diff --git a/.changeset/classify-index-failure-error-channel.md b/.changeset/classify-index-failure-error-channel.md
new file mode 100644
index 0000000000..0eb4e394ce
--- /dev/null
+++ b/.changeset/classify-index-failure-error-channel.md
@@ -0,0 +1,40 @@
+---
+"@objectstack/metadata-protocol": patch
+---
+
+fix(metadata-protocol): classify a failed index build from the ERROR, not its message (#6699)
+
+`classifyIndexFailure` — the function both runtime partial-index migrations in
+this package classify a failed `CREATE UNIQUE INDEX` with — carried its own
+private unique-violation vocabulary and answered from the **message channel
+only**. That made it the fifth such copy in the repo, and the one #6250's
+inventory missed: it lives in a package none of the other four touched, so it
+was never in that table and none of the queued follow-ups covered it.
+
+The first arm now delegates to `@objectstack/types`' `isUniqueViolationError`
+(#6250 / PR #6541) — the one named answer to "is this a unique-constraint
+violation?" — and `probeThenReplaceIndex` passes it the **caught error object**
+instead of `err.message`. A string-only swap would have compiled unchanged and
+kept the defect: the point of the shared predicate is the `code` / `errno` /
+`cause` channels, which unwrapping the message throws away.
+
+**What changes at runtime.** A driver that reports the conflict on `code` or
+`errno` while giving unhelpful prose — SQLite's `SQLITE_CONSTRAINT_UNIQUE`,
+MySQL's `ER_DUP_ENTRY` / errno `1062`, Postgres' SQLSTATE `23505`, or the
+condition one step down `error.cause` behind a pooled wrapper's `Write failed`
+— was classified `failed`. It is now `conflict`, which is the verdict that
+produces the report ADR-0120 D4 requires: the key that is not enforced, the
+query that lists the offending rows, and the pointer at `os migrate plan`.
+Every message-channel verdict is unchanged — the shared predicate's message
+limb covers all three shipped dialects' prose.
+
+**Two things deliberately preserved.** The arm order still checks the
+duplicate-row question BEFORE the dialect question, because MySQL's duplicate
+error mentions the key and some drivers wrap both facts in one string; and the
+dialect arm (`unsupported`) is still this module's own message-based
+vocabulary, since the shared predicate answers the first arm only and has no
+opinion about dialect support.
+
+`classifyIndexFailure`'s parameter widens from `string` to `unknown`, so every
+existing string call still compiles and is judged exactly as before. Callers
+holding a caught error should pass it directly rather than `err.message`.
diff --git a/packages/metadata-protocol/src/migrations/partial-index-probe.test.ts b/packages/metadata-protocol/src/migrations/partial-index-probe.test.ts
index 304f715e9e..c5d85da225 100644
--- a/packages/metadata-protocol/src/migrations/partial-index-probe.test.ts
+++ b/packages/metadata-protocol/src/migrations/partial-index-probe.test.ts
@@ -157,6 +157,97 @@ describe('probe-first partial index replacement (#6418)', () => {
         expect(classifyIndexFailure('disk I/O error')).toBe('failed');
     });
 
+    /**
+     * #6699 — the substance of the migration onto `@objectstack/types`'
+     * `isUniqueViolationError`: the conflict is judged on the channels the
+     * driver actually wrote it to, not on prose alone.
+     *
+     * Every message below is deliberately USELESS — none carries a word any
+     * unique-violation vocabulary's message limb looks for — so each case can
+     * pass ONLY by reading `code` / `errno`. The second assertion in each case
+     * proves that: run the same prose through the classifier on its own and the
+     * verdict is `failed`, which is what this module answered for all of them
+     * while it carried its own message-only regex.
+     */
+    it.each([
+        ['a SQLite extended result code', { code: 'SQLITE_CONSTRAINT_UNIQUE' }],
+        ["mysql2's symbolic name", { code: 'ER_DUP_ENTRY' }],
+        ['a bare MySQL errno', { errno: 1062 }],
+        ['a Postgres SQLSTATE', { code: '23505' }],
+    ])('reads a conflict off %s when the message says nothing (#6699)', (_label, channels) => {
+        const error = Object.assign(new Error('insert failed'), channels);
+        expect(classifyIndexFailure(error)).toBe('conflict');
+        expect(classifyIndexFailure(error.message)).toBe('failed');
+    });
+
+    it('follows a pooled wrapper down to the cause (#6699)', () => {
+        // Pool and query-builder layers re-throw with the original attached, so
+        // the only copy of the condition is one step down.
+        const wrapped = Object.assign(new Error('Write failed'), {
+            cause: Object.assign(new Error('insert failed'), { code: '23505' }),
+        });
+        expect(classifyIndexFailure(wrapped)).toBe('conflict');
+        expect(classifyIndexFailure(wrapped.message)).toBe('failed');
+    });
+
+    it('keeps the data verdict ahead of the dialect verdict on the OBJECT channel too (#6699)', () => {
+        // The arm order, re-pinned where widening the input could have broken
+        // it: a duplicate reported on `code`, wrapped by a layer whose prose is
+        // a dialect refusal. Judged on the message alone this is `unsupported`
+        // — "this database cannot build this index" for a real data conflict,
+        // exactly the misreport the ordering exists to prevent.
+        const both = Object.assign(new Error('near "WHERE": syntax error'), {
+            code: 'SQLITE_CONSTRAINT_UNIQUE',
+        });
+        expect(classifyIndexFailure(both)).toBe('conflict');
+        expect(classifyIndexFailure(both.message)).toBe('unsupported');
+        // …and on a single string carrying both facts, unchanged since #6418.
+        expect(
+            classifyIndexFailure('near "WHERE": syntax error — duplicate key value violates unique constraint'),
+        ).toBe('conflict');
+    });
+
+    it('a dialect refusal carrying its own code is still `unsupported` (#6699)', () => {
+        // Widening the input from `string` to the error object must not blind
+        // the second arm: MySQL's parse error has `code` and `errno` too, and
+        // neither is a unique-violation signal, so the verdict has to come from
+        // the message exactly as before.
+        const parseError = Object.assign(
+            new Error("You have an error in your SQL syntax ... near 'WHERE state'"),
+            { code: 'ER_PARSE_ERROR', errno: 1064 },
+        );
+        expect(classifyIndexFailure(parseError)).toBe('unsupported');
+        const io = Object.assign(new Error('disk I/O error'), { code: 'SQLITE_IOERR', errno: 10 });
+        expect(classifyIndexFailure(io)).toBe('failed');
+    });
+
+    it('the probe hands the ERROR to the classifier, not its message (#6699)', async () => {
+        // The threading pin, and the only test here that can see it: every
+        // assertion above still passes if `probeThenReplaceIndex` keeps
+        // unwrapping `err.message` before classifying. This one cannot — the
+        // verdict exists nowhere but on `code`.
+        const codeOnly: IndexExec = async (sql: string) => {
+            if (sql.startsWith('CREATE')) {
+                throw Object.assign(new Error('insert failed'), { code: 'SQLITE_CONSTRAINT_UNIQUE' });
+            }
+            return db.exec(sql);
+        };
+
+        const outcome = await probeThenReplaceIndex(codeOnly, {
+            indexName: REAL,
+            probeIndexName: PROBE,
+            buildSql,
+        });
+
+        expect(outcome.status).toBe('conflict');
+        expect(outcome.failedAt).toBe('probe');
+        // `detail` is unchanged — still the driver's own prose, for the operator.
+        expect(outcome.detail).toBe('insert failed');
+        // The probe is what failed, so the previous index is untouched.
+        expect(indexDdl(REAL)).toEqual(EXISTING_DDL);
+        expect(indexDdl(PROBE)).toBeUndefined();
+    });
+
     it('logProblem prefers error(), falls back to warn(), and tolerates neither', () => {
         const full = { warn: vi.fn(), error: vi.fn() };
         logProblem(full, 'msg', 'detail');
diff --git a/packages/metadata-protocol/src/migrations/partial-index-probe.ts b/packages/metadata-protocol/src/migrations/partial-index-probe.ts
index b407c37485..2c8a420a4b 100644
--- a/packages/metadata-protocol/src/migrations/partial-index-probe.ts
+++ b/packages/metadata-protocol/src/migrations/partial-index-probe.ts
@@ -37,6 +37,8 @@
  * a classified status plus the driver's own text and stays out of the way.
  */
 
+import { isUniqueViolationError } from '@objectstack/types';
+
 /** Raw-SQL seam. Mirrors `ensureOverlayIndex`: `raw()` first, `execute()` second. */
 export type IndexExec = (sql: string) => Promise<unknown>;
 
@@ -86,7 +88,26 @@ export type PartialIndexStatus =
     | 'failed';
 
 /**
- * Classify a failed `CREATE UNIQUE INDEX`.
+ * The text the DIALECT arm judges, from a thrown value of any shape.
+ *
+ * `message` first, because that is the channel a driver writes its refusal on
+ * and the only one this arm has ever read; `String()` only as the last resort
+ * — which is what a bare string resolves to unchanged, so a caller holding
+ * nothing but prose is judged exactly as before.
+ */
+function indexFailureText(error: unknown): string {
+    if (typeof error === 'string') return error;
+    if (typeof error === 'object' && error !== null) {
+        const { message } = error as { message?: unknown };
+        if (typeof message === 'string') return message;
+    }
+    return String(error);
+}
+
+/**
+ * Classify a failed `CREATE UNIQUE INDEX`, from the thrown ERROR (#6699).
+ *
+ * ## The two arms, and why the order is load-bearing
  *
  * Duplicate-row wording is checked BEFORE dialect wording: MySQL's duplicate
  * error mentions the key, and some drivers wrap both facts in one string, so
@@ -98,12 +119,35 @@ export type PartialIndexStatus =
  * split: no partial indexes at all, and (before 8.0.13 / on MariaDB) no
  * functional key parts for `COALESCE` parts. Both leave the same outcome — the
  * previous index stays — so one verdict is enough.
+ *
+ * ## Why the first arm is not this module's own regex any more
+ *
+ * "Is this a unique-constraint violation?" is one question, and #6250 gave it
+ * one named answer — `isUniqueViolationError` in `@objectstack/types`. This
+ * function carried a **fifth** private vocabulary for it (#6699, missed by that
+ * inventory because it lives in a package none of the other four touched), and
+ * the copy was strictly weaker in the way that inventory was about: it read the
+ * **message channel only**. A driver that reports the conflict on `code` /
+ * `errno` — SQLite's `SQLITE_CONSTRAINT_UNIQUE`, MySQL's `ER_DUP_ENTRY` /
+ * errno `1062`, Postgres' SQLSTATE `23505` — while giving unhelpful prose
+ * (`insert failed`, a pooled wrapper's `Write failed`, or the condition one step
+ * down `error.cause`) was classified `failed` here, where the shared predicate
+ * answers `true`. Same shape as the hole that made every MySQL conflict a 500
+ * in `mapDataError` before #6541.
+ *
+ * The predicate answers the FIRST arm only. It has no opinion about dialect
+ * support, so the second arm stays this module's own — and stays second.
+ *
+ * ⚠️ Pass the **error**, not `err.message`. A string still works (the predicate
+ * reads it on the message channel, and so does {@link indexFailureText}), but a
+ * caller that unwraps first throws away the `code` / `errno` / `cause` channels
+ * that are the whole reason this reads the object.
  */
-export function classifyIndexFailure(message: string): PartialIndexStatus {
-    if (/unique constraint failed|duplicate entry|duplicate key value|violates unique/i.test(message)) {
+export function classifyIndexFailure(error: unknown): PartialIndexStatus {
+    if (isUniqueViolationError(error)) {
         return 'conflict';
     }
-    if (/partial|where clause|near "where"|near 'where'|functional|syntax/i.test(message)) {
+    if (/partial|where clause|near "where"|near 'where'|functional|syntax/i.test(indexFailureText(error))) {
         return 'unsupported';
     }
     return 'failed';
@@ -169,9 +213,14 @@ export async function probeThenReplaceIndex(
     try {
         await exec(buildSql(probeIndexName));
     } catch (err: unknown) {
+        // `detail` is the OPERATOR-facing text and stays the driver's own prose.
+        // The VERDICT is taken from the error object itself, so a conflict
+        // reported on `code` / `errno` / `cause` with unhelpful prose is still
+        // classified as one (#6699) — unwrapping first is exactly what the
+        // migration onto the shared predicate exists to stop.
         const detail = err instanceof Error ? err.message : String(err);
         await dropIndexQuietly(exec, probeIndexName);
-        return { status: classifyIndexFailure(detail), detail, failedAt: 'probe' };
+        return { status: classifyIndexFailure(err), detail, failedAt: 'probe' };
     }
     await dropIndexQuietly(exec, probeIndexName);
 
diff --git a/packages/metadata-protocol/src/migrations/view-definition-active-index.test.ts b/packages/metadata-protocol/src/migrations/view-definition-active-index.test.ts
index 863f8a3fd1..af90d0400c 100644
--- a/packages/metadata-protocol/src/migrations/view-definition-active-index.test.ts
+++ b/packages/metadata-protocol/src/migrations/view-definition-active-index.test.ts
@@ -469,6 +469,14 @@ describe('sys_view_definition active-row uniqueness (#5839) on a NULL-safe key (
         // MariaDB's refusal of a functional key part, which #6417 introduces.
         expect(classifyIndexFailure('Functional index on a column is not supported')).toBe('unsupported');
         expect(classifyIndexFailure('disk I/O error')).toBe('failed');
+        // #6699: the same verdict off the `code` channel, with prose that
+        // carries no signal at all. Asserted through THIS module's re-export
+        // (the public `@objectstack/metadata-protocol` surface), because that is
+        // the export the classifier's own home is reached by — the full
+        // channel matrix lives in `partial-index-probe.test.ts`.
+        expect(
+            classifyIndexFailure(Object.assign(new Error('insert failed'), { code: 'ER_DUP_ENTRY' })),
+        ).toBe('conflict');
     });
 
     it('buildActiveIndexSql scopes rows AND spells the key NULL-safe', () => {