Variant->string drops the boolean tag (empty string, silently)

• Type: bug (silent wrong value) — Track A (the helper is VariantToStr in compiler/builtin/builtin.pas; a builtin change needs a re-pin)
• Found: 2026-08-13, while probing FPC's variant conversions for [[bug-p-variant-to-int-and-char-conversion-diverges-from-fpc]].
• Pre-existing — reproduced against the PINNED compiler, so it is NOT from [[bug-p-a-typecast-of-a-variant-reinterprets-it-instead-of-converting]]. That fix does make it reachable through one more spelling (String(v), which used to be a hard Error), which is how it surfaced.

program vs;
uses variants;
var v: Variant; s: AnsiString;
begin
  v := True;
  s := v;                    writeln('assign  s=[', s, ']');
  writeln('writeln v=[', v, ']');
end.

Root cause

VariantToStr (builtin.pas ~604) tests VType 1/2 (int), 3 (float), 5 (char), 6 (string), 8193 (promo), 0 (None) — and has no arm for 4, the boolean tag, so it falls through to the trailing else Result := ''. VariantTagName right below it does know t = 4 is 'a boolean', which is what makes the omission look like a slip rather than a decision.

The shape

Variant->TEXT has (at least) two independent sites: the writeln path renders a boolean correctly, VariantToStr does not. Same concept, two mechanisms, one of them with a hole — devdocs/dev/normalise-dont-special-case.md. Before closing, check whether the two can share one renderer rather than adding a fourth arm to the second copy; and grep for any other variant->text site (the NilPy side has pystr_of, which is deliberately separate and spells tags Python's way — that one is not a duplicate to merge).

Related: the same file's other missing-tag behaviour is worth a sweep while in there — VType 7 (object) also lands in the '' else.

Gate

make test + self-host fixedpoint, then re-pin (tools/testmgr.py --pin) — builtin changes do not reach the gate's fixedpoint until pinned. Add the repro above to the variant tests, diffed against FPC.

Progress (2026-08-13)

Fixed. VariantToStr grew the missing VT_BOOL arm ('True' / 'False', FPC's spelling). Done as a DEPENDENCY of [[bug-p-variant-to-int-and-char-conversion-diverges-from-fpc]] rather than on its own: FPC's Char(True) is 'T' — character 1 of the string form — so the --strict-fpc Char path could not be right until the string path was.

On the "should the two renderers merge?" question the ticket raised: not merged, deliberately. The writeln path and VariantToStr are not duplicates that drifted — writeln renders straight to the output buffer with no AnsiString temp, which is why it exists separately. Merging would put an allocation in every variant write. What the change DOES do is make the new VariantToCharFPC build on VariantToStr instead of walking the tags a third time, so the strict-Char rule and the string rule cannot drift apart.

VType 7 (object) still lands in the trailing '' — left as-is: unlike a boolean, an object has no obvious text form, and FPC raises rather than rendering. Filing that as a separate question would be inventing work; if it bites, the tag walk is one place.

Verified: s := v on a boolean now yields True/False, matching FPC; covered by the new rows in test/test_variant_typecast.pas (in make test).

Log

• 2026-08-13 — resolved, commit 86410fdd1.