Finalize on a bare dynamic array or variant

Split out of [[feature-a-implement-initialize-and-finalize-over-the-arc-helpers]] while landing it, 2026-08-21.

What works and what does not

An error, deliberately, not a no-op — a silent no-op on a managed type is the exact defect the parent ticket was filed for, so the gap says so out loud.

Why it stopped there

PXXDynArrayRelease(data, desc) needs the array's ELEMENT layout descriptor. Records reach theirs through IR_CONST_DATA with -(RECORD_RTTI_DATAREF_BASE + ci), resolved by the pre-link fixup pass in compiler.pas. A per-SYMBOL element descriptor is GetOrAllocSymRTTI(symIdx), which is codegen-level — and SYM_RTTI_DATAREF_BASE (300000) is declared in defs.inc but has no fixup branch, so there is no IR-level route to it.

So the work is one of:

1. Give SYM_RTTI_DATAREF_BASE a real fixup branch, making per-symbol descriptors reachable from IR the way record ones are. Probably the right fix — the constant already exists and reads as if it worked.
2. Or route the bare cases through the record path by other means.

Variants are cheaper and separable: Finalize(v) is PXXVarClear(@v) (already idempotent — it clears in place), and Initialize(v) is zeroing 16 bytes, since varEmpty is 0. That half could land on its own.

Why it is low prio

The gap is genuinely narrow: a record containing a dynamic array or variant is fully handled (PXXRecordFinalize walks to PXXDynArrayRelease with the sub-descriptor), and a bare one has a one-line workaround that does exactly what Finalize must — a := nil releases the old reference and nils, and VarClear(v) exists. Neither is on the parent ticket's motivating path (a GetMem'd record).

Gate

Track A: repro proving a bare dyn array's elements are released and a copy taken beforehand survives; test_initialize_finalize extended; self-host byte-identical.