A helper's const was visible as a bare global

Reported

Track T, test-core#src:test/test_record_helper_for_string_b331.pas, bad 2e7286e499d1, last good 9ac7e74e367b, 3 commits in range → ee388cf3a fix(P): a typed class/record const was global, so two owners shared one slot. Still red at HEAD when re-measured.

pascal26:71: error: undefined variable (BITS)

The test was wrong, not the compiler

Line 71 read:

  Writeln('bits:   ', BITS);                     { helper const (global scope) }

BITS is const BITS = 32; inside TU32Helper = record helper for Cardinal. The comment says the quiet part: the test asserted that a helper's const is visible as a plain global.

Oracle (FPC 3.2.2, on the type helper for LongWord spelling FPC accepts):

hc.pas(14,21) Error: (5000) Identifier not found "BITS"

FPC refuses the bare read and requires TU32Helper.BITS. So the red is a divergence being CLOSED, not opened. ee388cf3a gave typed class/record consts their own mangled backing symbol; the untyped helper const stopped being minted as a bare global on the way past.

Same shape as bug-a-a-units-mode-directive-turns-delphi-mode-off-for-the-program, found in the same tstate sweep: the commit that exposes a defect is not the commit that caused it, and in both cases reverting would have restored the bug. Worth stating together because two of two reds in one night had that shape, and the reflex on a bisect result is to revert the named commit.

Change

• The assertion moved to TU32Helper.BITS — what FPC accepts — with the history recorded at the site. The program's output is unchanged, so the Makefile's expected string did not move: the row was red on the COMPILE, and a stale expectation that still produces the right bytes is the kind that survives a long time.
• New test/test_helper_const_not_global_fail.pas asserts the bare form is REFUSED, in both arms — untyped (folded as a literal, no symbol) and typed (real storage, mangled backing name) reach a helper const through different machinery, so the regression can return on either. A sibling that only reads the const QUALIFIED cannot tell the difference, which is exactly how this went unnoticed.

No compiler change: the behaviour is already right.

Honest note on the negative control

There isn't one, and that is better said than implied. pinned does reject the new test, but on the line ABOVE — class method not found: WIDTH — because it predates qualified typed-const access entirely and never reaches the bare read. The test locks behaviour that is correct today; it is not evidence about yesterday.

Gate

make compiler/pascal26 + both rows + tools/gate.sh quick.