Foreslaa regel: xRec unreliable within same transaction as prior Insert or Modify

custom/knowledge/architecture/xrec-unreliable-within-same-transaction-as-prior-insert-or-modify.md

@@ -0,0 +1,85 @@
+---
+bc-version: [all]
+domain: architecture
+keywords: [xrec, trigger, onmodify, ondelete, transaction, test-codeunit, change-detection]
+technologies: [al]
+countries: [w1]
+application-area: [all]
+---
+
+# xRec is unreliable for change detection within the same uncommitted transaction as a prior Insert or Modify
+
+## Description
+
+The standard AL idiom for detecting whether a field changed inside `OnModify`
+(or `OnDelete`) is `if "Field" <> xRec."Field" then ...` — this is used
+throughout the base application and is normally trustworthy. But `xRec` was
+observed, empirically and reproducibly, to **mirror `Rec`'s new value instead
+of the record's stored value** when the `Modify` call runs against a record
+that was `Insert`ed (or previously `Modify`ed) earlier in the same
+**uncommitted transaction** — no explicit `Commit` in between.
+
+Confirmed via three successive diagnostic `Error()` calls inside the trigger
+itself: with the record obtained by `Get()` immediately before the field
+change (not even the same in-memory variable that did the `Insert`), `xRec`
+still showed the new value, identical to `Rec`. A **fresh `Record.Get()` call
+issued from inside the same trigger**, on the same primary key, correctly
+returned the true stored (old) value.
+
+This is exactly the shape of the most common AL test codeunit: `Insert` a
+record, then `Modify` it later in the same `[Test]` method to exercise an
+`OnModify` rule — no `Commit` between them, because test methods run inside
+an implicit rollback transaction. Any table trigger authored and validated
+only against that pattern will appear correct while silently never firing.
+
+## Rule
+
+A table trigger that needs a field's previously-stored value to decide
+whether something changed must not trust `xRec` alone if the record may have
+been `Insert`ed or `Modify`ed earlier in the same uncommitted transaction —
+re-fetch the comparison value explicitly with a fresh `Get()` call inside the
+trigger instead:
+
+```al
+trigger OnModify()
+var
+    Existing: Record "Some Table";
+begin
+    Existing.Get(PrimaryKeyFields);
+    if "Field" <> Existing."Field" then
+        Error(FieldLockedErr);
+end;
+```
+
+This costs one extra `Get()` per `Modify` call — negligible — and is correct
+regardless of whether a `Commit` happened since the record was last written.
+
+## What NOT to do
+
+- Do not assume `xRec` is trustworthy just because the trigger "looks like"
+  every other `xRec`-comparison trigger in the base app — those are typically
+  exercised through a client session where each user action commits, not
+  through code that inserts and modifies in one uncommitted batch
+- Do not "fix" a failing test by adding a `Commit()` between the `Insert` and
+  the `Modify` — that breaks the test's atomicity/rollback isolation and
+  papers over a real production risk: any code path that inserts and later
+  modifies a record without an intervening commit (batch jobs, import
+  routines, cascades) hits the same silent misdetection
+
+## Signal to watch for
+
+An `OnModify`/`OnDelete` trigger comparing `Rec`/`"Field"` against
+`xRec."Field"` to gate a rule (block a change, trigger a cascade, decide
+whether to log), where a test exercises it by inserting the record and then
+modifying it within the same `[Test]` method — and the test passes even
+though manual inspection shows the guarded condition should have fired. In
+review: any `<> xRec.` comparison inside a trigger, paired with a same-method
+Insert-then-Modify in its test.
+
+## Message to developer
+
+When an `xRec`-based trigger check appears to silently not fire in a test
+that inserts then modifies the same record: don't assume the guard condition
+or the test data is wrong — diagnose with a throwaway `Error()` printing both
+`xRec."Field"` and a fresh `Get()`'s value side by side. If they differ,
+switch the trigger to compare against the fresh `Get()`, not `xRec`.