A typed const array of string[N] is never initialised — silently

• Type: bug (silent wrong value) — Track A
• Found: 2026-08-10 by an FPC differential over the recursion / sorting / const-table surface. A lookup table of short strings is an everyday construct (month names, enum labels, keyword tables).
• Pre-existing: identical on pinned.

const Names: array[0..2] of string[8] = ('aa','bb','cc');
begin WriteLn('[', Names[0], ']'); end.

FPC prints [aa]. pxx prints eight NUL bytes and trailing garbage — no diagnostic.

Boundary (measured, fpc -O1 as oracle)

So it is exactly the FROZEN string forms, and the managed string sibling beside them works — the usual two-spellings split.

The equivalent VAR array is fine: var a: array[0..2] of string[8] then a[0] := 'aa' reads back aa. So the storage and the ordinary assignment path are healthy; only the const INITIALISATION is wrong.

Measurements to start from (mechanism NOT yet established — do not assume)

• SizeOf agrees between the const and the var array (24 for array[0..2] of string[8]), so the two symbols are laid out the same and the difference is in the initialisation, not the allocation.
• Length(Names[0]) on the const answers a value that looks like an ADDRESS (e.g. 4243361), while Length(a[0]) on the uninitialised var answers 0. Something pointer-shaped is reaching the slot.
• The parser DOES have a frozen-string arm for array-const elements: the element loop tests cElemTk in [tyString, tyAnsiString, tyFixedString, tyShortString] and records a Kind=1 (string-literal span) init. The Kind=1 emitter (parser.inc, the PendingInitFOff[i] = -3 or PendingInitKind[i] = 1 arm) builds an AN_STR_LIT tagged Ord(tyString).
• The nearby comment on the array-const path still says "Ordinal/Char/Int64 element types; string/float/record initializers remain a follow-up", which predates the Kind=1 arm — so the parse side was extended and the comment (and possibly the emit side) was not.

I did not confirm which of those is the fault, and a plausible-sounding story here would be worth nothing — see devdocs/dev/debugging-playbook.md. Dump the emitted init with PXXDBG=a.ast:main / a.ir:main and compare against the working a[0] := 'aa' lowering before concluding anything.

Note for whoever takes it

Nothing in lib/, compiler/, examples/ or test/ currently uses this construct (checked), so there is no in-repo code depending on today's behaviour — a diagnostic is a safe intermediate step if the real fix is large. Silent garbage is the one outcome that must not survive, since the array-of-record sibling right beside it already refuses what it cannot do (array-of-record constant with string fields must be global).

Gate

The six element types above matching FPC, plus a named-alias element and an N-D const array of string[N]; test/test_const_array_of_string.pas still green; self-host byte-identical.

Log

• 2026-08-11 — resolved, commit 6900505aa.