`Str` of a Boolean formats it as a digit

Track P — the Str intrinsic's formatter dispatch, pasparser_stmt.inc (the CaseEqual(name, 'Str') arm).
Found 2026-09-04 while measuring what a library writeln would have to reproduce for phase 3 of [[feature-writeln-as-library]]. Not looked for.

Repro

var s: ShortString; b: Boolean;
begin
  b := True; Str(b, s);
  writeln('Str=', s, ' writeln=', b);   { pxx: Str=1 writeln=TRUE }
end.                                    { fpc: Str=TRUE writeln=TRUE }

What it was

Str's dispatch chose between StrFloat, StrQWord and StrInt. Boolean was explicitly excluded from the StrQWord arm — correctly, since a Boolean is an unsigned ordinal and would otherwise have printed through the unsigned formatter — and then fell into StrInt, which prints digits.

StrBool has existed the whole time in compiler/builtin/builtin.pas and is correct (StrStrW('TRUE', width)). Nothing routed to it from Str. The write/writeln path already did, which is why the two disagreed.

Fix: one arm, claiming tyBoolean before the unsigned test. The <> tyBoolean exclusion in the StrQWord arm then became a condition that could never fire, and was deleted rather than left looking like a live guard.

Why it was findable

Directly above the fix sits the comment for bug-p-str-of-a-qword-formats-it-signed: "Str(q, s) on a QWord >= 2^63 produced -1 — a silent wrong VALUE, while writeln(q) two lines away was right." Identical shape, one type over, and that ticket's own fix is what put the dispatch table under a reader's eye. Str's table is a hand-written copy of write's, and it has now been short a case twice; the third renderer, array of const boxing, loses the same information for the sized booleans (filed as [[bug-a-the-sized-booleans-render-as-a-digit-in-both-str-and-writeln]]).

Test

test/test_str_of_boolean.pas, wired into test-core, byte-identical to fpc 3.2.2 -Mdelphi -O1.

The expected values are ones the broken path cannot produce (TRUE, not 1) — a row asserting 1 would have passed against the bug. Two properties of the file are deliberate:

It compares its whole output, not the tail line. The pre-fix compiler still printed the final STR BOOL OK, so a tail-only assertion — the shape the neighbouring elision tests use — would have been green against the bug.
The wt=/wf= rows render the same value through writeln. On the pre-fix compiler they stay GREEN while the Str rows go red, so the failure output shows the two-renderings split rather than just reporting a mismatch.

Control, run rather than assumed: with the arm removed and the old exclusion restored, 6 of 11 rows go RED (t=1, f=0, and all four width rows) and the writeln rows stay green. Restoring the fix returned the compiler to byte-identical sha c94252bb92cd, so the control changed nothing else.

Gate

make compiler/pascal26 — converged after 1 round(s), the recompute verb. tools/gate.sh quick 12/12 PASS with the FPC seed canary PASS rather than SKIP (gated before committing, which is the only condition under which it runs at all).

Not fixed here

WordBool/LongBool/ByteBool still print 1 from both renderers. That is a different defect with a different cause — they have no boolean-ness to dispatch on — and is filed separately.

Log

2026-09-04 — found, fixed, tested and closed in one pass, commit 14c8dcbb7. The fix and the close are the same commit; there is no separate close to cite. Test test/test_str_of_boolean.pas and the reverted-arm control are both in it.

Str of a Boolean formats it as a digit