fix(metadata-protocol): dotted-path SORT hint prescribes a stored field, not a formula (#6924) (#6996)

`assertSortFieldsExist` refuses a dotted `orderBy` and then told the author how
to fix it: "Denormalise the value onto '<object>' (a formula or rollup field
that copies it into a real column) and sort by that." Following that lands the
author back inside the exact silent degradation the refusal saved them from.

Measured on a REAL SqlDriver (better-sqlite3) and on InMemoryDriver, with a
`formula` field named directly in `orderBy` (non-dotted, so the gate lets it
through):

  control   orderBy title asc    -> A B C D E      a real column really sorts
  baseline  no sort              -> C A E B D      insertion order
  orderBy   <formula> asc        -> C A E B D  200 insertion order
  orderBy   <formula> desc       -> C A E B D  200 direction-blind

`SqlDriver.createColumn` returns early for `formula` (no column), sqlite answers
"no such column", the #3821 backstop retries WITHOUT the sort, and the response
is 200 with every row present in arbitrary order.

`rollup`/`summary` is dropped for a DIFFERENT reason, and the measurement
contradicts the reported diagnosis: a summary field does get a real, maintained
column (orderBy <summary> desc -> E D C B A over values 5 4 3 2 1). It simply
cannot do this job, since a rollup aggregates CHILD records and cannot carry a
looked-up parent's column onto the queried object.

This overturns #4256's recorded wording choice (closed `completed`), which
explicitly picked the "formula or rollup" phrasing as its remedy. The docs
callout at content/docs/protocol/objectql/query-syntax.mdx taught the same
denormalization, so code and docs agreed with each other about something untrue;
both move here. "Stored" is #6673's vocabulary for the identical correction on
the search axis.


Claude-Session: https://claude.ai/code/session_01W6bLax4KMrSfnE1ydFU8Dw

Co-authored-by: Claude <noreply@anthropic.com>

Author: os-zhuang

.changeset/sort-hint-prescribes-stored-field.md

@@ -0,0 +1,55 @@
+---
+"@objectstack/metadata-protocol": patch
+---
+
+fix(data): the dotted-path `400 INVALID_SORT` hint prescribes a **stored** field, not a formula (#6924)
+
+`assertSortFieldsExist` refuses a dotted `orderBy` (`?sort=account.company_name`)
+and then told the author how to fix it: *"Denormalise the value onto '<object>'
+(a formula or rollup field that copies it into a real column) and sort by that."*
+That prescription cannot be built. Following it lands the author back inside the
+exact silent degradation the refusal had just saved them from.
+
+Measured on a REAL `SqlDriver` (better-sqlite3) and on `InMemoryDriver`, with a
+`formula` field named directly in `orderBy` (non-dotted, so this gate lets it
+through):
+
+```
+control   orderBy title asc     -> A B C D E      a real column really sorts
+baseline  no sort               -> C A E B D      insertion order
+orderBy   <formula field> asc   -> C A E B D  200 insertion order
+orderBy   <formula field> desc  -> C A E B D  200 direction-blind
+```
+
+A `formula` field is virtual — `SqlDriver.createColumn` returns early for it and
+no column is created (sqlite answers `no such column`), the engine evaluates the
+expression *after* the driver returns, and the #3821 unknown-column backstop
+retries WITHOUT the sort. The response is `200`, every row present, order
+arbitrary: the failure mode #4226/#4256 exist to stop.
+
+The hint now reads: *"Denormalise the value onto '<object>' (a stored field,
+written when the source changes) and sort by that. Not a formula field: it is
+virtual, no driver materialises a column for one, and ORDER BY on it is silently
+dropped."* — "stored" being the same word #6673 landed for the identical
+correction on the search axis.
+
+`rollup`/`summary` is dropped from the hint for a different reason, and the
+measurement is worth recording because it contradicts the reported diagnosis: a
+`summary` field **does** get a real, maintained column (`orderBy <summary> desc`
+returned `E D C B A` over values `5 4 3 2 1`), so it is not unmaterializable. It
+simply cannot do this job — a rollup aggregates CHILD records
+(`count`/`sum`/`min`/`max`/`avg`) and so cannot carry a looked-up parent's column
+onto the queried object.
+
+**This overturns a recorded decision.** #4256 (closed `completed`) explicitly
+chose the "formula or rollup" wording as its remedy for dotted-path sort, and its
+own still-pending changeset (`sort-dotted-path-rejected.md`) describes it; that
+file is left as the accurate record of what #4256 shipped, and this entry
+supersedes its prescription. `content/docs/protocol/objectql/query-syntax.mdx`
+("Sorting on Related Fields") taught the same denormalization and is corrected in
+the same change, so code and docs stop agreeing with each other about something
+untrue.
+
+Not fixed here, filed separately: the platform still accepts a **non-dotted**
+`orderBy` naming a `formula` field and answers `200` in arbitrary order. That is
+an engine/driver-side refusal question, not hint text.

content/docs/protocol/objectql/query-syntax.mdx

@@ -531,12 +531,33 @@ no driver can order by it — `SqlDriver` would render it as
 `"account"."company_name"` against a table that was never joined, and until the
 path was refused, the unknown-column backstop retried **without the sort** and
 answered 200 with unordered rows. Denormalise the value onto the queried object
-(for example with a formula or rollup field) when you need to sort by it.
+as a **stored** field — one this object's own rows carry, written when the
+source changes — when you need to sort by it.
 
 Internal callers reaching `engine.find()` directly are unaffected: a dotted
 `orderBy` there still falls through to the driver backstop and orders nothing.
 </Callout>
 
+<Callout type="warn">
+**Do not denormalise onto a `formula` field to sort by it.** A `formula` field
+is virtual: no driver materialises a column for it (the engine evaluates it
+*after* the driver returns), so `ORDER BY` on one hits the same unknown-column
+backstop and is **silently dropped** — 200, every row present, arbitrary order.
+Measured on a real `SqlDriver` (better-sqlite3) and on `InMemoryDriver`: rows
+inserted `C A E B D` come back `C A E B D` for both `asc` and `desc`, while the
+same query on a stored column returns `A B C D E` / `E D C B A`.
+
+A `rollup`/`summary` field *does* get a real, maintained column and can be
+sorted on — but it aggregates **child** records (`count`/`sum`/`min`/`max`/
+`avg`), so it cannot carry a looked-up parent's column such as
+`account.company_name`. For that, write the value onto a stored field of the
+queried object and keep it in sync (a trigger or flow on the source record).
+
+This page taught the `formula`/`rollup` version until #6924, and so did the
+`400 INVALID_SORT` hint itself (#4256) — both are corrected together. The same
+correction on the search axis is #6673.
+</Callout>
+
 ---
 
 ## 4. Relationships (Expand)

packages/metadata-protocol/src/protocol.ts

@@ -4819,6 +4819,41 @@ export class ObjectStackProtocolImplementation implements
      * gets a message that says which relationship it tried to cross and
      * prescribes what `query-syntax.mdx` has prescribed since #4240:
      * denormalise the value onto the queried object and sort by that.
+     *
+     * [#6924] WHAT to denormalise onto was wrong, and this overturns #4256's
+     * own recorded wording. That issue chose "a formula or rollup field that
+     * copies it into a real column" — a prescription the platform cannot
+     * deliver, so the refusal handed the author a dead end at the exact moment
+     * they asked for help. Measured on a REAL `SqlDriver` (better-sqlite3) and
+     * on `InMemoryDriver`, with a `formula` field named directly (NOT dotted,
+     * so this gate lets it through):
+     *
+     * ```
+     * control  orderBy title asc    -> A B C D E      (a real column sorts)
+     * baseline no sort              -> C A E B D      (insertion order)
+     * orderBy  <formula field> asc  -> C A E B D  200 (insertion order)
+     * orderBy  <formula field> desc -> C A E B D  200 (direction-blind)
+     * ```
+     *
+     * No column exists to order by (`SqlDriver.createColumn` returns early for
+     * `formula`; sqlite answers `no such column`), the #3821 unknown-column
+     * backstop retries WITHOUT the sort, and the response is 200 with every
+     * row present in an arbitrary order — the very failure #4226/#4256 exist
+     * to stop. Following the old hint therefore landed the author back inside
+     * the defect they had just been refused for.
+     *
+     * `rollup`/`summary` was the other half of that wording and is NOT broken
+     * the same way — it does get a real, maintained column (`table.float`;
+     * measured: `orderBy <summary> desc` -> E D C B A over values 5 4 3 2 1).
+     * It is dropped from the hint because it cannot do THIS job: a rollup
+     * aggregates CHILD records (count/sum/min/max/avg), so it cannot carry a
+     * looked-up parent's column (`account.company_name`) onto this object.
+     * Wrong tool, not a broken one — naming it here still sends the author
+     * somewhere that cannot work.
+     *
+     * "Stored" is #6673's vocabulary for the same correction on the SEARCH
+     * axis (`validate-searchable-fields.ts`, "a stored text field"); the two
+     * axes deliberately say the same word.
      */
     private assertSortFieldsExist(object: string, orderBy: ReadonlyArray<{ field: string }>, param: string): void {
         if (orderBy.length === 0) return;
@@ -4855,8 +4890,10 @@ export class ObjectStackProtocolImplementation implements
                   + "not values inside them")
             + (dotted.length > 1 ? ` (also: ${dotted.slice(1).join(', ')})` : ''),
             {
-                hint: ` Denormalise the value onto '${object}' (a formula or rollup field that`
-                    + ' copies it into a real column) and sort by that.',
+                hint: ` Denormalise the value onto '${object}' (a stored field, written when the`
+                    + ' source changes) and sort by that. Not a formula field: it is virtual,'
+                    + ' no driver materialises a column for one, and ORDER BY on it is silently'
+                    + ' dropped.',
                 extra: { field: first, fields: dotted, object },
             },
         );

packages/objectql/src/query-expression-conformance.test.ts

@@ -450,9 +450,34 @@ describe('#4226 — sort / select / expand on the list path (real ObjectQL engin
             });
     });
 
-    it('the dotted rejection names the relationship it tried to cross and prescribes the fix', async () => {
-        await expect(protocol.findData({ object: 'showcase_task', query: { sort: 'project_id.name' } }))
-            .rejects.toThrow(/follows the relationship 'project_id'[\s\S]*formula or rollup/);
+    it('the dotted rejection names the relationship it tried to cross and prescribes a STORED field', async () => {
+        // [#6924] The prescription is part of the contract, not decoration: a
+        // refusal that hands the author an unbuildable fix is the same dead end
+        // as no hint at all. #4256 chose "a formula or rollup field that copies
+        // it into a real column"; measured on a REAL SqlDriver (better-sqlite3)
+        // and on InMemoryDriver, `orderBy` naming a `formula` field answers 200
+        // with the rows in INSERTION order, identically for asc and desc — no
+        // column exists, so the #3821 backstop retries without the sort. That
+        // is the exact silent degradation this gate exists to stop, so the old
+        // hint routed the author back into it.
+        const err: any = await protocol
+            .findData({ object: 'showcase_task', query: { sort: 'project_id.name' } })
+            .then(() => null, (e: unknown) => e);
+        expect(err).toBeTruthy();
+        // ADR-0112 envelope — a rejection case asserts code AND status, not
+        // merely that something was thrown.
+        expect(err.status).toBe(400);
+        expect(err.code).toBe('INVALID_SORT');
+        expect(err.message).toMatch(/follows the relationship 'project_id'/);
+        // The remedy must be a STORED field — #6673's vocabulary for the same
+        // correction on the SEARCH axis, deliberately the same word here.
+        expect(err.message).toMatch(/a stored field/);
+        // ...and the old prescription must be gone, not merely joined.
+        expect(err.message).not.toMatch(/formula or rollup/);
+        // `formula` may still appear — but only as the named trap, never as the
+        // thing to build. This is what separates the fix from a reword that
+        // keeps the dead end in a subordinate clause.
+        expect(err.message).toMatch(/Not a formula field/);
     });
 
     it('a dotted path under a non-reference head is refused on the same axis, minus the relationship claim', async () => {