),
+        windowSeconds: 300,
+        ...overrides,
+      };
+    }
+
+    beforeAll(async () => {
+      prisma = createTestPrisma();
+      await prisma.rateLimit.deleteMany({});
+      app = await createTestApp();
+      const admin = await makeUser('storage-admin', true);
+      adminId = admin.id;
+      adminCookie = admin.cookie;
+      const reader = await makeUser('storage-reader', false);
+      readerCookie = reader.cookie;
+    });
+
+    afterAll(async () => {
+      await prisma.readEvent.deleteMany({ where: { pondId } });
+      await prisma.auditEntry.deleteMany({ where: { action: 'read_trail.pruned' } });
+      await prisma.instanceSetting.deleteMany({ where: { key: 'readTrail.retentionDays' } });
+      await prisma.user.deleteMany({ where: { username: { contains: suffix } } });
+      await prisma.$disconnect();
+      await app.close();
+    });
+
+    it('prunes only events past its own period and audits the deletion', async () => {
+      await app.get(InstanceSettingsService).set('readTrail.retentionDays', 30, adminId);
+      const now = Date.now();
+      await prisma.readEvent.createMany({
+        data: [
+          eventRow({ occurredAt: new Date(now - 40 * MS_PER_DAY) }),
+          eventRow({ occurredAt: new Date(now - 31 * MS_PER_DAY) }),
+          eventRow({ occurredAt: new Date(now - 5 * MS_PER_DAY) }),
+        ],
+      });
+
+      const maintenance = app.get(ReadTrailMaintenanceService);
+      const pruned = await maintenance.pruneExpired();
+      expect(pruned).toBeGreaterThanOrEqual(2);
+
+      const remaining = await prisma.readEvent.findMany({ where: { pondId } });
+      expect(remaining).toHaveLength(1);
+
+      const audit = await prisma.auditEntry.findFirst({
+        where: { action: 'read_trail.pruned' },
+        orderBy: { at: 'desc' },
+      });
+      expect(audit).not.toBeNull();
+      expect(audit!.details).toMatchObject({ retentionDays: 30 });
+    });
+
+    it('answers "who read page X" and "what did user Y read" for Site Admins only', async () => {
+      const pageId = `page-${suffix}`;
+      await prisma.readEvent.createMany({
+        data: [
+          eventRow({ pageId, actorId: adminId, channel: 'export' }),
+          eventRow({ pageId: `other-${suffix}`, actorId: null }),
+        ],
+      });
+
+      const byPage = await api()
+        .get(`/api/v1/admin/system/read-events?pageId=${pageId}`)
+        .set('Cookie', adminCookie)
+        .expect(200);
+      expect(byPage.body.total).toBe(1);
+      expect(byPage.body.entries[0]).toMatchObject({
+        pageId,
+        channel: 'export',
+        windowSeconds: 300,
+      });
+      expect(byPage.body.entries[0].actor).toMatchObject({ id: adminId });
+
+      const byActor = await api()
+        .get(`/api/v1/admin/system/read-events?actor=storage-admin-${suffix}`)
+        .set('Cookie', adminCookie)
+        .expect(200);
+      expect(byActor.body.total).toBe(1);
+
+      // An unknown username matches nothing rather than everything.
+      const unknown = await api()
+        .get(`/api/v1/admin/system/read-events?actor=nobody-${suffix}`)
+        .set('Cookie', adminCookie)
+        .expect(200);
+      expect(unknown.body.total).toBe(0);
+
+      await api()
+        .get(`/api/v1/admin/system/read-events?pageId=${pageId}`)
+        .set('Cookie', readerCookie)
+        .expect(403);
+    });
+  });
+
+  describe('partitioned shape (fresh database, real migrations)', () => {
+    const baseUrl = process.env.TEST_DATABASE_URL!;
+    const dbName = `dorfteich_trail_${suffix}`;
+    let fresh: PrismaClient;
+    let maintenance: ReadTrailMaintenanceService;
+    const auditRecord = vi.fn().mockResolvedValue(undefined);
+    let retentionDays = 365;
+
+    function freshUrl(): string {
+      const url = new URL(baseUrl);
+      url.pathname = `/${dbName}`;
+      return url.toString();
+    }
+
+    beforeAll(async () => {
+      const admin = new PrismaClient({ datasourceUrl: baseUrl });
+      try {
+        await admin.$executeRawUnsafe(`CREATE DATABASE "${dbName}"`);
+      } finally {
+        await admin.$disconnect();
+      }
+      execFileSync(
+        process.execPath,
+        [join(API_ROOT, 'node_modules', 'prisma', 'build', 'index.js'), 'migrate', 'deploy'],
+        { env: { ...process.env, DATABASE_URL: freshUrl() }, stdio: 'pipe', cwd: API_ROOT },
+      );
+      fresh = new PrismaClient({ datasourceUrl: freshUrl() });
+      const settingsStub = {
+        get: async () => retentionDays,
+      } as unknown as InstanceSettingsService;
+      const auditStub = { record: auditRecord } as never;
+      const loggerStub = {
+        setContext: () => undefined,
+        debug: () => undefined,
+        info: () => undefined,
+        warn: () => undefined,
+        error: () => undefined,
+      } as unknown as PinoLogger;
+      maintenance = new ReadTrailMaintenanceService(
+        fresh as never,
+        settingsStub,
+        auditStub,
+        new ClockService(),
+        loggerStub,
+      );
+    }, 120_000);
+
+    afterAll(async () => {
+      await fresh.$disconnect();
+      const admin = new PrismaClient({ datasourceUrl: baseUrl });
+      try {
+        await admin.$executeRawUnsafe(`DROP DATABASE IF EXISTS "${dbName}" WITH (FORCE)`);
+      } finally {
+        await admin.$disconnect();
+      }
+    });
+
+    it('migrates read_events to a partitioned parent with current-month coverage', async () => {
+      const kind = await fresh.$queryRaw<{ relkind: string }[]>`
+        SELECT relkind::text FROM pg_class WHERE relname = 'read_events'`;
+      expect(kind[0]!.relkind).toBe('p');
+
+      const partitions = await fresh.$queryRaw<{ relname: string }[]>`
+        SELECT c.relname::text FROM pg_inherits i JOIN pg_class c ON c.oid = i.inhrelid
+         WHERE i.inhparent = 'read_events'::regclass ORDER BY c.relname`;
+      const names = partitions.map((p) => p.relname);
+      expect(names).toContain('read_events_default');
+      expect(names.some((n) => /^read_events_y\d{4}m\d{2}$/.test(n))).toBe(true);
+    });
+
+    it('keeps the dedup unique pair enforced per partition (P2002 on the duplicate)', async () => {
+      const row = {
+        actorId: null,
+        sessionKey: 'anon',
+        pageId: null,
+        pondId: 'pond-part',
+        channel: 'page_view',
+        classification: 'vs_nfd',
+        dedupKey: `dup-${suffix}`,
+        windowBucket: 42n,
+        windowSeconds: 300,
+      };
+      await fresh.readEvent.create({ data: row });
+      await expect(fresh.readEvent.create({ data: { ...row } })).rejects.toMatchObject({
+        code: 'P2002',
+      });
+    });
+
+    it('creates months ahead and prunes whole expired partitions by DROP, audited', async () => {
+      await maintenance.ensurePartitions();
+      const next = new Date();
+      const nextMonth = new Date(Date.UTC(next.getUTCFullYear(), next.getUTCMonth() + 2, 1));
+      const nextName = `read_events_y${nextMonth.getUTCFullYear()}m${String(
+        nextMonth.getUTCMonth() + 1,
+      ).padStart(2, '0')}`;
+      const created = await fresh.$queryRawUnsafe<{ relname: string }[]>(
+        `SELECT relname::text FROM pg_class WHERE relname = '${nextName}'`,
+      );
+      expect(created).toHaveLength(1);
+      // ...and each new partition carries its own dedup unique index.
+      const index = await fresh.$queryRawUnsafe<{ indexname: string }[]>(
+        `SELECT indexname::text FROM pg_indexes WHERE tablename = '${nextName}'
+          AND indexname = '${nextName}_dedup_key'`,
+      );
+      expect(index).toHaveLength(1);
+
+      // An old month: partition + one event well past retention.
+      await fresh.$executeRawUnsafe(
+        `CREATE TABLE "read_events_y2020m01" PARTITION OF "read_events"
+           FOR VALUES FROM ('2020-01-01') TO ('2020-02-01')`,
+      );
+      await fresh.readEvent.create({
+        data: {
+          occurredAt: new Date('2020-01-15T12:00:00Z'),
+          actorId: null,
+          sessionKey: 'anon',
+          pageId: null,
+          pondId: 'pond-part',
+          channel: 'export',
+          classification: 'vs_nfd',
+          dedupKey: `old-${suffix}`,
+          windowBucket: 1n,
+          windowSeconds: 300,
+        },
+      });
+
+      const pruned = await maintenance.pruneExpired();
+      expect(pruned).toBeGreaterThanOrEqual(1);
+      const gone = await fresh.$queryRawUnsafe<{ relname: string }[]>(
+        `SELECT relname::text FROM pg_class WHERE relname = 'read_events_y2020m01'`,
+      );
+      expect(gone).toHaveLength(0);
+      expect(auditRecord).toHaveBeenCalledWith(
+        expect.objectContaining({ action: 'read_trail.pruned' }),
+      );
+      // The current month's events survive.
+      const kept = await fresh.readEvent.count({ where: { dedupKey: `dup-${suffix}` } });
+      expect(kept).toBe(1);
+      retentionDays = 365; // restore for any later use
+    });
+  });
+});

apps/api/src/read-trail/read-trail.module.ts

@@ -1,10 +1,16 @@
-import { Global, Module } from '@nestjs/common';
+import { Global, Module, OnModuleInit } from '@nestjs/common';
 
 import { CommonModule } from '../common/common.module';
+import { SchedulerModule } from '../scheduler/scheduler.module';
+import { SchedulerService } from '../scheduler/scheduler.service';
 import { SettingsModule } from '../settings/settings.module';
 
+import { ReadTrailMaintenanceService } from './read-trail-maintenance.service';
 import { ReadTrailService } from './read-trail.service';
 
+/** Daily, per operations.md's maintenance-jobs table (issue #224). */
+const READ_TRAIL_MAINTENANCE_CADENCE_SECONDS = 24 * 60 * 60;
+
 /**
  * Global like AuditModule and for the same reason: the read-access trail
  * (issue #222, ADR 0023) cuts across every module that serves page content —
@@ -12,8 +18,23 @@ import { ReadTrailService } from './read-trail.service';
  */
 @Global()
 @Module({
-  imports: [CommonModule, SettingsModule],
-  providers: [ReadTrailService],
-  exports: [ReadTrailService],
+  imports: [CommonModule, SchedulerModule, SettingsModule],
+  providers: [ReadTrailService, ReadTrailMaintenanceService],
+  exports: [ReadTrailService, ReadTrailMaintenanceService],
 })
-export class ReadTrailModule {}
+export class ReadTrailModule implements OnModuleInit {
+  constructor(
+    private readonly scheduler: SchedulerService,
+    private readonly maintenance: ReadTrailMaintenanceService,
+  ) {}
+
+  onModuleInit(): void {
+    this.scheduler.register({
+      name: 'read-trail-maintenance',
+      cadenceSeconds: READ_TRAIL_MAINTENANCE_CADENCE_SECONDS,
+      run: async () => {
+        await this.maintenance.run();
+      },
+    });
+  }
+}

apps/api/src/settings/instance-settings.service.ts

@@ -56,6 +56,13 @@ export const INSTANCE_SETTINGS = {
   // 5 minutes keeps a live Yjs session (collab tokens every 60 s) at a
   // bounded ~12 events/hour/page while still evidencing distinct visits.
   'readTrail.dedupWindowMinutes': z.number().int().min(1).default(5),
+  // Read-trail retention (issue #224): days a read event is kept before the
+  // daily maintenance job removes it — deliberately independent of
+  // `audit.retentionDays` (#196), because volume, purpose and legal basis
+  // differ. The deletion itself is audited (`read_trail.pruned`), so a gap
+  // is always explainable. One year mirrors the audit default; shortening
+  // it is an operator decision under the purpose limitation (#225).
+  'readTrail.retentionDays': z.number().int().min(1).default(365),
   // Default VS-NfD classification for newly created pages (ADR 0022,
   // issue #204). An instance operated inside a classified environment sets
   // this to `vs_nfd` so nothing starts unmarked; inheritance from the

apps/web/e2e/system.spec.ts

@@ -20,8 +20,9 @@ test('lists maintenance jobs and triggers one manually', async ({ browser }) =>
   // Keep in sync with the scheduler registrations: trash-purge,
   // version-thinning, page-compaction, data-export-purge,
   // notification-digest, orphan-file-sweep (#194), audit-retention (#196),
-  // conversion-payload-prune (#233), mail-outbox-retention (#234).
-  await expect(jobsTable.locator('tbody tr')).toHaveCount(9);
+  // conversion-payload-prune (#233), mail-outbox-retention (#234),
+  // read-trail-maintenance (#224).
+  await expect(jobsTable.locator('tbody tr')).toHaveCount(10);
 
   const firstRow = jobsTable.locator('tbody tr').first();
   await firstRow.getByRole('button').click();

docs/architecture/adr/0023-read-access-audit-trail.md

@@ -89,6 +89,23 @@ traffic per open document.
   anonymous browsers apart would require fingerprinting (IP/UA), which the
   purpose limitation (#225) rules out.
 
+## Decisions taken in #224
+
+- **Monthly RANGE partitions** on `occurred_at`, plus a DEFAULT partition
+  as a safety net: a lagging maintenance job must never make classified
+  reads fail (the trail write is hard-failing — an ops miss must not
+  become an outage). Retention DROPs whole expired months without
+  scanning; the pruning run is audited (`read_trail.pruned`) under the
+  trail's own `readTrail.retentionDays` (default 365).
+- **The dedup unique pair lives per partition** (PostgreSQL cannot carry
+  it on the parent without the partition key). A dedup bucket spanning a
+  month boundary can therefore record one duplicate — over-recording is
+  acceptable, gaps are not.
+- **Query path is API-only** (`GET /admin/system/read-events`, Site
+  Admin): the trail is an examiner's tool, not a daily screen. Evidence
+  nobody can read is not evidence — the path exists; a panel does not.
+- **Growth measured**: ~1 MB per 1000 events including indexes.
+
 ## Consequences
 
 - The scope limit is the feature's strongest argument in the works-council

docs/architecture/audit-events.md

@@ -1,7 +1,7 @@
 # Audit event catalogue
 
-**Catalogue version 1.1 (2026-07-31; 1.1 adds `page.classification_*`,
-issue #205).**
+**Catalogue version 1.2 (2026-07-31; 1.2 adds `read_trail.pruned`,
+issue #224; 1.1 added `page.classification_*`, issue #205).**
 
 This is the operator-facing contract for the audit trail: every event id
 the application can emit, with its trigger, severity, actor/target
@@ -103,11 +103,12 @@ failure), `warning` = feeds detection (suspicious or destructive),
 
 ### Content integrity & lifecycle (`file.*`, `pond.*`, `audit.*`)
 
-| Id                      | Trigger                                                             | Severity | Actor                              | Target       | Fields                                                                                                                                   |
-| ----------------------- | ------------------------------------------------------------------- | -------- | ---------------------------------- | ------------ | ---------------------------------------------------------------------------------------------------------------------------------------- |
-| `file.integrity_failed` | Attachment download hash mismatch — fail-closed (issue #199)        | critical | `null` (any downloader; detection) | `attachment` | `pondId`, `expected` (stored sha256), `actual` (computed sha256)                                                                         |
-| `pond.purged`           | Pond irreversibly destroyed (manual or trash retention, issue #193) | notice   | admin, `null` when retention-run   | `pond`       | `trigger` (`manual` \| `retention`) plus per-object-type deletion counts (e.g. `pages`, `attachments`, … — informational, keys may grow) |
-| `audit.pruned`          | Audit retention deleted rows past the period (issue #196)           | info     | `null` (system)                    | —            | `count`, `cutoff` (ISO), `retentionDays`                                                                                                 |
+| Id                      | Trigger                                                              | Severity | Actor                              | Target       | Fields                                                                                                                                   |
+| ----------------------- | -------------------------------------------------------------------- | -------- | ---------------------------------- | ------------ | ---------------------------------------------------------------------------------------------------------------------------------------- |
+| `file.integrity_failed` | Attachment download hash mismatch — fail-closed (issue #199)         | critical | `null` (any downloader; detection) | `attachment` | `pondId`, `expected` (stored sha256), `actual` (computed sha256)                                                                         |
+| `pond.purged`           | Pond irreversibly destroyed (manual or trash retention, issue #193)  | notice   | admin, `null` when retention-run   | `pond`       | `trigger` (`manual` \| `retention`) plus per-object-type deletion counts (e.g. `pages`, `attachments`, … — informational, keys may grow) |
+| `audit.pruned`          | Audit retention deleted rows past the period (issue #196)            | info     | `null` (system)                    | —            | `count`, `cutoff` (ISO), `retentionDays`                                                                                                 |
+| `read_trail.pruned`     | Read-trail retention removed events past its own period (issue #224) | info     | `null` (system)                    | —            | `count`, `cutoff` (ISO), `retentionDays`                                                                                                 |
 
 ### Public API (`api.*`)
 

docs/architecture/data-model.md

@@ -186,6 +186,40 @@ Cached counters per pond: `storage_bytes_used`, `editor_count`,
 - `jobs`: conversion jobs for import/export (ADR 0009) and maintenance jobs
   (compaction, trash purge, quota reconciliation) with status + timestamps.
 
+## Read-access trail (`read_events`, issues #222–#224, ADR 0023)
+
+One row per read of a `classification = vs_nfd` page per channel and dedup
+window — separate from `audit_log` because volume, purpose and legal basis
+all differ.
+
+| Column                       | Notes                                                                                |
+| ---------------------------- | ------------------------------------------------------------------------------------ |
+| `id`, `occurred_at`          | composite PK (the partition key must be part of it)                                  |
+| `actor_id`                   | reader, null = anonymous; **no FK** — evidence outlives accounts                     |
+| `session_key`                | `session:<id>` / `token:<id>` / `job:<id>` / `anon`                                  |
+| `page_id`, `pond_id`         | plain ids, **no FK** — evidence outlives purges                                      |
+| `channel`                    | `page_view` / `no_js_shell` / `public_api` / `attachment` / `export` / `collab_join` |
+| `classification`             | at read time; reclassification never rewrites history                                |
+| `dedup_key`, `window_bucket` | unique pair per partition — race-free dedup (#223)                                   |
+| `window_seconds`             | the row represents up to this many seconds of access                                 |
+| `details`                    | small context (export format, attachment id) — never content                         |
+
+Storage shape (issue #224): RANGE-partitioned by month on `occurred_at`,
+with a DEFAULT partition as safety net (a lagging maintenance job must
+never make classified reads fail — the trail write is hard-failing by
+design). The daily `read-trail-maintenance` job creates months ahead
+(each with its per-partition dedup index) and applies the trail's own
+retention `readTrail.retentionDays` (default 365): whole expired months
+are DROPped without scanning, remainders deleted by range; every pruning
+run is audited as `read_trail.pruned`. Growth, measured on PG 17:
+**~1 MB per 1000 events** including indexes (≈970 bytes/row) — a
+100-person instance at hundreds of classified reads/day stays in the tens
+of MB per year.
+
+Query path: `GET /admin/system/read-events` (Site Admin, filters
+`pageId`, `actor`, `channel`, `from`/`to`) — deliberately API-only, no
+panel: the trail is an examiner's tool, not a daily screen.
+
 ## Comments & notifications (later milestone)
 
 - `comments`: `page_id`, `author_id`, `body` (Markdown), `anchor`

docs/architecture/operations.md

@@ -109,6 +109,7 @@ monitoring, structured logs, backup alerting — no dedicated metrics stack.
 | conversion payload prune | daily                   | null finished conversion jobs' raw bytes (#233)     |
 | mail outbox retention    | daily                   | delete sent/failed outbox rows past period (#234)   |
 | mail outbox retry        | every minute            | e-mail delivery with backoff                        |
+| read-trail maintenance   | daily                   | partition months ahead + prune `read_events` (#224) |
 
 Job outcomes are visible in the Site Admin UI (last run, status) — that
 panel is the operator's single glance for instance health.

docs/vs-nfd/20-massnahmenplan.md

@@ -136,7 +136,7 @@ Protokolliert werden Lesezugriffe **ausschließlich** auf Seiten mit
       Attachment-Download, Export, No-JS-Shell, Collab-WS-Join · 4 AT · #222
 - [x] Dedup-Fenster (eine Sitzung + eine Seite innerhalb N Minuten = ein
       Ereignis), sonst erzeugt Yjs-Sync eine Ereignisflut · 2 AT · #223
-- [ ] Getrennte Tabelle mit eigener Retention und Partitionierung · 2 AT · #224
+- [x] Getrennte Tabelle mit eigener Retention und Partitionierung · 2 AT · #224
 - [ ] Abschaltbar, Zweckbindung dokumentiert · 1–2 AT · #225
 
 Vorteil über den Aufwand hinaus: Die Zweckbindung ist sauber begründbar

docs/vs-nfd/50-haertungsleitfaden.md

@@ -40,6 +40,7 @@ Settings-Cache ist in-process (operations.md).
 | `backup.nextcloud.enabled`                                                                                  | `false`                                 | `false`        | „Backup nur lokal": kein Anwendungs-Upload von Restore-Sets zu Drittdiensten. Fernspiegel regelt ausschließlich die Deploy-Allowlist (1.2).                                                                                                                                                                    |
 | `trash.retentionDays`, `audit.retentionDays`, `conversion.payloadRetentionDays`, `mail.outboxRetentionDays` | Defaults (30/365/30/30)                 | ebd.           | Aufbewahrung bewusst begrenzt; Verkürzung nach Betreiber-Löschkonzept zulässig (Betriebshandbuch §5).                                                                                                                                                                                                          |
 | `readTrail.dedupWindowMinutes`                                                                              | Default (5)                             | `5`            | Dedup-Fenster des Lesetrails (#223): je (Sitzung, Seite, Kanal) ein Ereignis pro Fenster — begrenzt die Ereignisflut einer Live-Sitzung auf ~12/h. Kleiner = feineres Protokoll und mehr Zeilen; Änderung mit dem Zweckbindungs-Dokument (#225) abstimmen.                                                     |
+| `readTrail.retentionDays`                                                                                   | Default (365)                           | `365`          | eigene Aufbewahrung des Lesetrails (#224), bewusst getrennt von `audit.retentionDays`; Löschläufe sind selbst auditiert (`read_trail.pruned`). Dauer mit der Zweckbindung (#225) und dem Betreiber-Löschkonzept abstimmen.                                                                                     |
 | `legal.imprint`, `legal.privacyPolicy`                                                                      | befüllt                                 | leer           | Betreiberpflicht; leere Seiten zeigen einen Warnbanner.                                                                                                                                                                                                                                                        |
 
 ### 1.2 Deploy-Konfiguration (`.env` / Compose — nur Plattformzugriff, bewusst nicht per Admin-UI)

packages/shared/src/system.ts

@@ -164,6 +164,49 @@ export interface AuditListView {
   total: number;
 }
 
+/**
+ * Site-Admin query path over the read-access trail (issue #224, ADR 0023) —
+ * evidence nobody can read is not evidence. Answers the two expected
+ * questions: "who read page X" (pageId) and "what did user Y read" (actor),
+ * both within a period.
+ */
+export const READ_EVENT_PAGE_SIZE = 50;
+
+export const readEventListQuerySchema = z.object({
+  /** Page id the events belong to. */
+  pageId: z.string().trim().min(1).optional(),
+  /** Exact username of the reading user. */
+  actor: z.string().trim().min(1).optional(),
+  channel: z.string().trim().min(1).optional(),
+  from: z.coerce.date().optional(),
+  to: z.coerce.date().optional(),
+  page: z.coerce.number().int().min(1).default(1),
+});
+
+export type ReadEventListQuery = z.infer<typeof readEventListQuerySchema>;
+
+export interface ReadEventView {
+  id: string;
+  occurredAt: string;
+  actor: AuditActorView | null;
+  /** Null for pond-level attachments (no page); the details name the file. */
+  pageId: string | null;
+  pondId: string;
+  channel: string;
+  classification: string;
+  /** Seconds the event's dedup window spans (#223) — the row represents up
+   * to this much access time, not a single request. */
+  windowSeconds: number;
+  details: Record<string, unknown> | null;
+}
+
+export interface ReadEventListView {
+  entries: ReadEventView[];
+  page: number;
+  pageCount: number;
+  total: number;
+}
+
 export interface StoragePondView {
   pondId: string;
   name: string;