diff --git a/.changeset/showcase-inquiry-purge-bulk-intent.md b/.changeset/showcase-inquiry-purge-bulk-intent.md
new file mode 100644
index 0000000000..63167f2e02
--- /dev/null
+++ b/.changeset/showcase-inquiry-purge-bulk-intent.md
@@ -0,0 +1,35 @@
+---
+---
+
+fix(showcase): `showcase_inquiry_purge` 的谓词删除节点声明 `multi: true`（#5225）
+
+`examples/app-showcase/src/automation/flows/index.ts` 里 `InquiryPurgeFlow` 的
+`purge` 节点按谓词 `{ status: 'closed' }` 批量删除,却没有声明批量意图。数据引擎只
+在 `filter` 用标量 `id` 点名一行时才接受无 `options.multi` 的写入,于是这条流的每
+一次运行都在该节点失败:
+
+```
+Node 'purge' failed: delete_record(showcase_inquiry) failed:
+Delete requires an ID or options.multi=true
+```
+
+`acted: 0` —— 声明式端点 `POST /api/v1/apps/showcase/inquiries/purge` 与内建触发
+路由 `POST /api/v1/automation/showcase_inquiry_purge/trigger` 两条路径同一个签名。
+也就是说 `src/coverage.ts` 声称由本流演示的 CRUD 四件套的 delete 半边,从写下的那
+天起就是 declared ≠ enforced(PD #10),直到 #5112 的真机 boot 探针打到它才浮出来。
+
+修法是**补一个声明**,不是改写流程:在 #5393(PR #5485)之前,节点 config 上根本
+不存在任何批量意图的拼写,这正是第 3 轮分诊拒绝 get→loop→逐 id 删的原因(PD #5 的
+workaround)。`multi` 落地之后,一行声明就是长期正确的形状。
+
+⚠️ `filter` 在这里不是可有可无的修饰:`multi: true` 而 `filter` 缺失或为空 = 声明
+式整表删除。本节点是「批量意图 + 谓词边界」的参考样本,也是 #5482 authoring 期
+lint 规则未来的「必须零告警」验收样本。
+
+新增 `examples/app-showcase/test/predicate-write-bulk-intent.test.ts`:把上述规则
+陈述为覆盖**全部** `delete_record` / `update_record` 节点的双向不变量(谓词写必须
+声明 `multi: true`;`multi: true` 必须带非空 `filter`),并深走 ADR-0031 结构化容器
+——`showcase_task_crm_sync` 的 `catch` 区里就藏着一个 `update_record`,只扫顶层
+`nodes` 会漏掉它。
+
+仅改示例应用(`examples/app-showcase` 为 private 包),不发布任何包。
diff --git a/examples/app-showcase/src/automation/flows/index.ts b/examples/app-showcase/src/automation/flows/index.ts
index 86266a7b1a..13bf326f8e 100644
--- a/examples/app-showcase/src/automation/flows/index.ts
+++ b/examples/app-showcase/src/automation/flows/index.ts
@@ -1385,8 +1385,9 @@ export const InboundTaskWebhookFlow = defineFlow({
  * ReassignWizardFlow · get + delete: here). A janitor flow: fetch the
  * already-closed inquiries (records mode), gate on whether any exist, delete
  * by the same filter, and report. Config keys follow the executor contract
- * exactly — `objectName` + `filter` (Prime Directive #12: no
- * `object`/`filters` aliases). `runAs: 'system'` because a janitor acts
+ * exactly — `objectName` + `filter` + the declared bulk intent `multi`
+ * (Prime Directive #12: no `object`/`filters` aliases). `runAs: 'system'`
+ * because a janitor acts
  * across owners; autolaunched with no record trigger — invoke it on demand
  * (API/subflow) rather than on every write.
  */
@@ -1421,7 +1422,21 @@ export const InquiryPurgeFlow = defineFlow({
       id: 'purge',
       type: 'delete_record',
       label: 'Delete them',
-      config: { objectName: 'showcase_inquiry', filter: { status: 'closed' } },
+      // `multi: true` is what makes this a PREDICATE delete — and without it the
+      // node had never deleted anything: the data engine accepts a delete only
+      // when `filter` names one row by scalar `id`, so every run of this flow
+      // failed here with `Delete requires an ID or options.multi=true` and
+      // reported `acted: 0` (#5225, found by the #5112 boot probes). No bulk
+      // spelling existed on this node's config at all until #5393/PR #5485
+      // declared one; the engine's refusal was the contract working, not a bug
+      // to route around — which is why the fix is this declaration and not a
+      // get→loop→delete-by-id rewrite (PD #5).
+      //
+      // ⚠️ `filter` is NOT optional decoration here: `multi: true` with an
+      // absent or empty `filter` is a declared WHOLE-OBJECT delete. This node is
+      // the reference for "bulk intent, bounded by a predicate" — the shape the
+      // #5482 lint rule must leave at zero warnings.
+      config: { objectName: 'showcase_inquiry', filter: { status: 'closed' }, multi: true },
     },
     {
       id: 'report',
diff --git a/examples/app-showcase/test/predicate-write-bulk-intent.test.ts b/examples/app-showcase/test/predicate-write-bulk-intent.test.ts
new file mode 100644
index 0000000000..0bf2f0b863
--- /dev/null
+++ b/examples/app-showcase/test/predicate-write-bulk-intent.test.ts
@@ -0,0 +1,215 @@
+// Copyright (c) 2026 ObjectStack. Licensed under the Apache-2.0 license.
+
+/**
+ * [#5225 / #5393] Every predicate write this app ships declares its bulk intent.
+ *
+ * ## The defect this pins
+ *
+ * `showcase_inquiry_purge`'s `delete_record` node deleted by the predicate
+ * `{ status: 'closed' }` and declared no bulk intent. The data engine accepts a
+ * write without `options.multi` only when `filter` names ONE row by a scalar
+ * `id`, so every run of the flow — both the declarative endpoint
+ * (`POST /api/v1/apps/showcase/inquiries/purge`) and the built-in trigger route
+ * — failed on that node with
+ *
+ *     Node 'purge' failed: delete_record(showcase_inquiry) failed:
+ *     Delete requires an ID or options.multi=true
+ *
+ * and reported `acted: 0`. The showcase's own coverage manifest claims
+ * `delete_record` is demonstrated by this flow, so the delete half of the CRUD
+ * quartet was `declared ≠ enforced` (PD #10) from the day it was written until
+ * #5112's boot probes hit it.
+ *
+ * The fix is a DECLARATION, not a rewrite: until #5393 (PR #5485) no spelling of
+ * bulk intent existed on the node config at all, which is why the third triage
+ * round correctly refused to route around the engine with a
+ * get→loop→delete-by-id rewrite (PD #5 workaround). With `multi` declared, the
+ * one-line fix is the long-term-correct shape.
+ *
+ * ## Why this file sweeps instead of asserting one node
+ *
+ * A test naming only the purge node would go green for the wrong reason the day
+ * someone adds a second predicate write. So the invariant below is stated over
+ * EVERY `delete_record` / `update_record` node in `allFlows`, and it is
+ * two-sided — which matters, because `multi` cuts both ways:
+ *
+ *   - a predicate write WITHOUT `multi: true` is refused by the engine at run
+ *     time (the #5225 failure, silent in every unit test that fakes the engine);
+ *   - `multi: true` with an absent or empty `filter` is a declared WHOLE-OBJECT
+ *     write — every row, by declaration. Authoring-time linting for that shape
+ *     is queued as #5482, and this app is meant to be its "must be zero
+ *     warnings" sample, so the second side is asserted here too.
+ *
+ * Node configs are additionally driven through the REAL spec schemas rather than
+ * inspected as plain objects: the claim is about a VALUE verdict (`multi` is
+ * `true`, `filter` is a non-empty predicate), not merely about a key being an
+ * authorable surface, so full `safeParse` green is the right bar.
+ */
+
+import { describe, it, expect } from 'vitest';
+import { DeleteRecordConfigSchema, UpdateRecordConfigSchema } from '@objectstack/spec/automation';
+
+import { allFlows } from '../src/automation/flows/index.js';
+
+type NodeLike = { id?: string; type?: string; config?: Record<string, unknown> };
+type FlowLike = { name?: string; nodes?: NodeLike[] };
+
+const WRITE_SCHEMAS = {
+  delete_record: DeleteRecordConfigSchema,
+  update_record: UpdateRecordConfigSchema,
+} as const;
+
+type WriteNodeType = keyof typeof WRITE_SCHEMAS;
+
+interface WriteNode {
+  flow: string;
+  node: string;
+  type: WriteNodeType;
+  config: Record<string, unknown>;
+}
+
+/**
+ * Collect write nodes by walking the flow DEEPLY, not just its top-level
+ * `nodes` array.
+ *
+ * This app nests real write nodes inside ADR-0031 structured containers — the
+ * `catch` region of `showcase_task_crm_sync`'s try/catch holds an
+ * `update_record`, and branch/loop bodies elsewhere hold others. A flat scan of
+ * `flow.nodes` silently skips every one of them, which would leave the guard
+ * below passing while the exact class of defect it exists to catch hid one
+ * level down. So the walk is generic over the object graph rather than a list
+ * of container key names (`try`/`catch`/`body`/`branches`/…) that a new
+ * container shape could quietly fall outside of.
+ */
+function collectWriteNodes(flowName: string, value: unknown, out: WriteNode[]): void {
+  if (Array.isArray(value)) {
+    for (const entry of value) collectWriteNodes(flowName, entry, out);
+    return;
+  }
+  if (!value || typeof value !== 'object') return;
+
+  const node = value as NodeLike;
+  const type = node.type as WriteNodeType | undefined;
+  if (typeof type === 'string' && type in WRITE_SCHEMAS && node.id !== undefined) {
+    out.push({
+      flow: flowName,
+      node: String(node.id),
+      type,
+      config: (node.config ?? {}) as Record<string, unknown>,
+    });
+  }
+
+  for (const child of Object.values(value as Record<string, unknown>)) {
+    collectWriteNodes(flowName, child, out);
+  }
+}
+
+const writeNodes: WriteNode[] = [];
+for (const flow of allFlows as unknown as FlowLike[]) {
+  collectWriteNodes(String(flow.name), flow.nodes, writeNodes);
+}
+
+/**
+ * Does this filter name exactly one row the way the engine's non-`multi` path
+ * requires — a SCALAR `id`? `{ id: { $in: [...] } }` does not qualify (the
+ * engine refuses it), and neither does any other predicate.
+ *
+ * `{recordId}` / `{record.id}` templates count: they interpolate to one scalar
+ * id, and #3810 already refuses the node outright when such a template erases
+ * to nothing, so a "scalar" that vanished never reaches the write.
+ */
+function namesOneRowById(filter: unknown): boolean {
+  if (!filter || typeof filter !== 'object') return false;
+  const keys = Object.keys(filter as Record<string, unknown>);
+  if (keys.length !== 1 || keys[0] !== 'id') return false;
+  const id = (filter as { id: unknown }).id;
+  return typeof id === 'string' || typeof id === 'number';
+}
+
+function isNonEmptyPredicate(filter: unknown): boolean {
+  return (
+    !!filter
+    && typeof filter === 'object'
+    && !Array.isArray(filter)
+    && Object.keys(filter as Record<string, unknown>).length > 0
+  );
+}
+
+describe('[#5225] showcase predicate writes declare bulk intent', () => {
+  it('the app really does ship write nodes — this suite is not vacuous', () => {
+    // If a refactor drops every CRUD write node, the per-node cases below would
+    // pass by iterating nothing, which is exactly how #5225 hid for so long.
+    expect(writeNodes.length).toBeGreaterThan(0);
+    expect(writeNodes.some((n) => n.type === 'delete_record')).toBe(true);
+    expect(writeNodes.some((n) => n.type === 'update_record')).toBe(true);
+  });
+
+  it('reaches write nodes nested inside structured containers', () => {
+    // `record_failure` lives in the `catch` region of `showcase_task_crm_sync`,
+    // not in its top-level `nodes`. A flat walk finds everything else and misses
+    // exactly this one, so naming it is what keeps the collector honest — a
+    // regression to `flow.nodes` alone fails here rather than silently shrinking
+    // the sweep's coverage.
+    expect(writeNodes.map((n) => n.node)).toContain('record_failure');
+  });
+
+  describe.each(writeNodes)('$flow / $node ($type)', ({ type, config }) => {
+    it('parses green against the real spec schema', () => {
+      const result = WRITE_SCHEMAS[type].safeParse(config);
+      expect(result.success ? null : JSON.stringify(result.error?.issues)).toBeNull();
+    });
+
+    it('either names one row by scalar id, or declares `multi: true`', () => {
+      // The engine's rule, restated as the authoring rule. A node that satisfies
+      // neither branch is the #5225 shape: it parses, it publishes, and it fails
+      // on every single execution with `requires an ID or options.multi=true`.
+      const single = namesOneRowById(config.filter);
+      expect(single || config.multi === true).toBe(true);
+    });
+
+    it('never declares `multi: true` without a bounding filter', () => {
+      // `multi: true` + absent/empty filter = a declared whole-object write. It
+      // is a legal thing to author deliberately, and it is NOT something this
+      // reference app should ever demonstrate by accident — #5482's lint rule
+      // uses this app as its zero-warning sample.
+      if (config.multi === true) {
+        expect(isNonEmptyPredicate(config.filter)).toBe(true);
+      }
+    });
+  });
+});
+
+describe('[#5225] the purge flow specifically — the node that never deleted anything', () => {
+  const purge = writeNodes.find((n) => n.flow === 'showcase_inquiry_purge' && n.node === 'purge');
+
+  it('is still the delete half of the CRUD quartet src/coverage.ts claims', () => {
+    // coverage.ts names `get+delete: InquiryPurgeFlow` under flowNodeTypes. If
+    // this node is ever renamed or retyped, that claim needs re-checking rather
+    // than this file silently finding nothing.
+    expect(purge).toBeDefined();
+    expect(purge!.type).toBe('delete_record');
+  });
+
+  it('deletes closed inquiries by predicate, with bulk intent declared', () => {
+    expect(purge!.config).toMatchObject({
+      objectName: 'showcase_inquiry',
+      filter: { status: 'closed' },
+      multi: true,
+    });
+    // Not `{ id: … }` — the point of the node is the predicate path, so the
+    // scalar-id escape must NOT be what makes the sweep above pass for it.
+    expect(namesOneRowById(purge!.config.filter)).toBe(false);
+  });
+
+  it('is refused by the engine contract the moment `multi` is dropped', () => {
+    // Reverse verification, direction decided up front: removing the
+    // declaration must land the node back in the branch that produced
+    // `Delete requires an ID or options.multi=true` / `acted: 0`. The schema
+    // still accepts the stripped config — `multi` is optional by design, since
+    // omitting it is a valid deliberate choice — so the regression this pins is
+    // an EXECUTION one, and the sweep rule above is what catches it statically.
+    const { multi: _multi, ...withoutIntent } = purge!.config;
+    expect(DeleteRecordConfigSchema.safeParse(withoutIntent).success).toBe(true);
+    expect(namesOneRowById(withoutIntent.filter) || withoutIntent.multi === true).toBe(false);
+  });
+});