Assigning a dynamic array of interfaces is lowered as an interface assign

• Track A (compiler/ir.inc assignment path; compiler/symtab.inc helper).
• Found 2026-08-20 by an FPC differential probe of interfaces held in containers.

Measured

var a, b: array of IFoo;
begin
  SetLength(a, 1);
  a[0] := TFoo.Create('z');
  writeln(a[0].Name);    { prints z }
  b := a;                { SIGSEGV }

Identical on pinned and on HEAD before the fix. FPC runs all of them.

Cause

An array's TypeKind IS its element kind. So b := a arrives at the assignment path with lhsTk = tyRecord, ResolveNodeRec answers IFoo, and the COM-interface arm concluded the destination was an interface VALUE. It emitted

PXXIntfAssign(@b, @a, ifaceId)

which reads a dynamic-array handle as an interface fat pointer and dispatches _Release through whatever the first word points at.

That the all-nil case crashed too is the tell: nothing about the elements is involved — the ARRAY HANDLE itself is being treated as an object reference.

Fix

isComIntf now also requires that the LHS is not a whole array, via a new ASTNodeIsWholeArray. An array assignment is a whole-array copy and belongs to the paths below, whatever its elements are.

Verified that the neighbouring shapes take the correct path: a static array of interfaces copies by value, a record containing an interface array copies, and g.arr := h.arr (a field that is a whole interface array) works.

The pattern, fourth time today

This is the same mistake as bug-a-a-local-array-of-interfaces-is-not-zero-initialised and bug-a-a-record-with-an-interface-field-is-not-zero-initialised, seen from the other side. Those asked "is this refcounted?" of a CONTAINER and got False when they needed True; this one asked it of an array and got True when it needed False. The container's shape and the member's kind are different questions, and TypeKind cannot tell them apart because an array borrows its element's kind. Every pass that reads TypeKind = tyRecord on a possibly-array symbol is suspect; the helpers accumulated today (RecIsComInterface, SymElemIsComInterface, ASTNodeIsWholeArray) are the beginning of asking it properly.

Found in passing — NOT fixed here

Three more, each measured, each filed:

• bug-a-setlength-shrink-does-not-release-dropped-interface-elements — shrinking 4→2 destroys 0 where FPC destroys 2.
• bug-a-a-local-dynamic-array-of-interfaces-is-not-released-at-scope-exit — a routine filling 2 elements and returning destroys 0.
• bug-a-two-function-result-interfaces-into-a-local-dyn-array-segfault — the sharpest of the three, and a crash.

Test

test/test_dynarray_of_interfaces_assign.pas — 6/6, identical to FPC. The assign, the all-nil assign, handle aliasing, static-array copy-by-value, a record holding an interface array, and the integer/string element types that always worked (so a fix cannot break them). The pinned binary segfaults immediately.

Gate

make compiler/pascal26 (fixedpoint) + tools/gate.sh quick.