The question

tyRecord describes the STORAGE of three unrelated things: a real record, an interface (a 16-byte {IMT, instance} fat pointer) and a method pointer (a 16-byte {Code, Data} pair). So every rule that refuses something because the destination is tyRecord must first ask "is this actually a record?" — and that question has two conditions, always the same two.

The third row is the defect, filed and fixed as the method-pointer half of bug-p-methodptr-nil-assign: the three arms added by fa8f2424d inherited one of the two conditions, so c.OnHit := nil and arr[1] := nil were refused with cannot assign Pointer to record while ev := nil kept working.

That fix is a one-line bail and it is the right fix. This ticket is the observation underneath it: the pair had no name, so a new call site had to re-derive it, and re-derived half.

Shape

One predicate — RecIsFatPointer(rec: Integer): Boolean, MethodPtrRecId or UClsIsInterface, both with the REC_NONE and MAX_UCLASS guards the existing copies already carry — called from all three rows. Note the AN_IDENT row asks its half a different way (SymProcSig >= 0, which bails on any procvar, not just a method pointer); check before collapsing it whether that breadth is load-bearing there or incidental, because a plain procvar is spelled tyPointer and the record rule cannot fire on it.

Do not treat "they agree today" as evidence. Row three agreed with row one about interfaces and disagreed about method pointers, and the disagreement was a compile error on ordinary code (OnClick := nil is how you detach a handler). Build the control the way refactor-a-two-predicates-answer-what-a-caret-yields did: a probe per condition per site, run against a binary with the predicate and one without, so the takeover is positive rather than an absence of complaints.

Gate

Track A: make compiler/pascal26 (fixedpoint) + tools/gate.sh quick, pinned by test/test_methodptr_nil_assign.pas and the assign fail/ok pair (test_assign_lvalue_shapes_fail / _ok, whose 12-row count is itself the assertion).

Rejected — it had already landed when this was written

2026-08-30, frankA, the same hour. d5fd2a6ca (frankS) is exactly the shape this ticket asked for:

• symtab.inc:8116 — RecIsReferenceShaped(rec), MethodPtrRecId or UClsIsInterface, one name for the pair.
• symtab.inc:8170 — ProcParamIsNilable delegates to it.
• ir.inc:106 and ir.inc:151 — both AssignSideKind arms call it, so the drift this ticket was about is gone rather than documented.

Measured rather than read off the diff: rebuilt at HEAD (fixedpoint, 2 rounds) and ran test/test_methodptr_nil_assign.pas — compiles, and all three cleared shapes report assigned=FALSE after being armed and called.

Why it was filed anyway, which is the part worth keeping

I read symtab.inc:8130-8141 at the sha in my checkout and wrote the ticket from it. The predicate had existed for forty minutes. git fetch/pull before filing — not merely before writing — is the cheap fix, and it is the same error the coordinator made tonight from the other side: an absence read as a fact about the repo when it was a fact about one working tree.

A ticket asserts a present-tense claim about the codebase. That makes filing one a measurement, subject to the same staleness rule as any other, and the board is where a stale claim is most expensive: nobody re-derives it, they act on it. This one would have sent its resolver to write a predicate that was already three call sites deep.